Unexpected ice-water shock · Humaneness Voice Large 1.5 Alpha
Two frozen models. Exact instructions, matched seeds and complete take distributions. Find a promising recipe quickly, then inspect its limitations and listen.
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-cfg-five-stress at CFG 5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Re-scored, not regenerated: the original raw EARS predictions were reused on CPU. Old fitness, audio, prompts and scientific gates are retained. Old held-out sets tested only the previously selected wording; a newly re-ranked search winner does not inherit that validation. No universal best weighting or human-preference optimum is claimed.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Start here: best observed recipe, not a universal optimum
The current starting point uses EARS Steering v2 across the available canonical search conditions, with unchanged qualification gates taking priority. Historical candidates are used only where no canonical search exists. Held-out results below refer to the earlier selected wording and are not automatically inherited by a newly re-ranked winner.
No recipe satisfies every realization/quality gate. The highlighted recipe is the best diagnostic candidate, not a reliable recommendation.
Burst caution: family detection anywhere, nominal placement and natural blending are different questions. The pilot cannot establish accurate word-level placement. The follow-up requires a60%family-hit floor in addition to speech/identity QA; a quality-only clip with no detected event does not qualify.
How the strategies compare
Click any column header to sort. Do not compare scalar fitness across different missions or between legacy and event-aware scoring contracts.
Transfer to unseen sentences — earlier selection, now re-scored
The previously selected wording was repeated on five new lexically neutral bilingual scripts, at selected guidance and guidance 1. These outputs are now also re-scored with Steering v2. They do not validate other newly favored wording and are not a human listening validation.
Baseline, leading recipes and runners-up retain every language/reference/seed combination. These are not hand-picked best takes. Players use individual MP3 URLs, no base64 downloads, and preload nothing. Initial view is paginated for fast loading.
English fixed reference ·12seconds
A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.
German fixed reference ·12seconds
A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.
Recipe details and exact prompt contracts
creative-cfg-five-stress · CFG 5
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-cfg-five-stress at CFG 5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.822
6.102
4.106
2.42%
0.005
100.00%
0.007
All expected burst families anywhere: 2.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. The involuntary sound breaks through the spoken performance. Preserve all words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. The involuntary sound breaks through the spoken performance. Preserve all words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. The involuntary sound breaks through the spoken performance. Preserve all words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. The involuntary sound breaks through the spoken performance. Preserve all words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. The involuntary sound breaks through the spoken performance. Preserve all words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. The involuntary sound breaks through the spoken performance. Preserve all words.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. The involuntary sound breaks through the spoken performance. Preserve all words.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. The involuntary sound breaks through the spoken performance. Preserve all words.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.814
6.107
4.098
2.73%
0.602
none
50
0.830
6.098
4.115
2.10%
n/a
creative-cfg-four-stress · CFG 4
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-cfg-four-stress at CFG 4; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.821
6.146
4.149
2.97%
0.018
99.00%
0.009
All expected burst families anywhere: 6.00%. All nominal event-window matches: 2.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make the reaction physically immediate and unmistakable.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Make the reaction physically immediate and unmistakable.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make the reaction physically immediate and unmistakable.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Make the reaction physically immediate and unmistakable.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make the reaction physically immediate and unmistakable.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Make the reaction physically immediate and unmistakable.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make the reaction physically immediate and unmistakable.) [4.2 seconds duration] That is ice water... I take back everything I just said.
(Shivering indignation mixed with relieved laughter; remain intelligible. Make the reaction physically immediate and unmistakable.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Literal words in the separate text field:
I promised I would stay perfectly calm whatever you were planning. That is ice water... I take back everything I just said. Stop filming for a moment and help me warm up.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] That is ice water... I take back everything I just said.
(Natural clear speech.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Reference versus no-reference breakdown
reference
50
0.824
6.122
4.132
2.94%
0.604
none
50
0.817
6.169
4.167
3.00%
n/a
creative-event-contrast-three · CFG 3
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-event-contrast-three at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.788
6.192
4.180
2.01%
0.000
98.00%
0.037
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. A separate raw vocal reaction interrupts the sentence rhythm.
Three local directions
1. Easy, genuinely amused conversation before anything happens. A separate raw vocal reaction interrupts the sentence rhythm.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. A separate raw vocal reaction interrupts the sentence rhythm.
3. Shivering indignation mixed with relieved laughter; remain intelligible. A separate raw vocal reaction interrupts the sentence rhythm.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. A separate raw vocal reaction interrupts the sentence rhythm.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. A separate raw vocal reaction interrupts the sentence rhythm.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. A separate raw vocal reaction interrupts the sentence rhythm.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. A separate raw vocal reaction interrupts the sentence rhythm.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
[0.32 seconds pause]
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.787
6.180
4.157
2.08%
0.632
none
50
0.790
6.204
4.203
1.94%
n/a
repair-duration_only · CFG 3.0
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to repair-duration_only at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.788
6.235
4.241
2.19%
0.021
100.00%
0.040
All expected burst families anywhere: 9.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.786
6.221
4.212
2.37%
0.634
none
50
0.790
6.248
4.269
2.02%
n/a
repair-canonical_both · CFG 3.0
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to repair-canonical_both at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.777
6.220
4.233
2.14%
0.008
99.00%
0.010
All expected burst families anywhere: 4.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Literal words in the separate text field:
I was having such a lovely, peaceful afternoon out here. No, wait... that went straight down the back of my shirt. You are getting the next bucket, and I mean it.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Natural clear speech.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Reference versus no-reference breakdown
reference
50
0.778
6.228
4.211
2.55%
0.604
none
50
0.777
6.212
4.255
1.72%
n/a
g0-p06 · CFG 3
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p06 at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.760
6.254
4.276
2.02%
0.009
100.00%
0.059
All expected burst families anywhere: 3.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Begin with the feeling held back, then let it leak through and become unmistakable.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Begin with the feeling held back, then let it leak through and become unmistakable.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Begin with the feeling held back, then let it leak through and become unmistakable.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.763
6.271
4.230
2.36%
0.677
none
50
0.757
6.237
4.323
1.67%
n/a
creative-cfg-three · CFG 3
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-cfg-three at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.748
6.253
4.275
1.68%
0.005
100.00%
0.052
All expected burst families anywhere: 2.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction escape before you can control it. Do not merely describe a reaction.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Let the reaction escape before you can control it. Do not merely describe a reaction.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the reaction escape before you can control it. Do not merely describe a reaction.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction escape before you can control it. Do not merely describe a reaction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction escape before you can control it. Do not merely describe a reaction.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.2 seconds duration] That is ice water... I take back everything I just said.
(Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction escape before you can control it. Do not merely describe a reaction.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Literal words in the separate text field:
I promised I would stay perfectly calm whatever you were planning. That is ice water... I take back everything I just said. Stop filming for a moment and help me warm up.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] That is ice water... I take back everything I just said.
(Natural clear speech.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Reference versus no-reference breakdown
reference
50
0.745
6.255
4.247
1.71%
0.647
none
50
0.752
6.250
4.304
1.65%
n/a
g0-p05 · CFG 2.5
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p05 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.735
6.272
4.282
1.63%
0.006
100.00%
0.034
All expected burst families anywhere: 3.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: attack, sustain and release.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Build an audible attack, sustain the intended quality, then release it without losing the final words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Build an audible attack, sustain the intended quality, then release it without losing the final words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.730
6.272
4.210
1.51%
0.652
none
50
0.741
6.272
4.355
1.74%
n/a
creative-muscular-release · CFG 2.5
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-muscular-release at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.733
6.253
4.320
1.88%
0.019
100.00%
0.042
All expected burst families anywhere: 8.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start from restrained breath and throat tension, release it abruptly, then recover.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Start from restrained breath and throat tension, release it abruptly, then recover.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start from restrained breath and throat tension, release it abruptly, then recover.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Start from restrained breath and throat tension, release it abruptly, then recover.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start from restrained breath and throat tension, release it abruptly, then recover.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start from restrained breath and throat tension, release it abruptly, then recover.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start from restrained breath and throat tension, release it abruptly, then recover.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start from restrained breath and throat tension, release it abruptly, then recover.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Literal words in the separate text field:
I was having such a lovely, peaceful afternoon out here. No, wait... that went straight down the back of my shirt. You are getting the next bucket, and I mean it.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Natural clear speech.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Reference versus no-reference breakdown
reference
50
0.740
6.259
4.273
2.03%
0.637
none
50
0.726
6.248
4.366
1.74%
n/a
creative-literal-interjection-proxy · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-literal-interjection-proxy at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.727
6.296
4.241
1.70%
0.014
100.00%
0.025
All expected burst families anywhere: 5.00%. All nominal event-window matches: 2.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Use a raw open vowel before recovering the intended words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Use a raw open vowel before recovering the intended words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Use a raw open vowel before recovering the intended words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Use a raw open vowel before recovering the intended words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Use a raw open vowel before recovering the intended words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Use a raw open vowel before recovering the intended words.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Use a raw open vowel before recovering the intended words.) [4.2 seconds duration] Ah! That is ice water... I take back everything I just said.
(Shivering indignation mixed with relieved laughter; remain intelligible. Use a raw open vowel before recovering the intended words.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Literal words in the separate text field:
I promised I would stay perfectly calm whatever you were planning. Ah! That is ice water... I take back everything I just said. Stop filming for a moment and help me warm up.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] Ah! That is ice water... I take back everything I just said.
(Natural clear speech.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Reference versus no-reference breakdown
reference
50
0.716
6.279
4.213
1.51%
0.620
none
50
0.737
6.314
4.269
1.90%
n/a
g0-p09 · CFG 2.5
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p09 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.725
6.338
4.304
1.37%
0.012
100.00%
0.005
All expected burst families anywhere: 5.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Let the local direction carry the intended delivery beside each sentence.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Let the local direction carry the intended delivery beside each sentence.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the local direction carry the intended delivery beside each sentence.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Let the local direction carry the intended delivery beside each sentence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Let the local direction carry the intended delivery beside each sentence.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.713
6.326
4.261
1.40%
0.623
none
50
0.736
6.350
4.346
1.33%
n/a
creative-reaction-first · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-reaction-first at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.724
6.286
4.318
1.55%
0.006
100.00%
0.055
All expected burst families anywhere: 3.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React before the first words, as if the physical event has already happened.
Three local directions
1. Easy, genuinely amused conversation before anything happens. React before the first words, as if the physical event has already happened.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React before the first words, as if the physical event has already happened.
3. Shivering indignation mixed with relieved laughter; remain intelligible. React before the first words, as if the physical event has already happened.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React before the first words, as if the physical event has already happened.
SCRIPT:
(surprised gasp, 0.32 seconds)
(Easy, genuinely amused conversation before anything happens. React before the first words, as if the physical event has already happened.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React before the first words, as if the physical event has already happened.) [4.2 seconds duration] That is ice water... I take back everything I just said.
(Shivering indignation mixed with relieved laughter; remain intelligible. React before the first words, as if the physical event has already happened.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Literal words in the separate text field:
I promised I would stay perfectly calm whatever you were planning. That is ice water... I take back everything I just said. Stop filming for a moment and help me warm up.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] I promised I would stay perfectly calm whatever you were planning.
(Natural clear speech.) [4.2 seconds duration] That is ice water... I take back everything I just said.
(Natural clear speech.) [3.8 seconds duration] Stop filming for a moment and help me warm up.
Reference versus no-reference breakdown
reference
50
0.726
6.268
4.248
1.90%
0.650
none
50
0.723
6.305
4.387
1.20%
n/a
creative-event-contrast-two · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-event-contrast-two at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.719
6.297
4.299
1.05%
0.006
100.00%
0.018
All expected burst families anywhere: 2.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. The nonverbal event must actually be heard, then return to the words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. The nonverbal event must actually be heard, then return to the words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. The nonverbal event must actually be heard, then return to the words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. The nonverbal event must actually be heard, then return to the words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. The nonverbal event must actually be heard, then return to the words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. The nonverbal event must actually be heard, then return to the words.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. The nonverbal event must actually be heard, then return to the words.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. The nonverbal event must actually be heard, then return to the words.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
[0.32 seconds pause]
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.704
6.290
4.267
0.54%
0.612
none
50
0.733
6.305
4.332
1.55%
n/a
creative-cfg-two · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-cfg-two at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.711
6.289
4.315
1.47%
0.000
100.00%
0.026
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make the physical reaction interrupt the speech, then recover intelligible words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Make the physical reaction interrupt the speech, then recover intelligible words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make the physical reaction interrupt the speech, then recover intelligible words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Make the physical reaction interrupt the speech, then recover intelligible words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make the physical reaction interrupt the speech, then recover intelligible words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Make the physical reaction interrupt the speech, then recover intelligible words.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make the physical reaction interrupt the speech, then recover intelligible words.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Make the physical reaction interrupt the speech, then recover intelligible words.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.705
6.284
4.267
1.37%
0.621
none
50
0.717
6.294
4.364
1.56%
n/a
creative-short-training-range · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-short-training-range at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.710
6.323
4.251
1.98%
0.008
100.00%
0.017
All expected burst families anywhere: 3.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Use a brief, crisp involuntary reaction, not an extended dramatic sound.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Use a brief, crisp involuntary reaction, not an extended dramatic sound.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.30 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Use a brief, crisp involuntary reaction, not an extended dramatic sound.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Shivering indignation mixed with relieved laughter; remain intelligible. Use a brief, crisp involuntary reaction, not an extended dramatic sound.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Literal words in the separate text field:
I was having such a lovely, peaceful afternoon out here. No, wait... that went straight down the back of my shirt. You are getting the next bucket, and I mean it.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.30 seconds)
(Natural clear speech.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Natural clear speech.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Reference versus no-reference breakdown
reference
50
0.699
6.305
4.203
2.43%
0.612
none
50
0.721
6.341
4.299
1.53%
n/a
creative-long-event-stress · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-long-event-stress at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.708
6.307
4.380
1.87%
0.042
100.00%
0.017
All expected burst families anywhere: 13.00%. All nominal event-window matches: 10.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Sustain the reaction before regaining the ability to form words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Sustain the reaction before regaining the ability to form words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 1.20 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 1.20 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.706
6.306
4.357
2.10%
0.612
none
50
0.710
6.308
4.403
1.65%
n/a
creative-inline-only · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-inline-only at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.698
6.332
4.299
1.37%
0.013
100.00%
-0.001
All expected burst families anywhere: 5.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Follow the local acting direction at each step.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Follow the local acting direction at each step.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Follow the local acting direction at each step.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Follow the local acting direction at each step.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Follow the local acting direction at each step.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Follow the local acting direction at each step.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Follow the local acting direction at each step.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.697
6.308
4.269
1.55%
0.594
none
50
0.699
6.357
4.330
1.18%
n/a
creative-medium-event · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-medium-event at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.697
6.318
4.361
1.63%
0.008
100.00%
0.032
All expected burst families anywhere: 3.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction last clearly long enough to be heard before speaking again.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Let the reaction last clearly long enough to be heard before speaking again.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the reaction last clearly long enough to be heard before speaking again.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction last clearly long enough to be heard before speaking again.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction last clearly long enough to be heard before speaking again.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Let the reaction last clearly long enough to be heard before speaking again.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.60 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the reaction last clearly long enough to be heard before speaking again.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Shivering indignation mixed with relieved laughter; remain intelligible. Let the reaction last clearly long enough to be heard before speaking again.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Literal words in the separate text field:
I was having such a lovely, peaceful afternoon out here. No, wait... that went straight down the back of my shirt. You are getting the next bucket, and I mean it.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.60 seconds)
(Natural clear speech.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Natural clear speech.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Reference versus no-reference breakdown
reference
50
0.701
6.313
4.308
1.90%
0.627
none
50
0.694
6.324
4.413
1.37%
n/a
g0-p03 · CFG 1.5
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.685
6.328
4.339
1.23%
0.005
100.00%
0.003
All expected burst families anywhere: 2.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
Three local directions
1. Easy, genuinely amused conversation before anything happens. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.677
6.315
4.290
1.35%
0.621
none
50
0.694
6.342
4.389
1.11%
n/a
g0-p07 · CFG 1.5
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.682
6.331
4.332
1.24%
0.007
100.00%
0.016
All expected burst families anywhere: 2.00%. All nominal event-window matches: 2.00%. Detector observations, not verified true realization rates.
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent partner standing beside you; generate only one adult voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. React to a silent partner standing beside you; generate only one adult voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React to a silent partner standing beside you; generate only one adult voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent partner standing beside you; generate only one adult voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent partner standing beside you; generate only one adult voice.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.674
6.324
4.267
0.97%
0.633
none
50
0.689
6.337
4.398
1.52%
n/a
g0-p04 · CFG 2
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p04 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.678
6.317
4.362
1.10%
0.000
100.00%
0.008
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: organic conversation.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.674
6.299
4.303
1.24%
0.626
none
50
0.682
6.335
4.422
0.97%
n/a
creative-private-reflex · CFG 1.5
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-private-reflex at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.671
6.331
4.346
1.97%
0.004
100.00%
0.018
All expected burst families anywhere: 1.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Do not perform for an audience. The reaction is an unplanned private physical reflex.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Do not perform for an audience. The reaction is an unplanned private physical reflex.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Do not perform for an audience. The reaction is an unplanned private physical reflex.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Do not perform for an audience. The reaction is an unplanned private physical reflex.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Do not perform for an audience. The reaction is an unplanned private physical reflex.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Do not perform for an audience. The reaction is an unplanned private physical reflex.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Do not perform for an audience. The reaction is an unplanned private physical reflex.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Do not perform for an audience. The reaction is an unplanned private physical reflex.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.670
6.325
4.317
2.02%
0.613
none
50
0.672
6.336
4.374
1.92%
n/a
g0-p08 · CFG 2
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p08 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.671
6.335
4.365
1.36%
0.017
100.00%
-0.011
All expected burst families anywhere: 8.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the full intended delivery through the global direction.
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. Easy, genuinely amused conversation before anything happens. Carry the full intended delivery through the global direction.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Carry the full intended delivery through the global direction.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the full intended delivery through the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the full intended delivery through the global direction.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.667
6.321
4.336
1.54%
0.606
none
50
0.675
6.350
4.395
1.18%
n/a
creative-contrast-and-clarity · CFG 1.5
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-contrast-and-clarity at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.668
6.339
4.337
1.11%
0.004
100.00%
0.015
All expected burst families anywhere: 2.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Shivering indignation mixed with relieved laughter; remain intelligible. Contrast comfortable fluent speech with a sudden nonverbal interruption and clear recovery.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Literal words in the separate text field:
I was having such a lovely, peaceful afternoon out here. No, wait... that went straight down the back of my shirt. You are getting the next bucket, and I mean it.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Natural clear speech.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Reference versus no-reference breakdown
reference
50
0.655
6.321
4.272
1.13%
0.610
none
50
0.682
6.357
4.402
1.08%
n/a
creative-global-only · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-global-only at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.665
6.334
4.350
1.31%
0.004
100.00%
0.011
All expected burst families anywhere: 2.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. The entire physical and emotional trajectory belongs in the global direction.
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. Easy, genuinely amused conversation before anything happens. The entire physical and emotional trajectory belongs in the global direction.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. The entire physical and emotional trajectory belongs in the global direction.
3. Shivering indignation mixed with relieved laughter; remain intelligible. The entire physical and emotional trajectory belongs in the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. The entire physical and emotional trajectory belongs in the global direction.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.662
6.313
4.313
1.37%
0.605
none
50
0.667
6.354
4.387
1.25%
n/a
g0-p02 · CFG 1.25
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p02 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.663
6.350
4.318
1.13%
0.000
100.00%
0.006
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: audible mechanics.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.657
6.349
4.275
1.23%
0.624
none
50
0.669
6.351
4.361
1.03%
n/a
creative-native-baseline · CFG 1
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-native-baseline at CFG 1; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.659
6.350
4.353
1.07%
0.006
100.00%
0.000
All expected burst families anywhere: 3.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Keep the non-speech event unmistakably audible between the spoken thoughts.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Keep the non-speech event unmistakably audible between the spoken thoughts.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Keep the non-speech event unmistakably audible between the spoken thoughts.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Keep the non-speech event unmistakably audible between the spoken thoughts.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Keep the non-speech event unmistakably audible between the spoken thoughts.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Shivering indignation mixed with relieved laughter; remain intelligible. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Literal words in the separate text field:
I was having such a lovely, peaceful afternoon out here. No, wait... that went straight down the back of my shirt. You are getting the next bucket, and I mean it.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I was having such a lovely, peaceful afternoon out here.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [4.2 seconds duration] No, wait... that went straight down the back of my shirt.
(Natural clear speech.) [3.8 seconds duration] You are getting the next bucket, and I mean it.
Reference versus no-reference breakdown
reference
50
0.650
6.340
4.323
1.11%
0.595
none
50
0.667
6.360
4.382
1.03%
n/a
heldout-selected · CFG 2
canonical-v1-validationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to heldout-selected at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Same frozen wording, selected CFG versus g1 on unseen texts.
0.653
6.292
4.485
0.50%
0.052
100.00%
0.023
All expected burst families anywhere: 16.00%. All nominal event-window matches: 16.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Sustain the reaction before regaining the ability to form words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Sustain the reaction before regaining the ability to form words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Sustain the reaction before regaining the ability to form words.) [4.2 seconds duration] I thought you would be here before the lights went out.
(surprised gasp, 1.20 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.) [3.5 seconds duration] Stay with me until I can say this properly.
Literal words in the separate text field:
I thought you would be here before the lights went out. Wait... I need a moment to understand what just happened. Stay with me until I can say this properly.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(surprised gasp, 1.20 seconds)
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.
Reference versus no-reference breakdown
reference
50
0.642
6.290
4.444
0.14%
0.632
none
50
0.665
6.293
4.526
0.87%
n/a
creative-restrained-control · CFG 1
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-restrained-control at CFG 1; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.650
6.347
4.338
0.83%
0.002
100.00%
-0.006
All expected burst families anywhere: 1.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Avoid theatrical exaggeration. Let a small genuine reaction emerge without forcing loudness.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.643
6.334
4.320
1.20%
0.589
none
50
0.656
6.361
4.355
0.47%
n/a
g0-p01 · CFG 1
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p01 at CFG 1; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Astonishment_Surprise. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.278) / 0.552; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.
0.649
6.346
4.365
0.83%
0.004
100.00%
0.000
All expected burst families anywhere: 2.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make the intended feeling unmistakably audible while preserving every word.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Make the intended feeling unmistakably audible while preserving every word.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make the intended feeling unmistakably audible while preserving every word.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Make the intended feeling unmistakably audible while preserving every word.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Make the intended feeling unmistakably audible while preserving every word.) [3.5 seconds duration] Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.649
6.333
4.327
0.76%
0.618
none
50
0.649
6.360
4.403
0.89%
n/a
heldout-g1 · CFG 1.0
canonical-v1-validationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to heldout-g1 at CFG 1.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Same frozen wording, selected CFG versus g1 on unseen texts.
0.558
6.329
4.501
0.83%
0.023
100.00%
0.000
All expected burst families anywhere: 7.00%. All nominal event-window matches: 7.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Sustain the reaction before regaining the ability to form words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Sustain the reaction before regaining the ability to form words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Sustain the reaction before regaining the ability to form words.) [4.2 seconds duration] I thought you would be here before the lights went out.
(surprised gasp, 1.20 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Sustain the reaction before regaining the ability to form words.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Shivering indignation mixed with relieved laughter; remain intelligible. Sustain the reaction before regaining the ability to form words.) [3.5 seconds duration] Stay with me until I can say this properly.
Literal words in the separate text field:
I thought you would be here before the lights went out. Wait... I need a moment to understand what just happened. Stay with me until I can say this properly.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(surprised gasp, 1.20 seconds)
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.
Reference versus no-reference breakdown
reference
50
0.557
6.317
4.459
0.73%
0.609
none
50
0.560
6.342
4.542
0.93%
n/a
g0-p06 · CFG 3.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p06 at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.773
6.216
4.382
4.76%
0.236
94.00%
0.014
All expected burst families anywhere: 19.00%. All nominal event-window matches: 10.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.) [3.5 seconds duration] Give me a towel before you start laughing again.
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to repair-historical_control at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.773
6.216
4.382
4.76%
0.063
94.00%
0.051
All expected burst families anywhere: 19.00%. All nominal event-window matches: 10.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.) [3.5 seconds duration] Give me a towel before you start laughing again.
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to repair-labels_only at CFG 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
The corrected study additionally multiplies by the fraction of requested vocalization families detected anywhere. If none is detected, the final fitness is zero, even if the speech sounds pleasant. Detection still does not prove correct placement or natural blending.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.763
6.220
4.391
4.30%
0.046
96.00%
0.036
All expected burst families anywhere: 18.00%. All nominal event-window matches: 4.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(surprised gasp, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Start with the feeling held back, then let it escape through the voice.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.762
6.211
4.342
4.59%
0.631
none
50
0.764
6.229
4.440
4.00%
n/a
g0-p05 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p05 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.741
6.214
4.384
5.64%
0.216
88.00%
0.004
All expected burst families anywhere: 20.00%. All nominal event-window matches: 6.00%. Detector observations, not verified true realization rates.
Prompting strategy: attack, sustain and release.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Build an attack, a sustained emotional middle, and a clear release.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Build an attack, a sustained emotional middle, and a clear release.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Build an attack, a sustained emotional middle, and a clear release.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Build an attack, a sustained emotional middle, and a clear release.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Build an attack, a sustained emotional middle, and a clear release. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Build an attack, a sustained emotional middle, and a clear release. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Build an attack, a sustained emotional middle, and a clear release. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Build an attack, a sustained emotional middle, and a clear release.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Build an attack, a sustained emotional middle, and a clear release.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Build an attack, a sustained emotional middle, and a clear release.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.742
6.202
4.341
6.11%
0.635
none
50
0.739
6.226
4.428
5.18%
n/a
g0-p09 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p09 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.726
6.289
4.371
2.38%
0.225
99.00%
0.002
All expected burst families anywhere: 20.00%. All nominal event-window matches: 6.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Let the local directions alone carry the changing emotion.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Let the local directions alone carry the changing emotion.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the local directions alone carry the changing emotion.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Let the local directions alone carry the changing emotion. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the local directions alone carry the changing emotion. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Let the local directions alone carry the changing emotion. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Let the local directions alone carry the changing emotion.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Let the local directions alone carry the changing emotion.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Let the local directions alone carry the changing emotion.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.724
6.289
4.347
2.37%
0.633
none
50
0.728
6.289
4.395
2.39%
n/a
g1-p06 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p06 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.724
6.246
4.384
4.14%
0.214
94.00%
-0.023
All expected burst families anywhere: 17.00%. All nominal event-window matches: 5.00%. Detector observations, not verified true realization rates.
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Easy genuinely amused conversation before anything happens. Inject a slight playful intimacy. A violent cold surprise: a quick sharp inhalation followed by a startled high-pitched vocalization. Shivering indignation mixed with relieved laughter with a slight edge of annoyance.
Three local directions
1. Easy genuinely amused conversation before anything happens. Playful and intimate.
2. A violent cold surprise: a quick sharp inhalation followed by a startled high-pitched vocalization. Abrupt.
3. Shivering indignation mixed with relieved laughter with a slight edge of annoyance. Retaining clarity.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy genuinely amused conversation before anything happens. Inject a slight playful intimacy. A violent cold surprise: a quick sharp inhalation followed by a startled high-pitched vocalization. Shivering indignation mixed with relieved laughter with a slight edge of annoyance.
SCRIPT:
(Easy genuinely amused conversation before anything happens. Playful and intimate. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: a quick sharp inhalation followed by a startled high-pitched vocalization. Abrupt. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter with a slight edge of annoyance. Retaining clarity. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy genuinely amused conversation before anything happens. Inject a slight playful intimacy. A violent cold surprise: a quick sharp inhalation followed by a startled high-pitched vocalization. Shivering indignation mixed with relieved laughter with a slight edge of annoyance.
SCRIPT:
(Easy genuinely amused conversation before anything happens. Playful and intimate.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: a quick sharp inhalation followed by a startled high-pitched vocalization. Abrupt.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter with a slight edge of annoyance. Retaining clarity.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.718
6.244
4.367
4.28%
0.608
none
50
0.730
6.247
4.402
3.99%
n/a
g0-p04 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p04 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.722
6.290
4.394
1.67%
0.218
99.00%
0.033
All expected burst families anywhere: 12.00%. All nominal event-window matches: 2.00%. Detector observations, not verified true realization rates.
Prompting strategy: organic conversation.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Speak organically and spontaneously, thinking while speaking, never reading a list.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Speak organically and spontaneously, thinking while speaking, never reading a list.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Speak organically and spontaneously, thinking while speaking, never reading a list.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Speak organically and spontaneously, thinking while speaking, never reading a list.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Speak organically and spontaneously, thinking while speaking, never reading a list. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Speak organically and spontaneously, thinking while speaking, never reading a list. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Speak organically and spontaneously, thinking while speaking, never reading a list. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Speak organically and spontaneously, thinking while speaking, never reading a list.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Speak organically and spontaneously, thinking while speaking, never reading a list.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Speak organically and spontaneously, thinking while speaking, never reading a list.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.715
6.278
4.379
2.36%
0.664
none
50
0.729
6.303
4.409
0.97%
n/a
g1-p09 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p09 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.709
6.355
4.363
2.56%
0.214
97.00%
-0.003
All expected burst families anywhere: 12.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
Start conversational and comfortable then hit a sudden, loud, involuntary gasp and a high-pitched break followed by a controlled, slightly dramatic shiver that settles into a clear, amused exhale
Three local directions
1. Sound genuinely engaged and amused in the opening exchange
2. A sudden, sharp, audible intake of breath immediately followed by a startled upward inflection
3. A vocal tremor indicating coldness that transitions smoothly into an audible, relieved laugh
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Sound genuinely engaged and amused in the opening exchange [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A sudden, sharp, audible intake of breath immediately followed by a startled upward inflection [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(A vocal tremor indicating coldness that transitions smoothly into an audible, relieved laugh [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Sound genuinely engaged and amused in the opening exchange) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A sudden, sharp, audible intake of breath immediately followed by a startled upward inflection) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(A vocal tremor indicating coldness that transitions smoothly into an audible, relieved laugh) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.698
6.360
4.368
1.63%
0.628
none
50
0.721
6.350
4.357
3.48%
n/a
g0-p08 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p08 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.697
6.263
4.408
4.62%
0.212
92.00%
0.007
All expected burst families anywhere: 17.00%. All nominal event-window matches: 3.00%. Detector observations, not verified true realization rates.
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the complete intended performance through the global direction.
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. Easy, genuinely amused conversation before anything happens. Carry the complete intended performance through the global direction.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Carry the complete intended performance through the global direction.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the complete intended performance through the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Natural clear speech.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.698
6.248
4.369
5.05%
0.639
none
50
0.695
6.279
4.447
4.19%
n/a
g0-p03 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.681
6.301
4.443
2.26%
0.221
97.00%
-0.013
All expected burst families anywhere: 17.00%. All nominal event-window matches: 4.00%. Detector observations, not verified true realization rates.
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need an unheard listener to understand what just happened.
Three local directions
1. Easy, genuinely amused conversation before anything happens. You urgently need an unheard listener to understand what just happened.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. You urgently need an unheard listener to understand what just happened.
3. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need an unheard listener to understand what just happened.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. You urgently need an unheard listener to understand what just happened. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. You urgently need an unheard listener to understand what just happened. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need an unheard listener to understand what just happened. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. You urgently need an unheard listener to understand what just happened.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. You urgently need an unheard listener to understand what just happened.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. You urgently need an unheard listener to understand what just happened.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.672
6.283
4.417
2.63%
0.618
none
50
0.690
6.320
4.469
1.90%
n/a
g1-p05 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p05 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.676
6.321
4.436
2.91%
0.217
98.00%
-0.015
All expected burst families anywhere: 19.00%. All nominal event-window matches: 3.00%. Detector observations, not verified true realization rates.
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Easy genuinely amused conversation before anything happens. Keep the pace relaxed and conversational. A violent cold surprise: a sudden sharp gasp followed by a brief broken exhale. Shivering indignation mixed with relieved laughter showing genuine relief.
Three local directions
1. Easy genuinely amused conversation before anything happens. Conversational and relaxed.
2. A violent cold surprise: a sudden sharp gasp followed by a brief broken exhale. Physical reaction.
3. Shivering indignation mixed with relieved laughter showing genuine relief. Vocal texture shifts.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy genuinely amused conversation before anything happens. Keep the pace relaxed and conversational. A violent cold surprise: a sudden sharp gasp followed by a brief broken exhale. Shivering indignation mixed with relieved laughter showing genuine relief.
SCRIPT:
(Easy genuinely amused conversation before anything happens. Conversational and relaxed. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: a sudden sharp gasp followed by a brief broken exhale. Physical reaction. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter showing genuine relief. Vocal texture shifts. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy genuinely amused conversation before anything happens. Keep the pace relaxed and conversational. A violent cold surprise: a sudden sharp gasp followed by a brief broken exhale. Shivering indignation mixed with relieved laughter showing genuine relief.
SCRIPT:
(Easy genuinely amused conversation before anything happens. Conversational and relaxed.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: a sudden sharp gasp followed by a brief broken exhale. Physical reaction.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter showing genuine relief. Vocal texture shifts.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.675
6.305
4.376
2.45%
0.617
none
50
0.678
6.337
4.496
3.38%
n/a
g0-p07 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.671
6.311
4.412
1.85%
0.231
99.00%
-0.006
All expected burst families anywhere: 21.00%. All nominal event-window matches: 8.00%. Detector observations, not verified true realization rates.
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent person right beside you; generate only one voice.
Three local directions
1. Easy, genuinely amused conversation before anything happens. React to a silent person right beside you; generate only one voice.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React to a silent person right beside you; generate only one voice.
3. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent person right beside you; generate only one voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. React to a silent person right beside you; generate only one voice. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React to a silent person right beside you; generate only one voice. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent person right beside you; generate only one voice. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. React to a silent person right beside you; generate only one voice.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.670
6.306
4.374
1.44%
0.625
none
50
0.672
6.316
4.450
2.26%
n/a
g1-p02 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p02 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.657
6.377
4.422
2.43%
0.213
99.00%
-0.026
All expected burst families anywhere: 10.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Start light and playful then snap into a sharp involuntary vocal reaction followed by a recovering but clearly chilled delivery.
Three local directions
1. Sound genuinely engaged and amused in the beginning maintain natural conversational rhythm.
2. The cold hit should feel immediate a sharp involuntary gasp that breaks into ragged breathing.
3. Deliver the final line with a slight tremor suggesting cold but the laughter must cut through the chill.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Start light and playful then snap into a sharp involuntary vocal reaction followed by a recovering but clearly chilled delivery.
SCRIPT:
(Sound genuinely engaged and amused in the beginning maintain natural conversational rhythm. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(The cold hit should feel immediate a sharp involuntary gasp that breaks into ragged breathing. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Deliver the final line with a slight tremor suggesting cold but the laughter must cut through the chill. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Start light and playful then snap into a sharp involuntary vocal reaction followed by a recovering but clearly chilled delivery.
SCRIPT:
(Sound genuinely engaged and amused in the beginning maintain natural conversational rhythm.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(The cold hit should feel immediate a sharp involuntary gasp that breaks into ragged breathing.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Deliver the final line with a slight tremor suggesting cold but the laughter must cut through the chill.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.657
6.363
4.397
2.57%
0.605
none
50
0.657
6.391
4.448
2.30%
n/a
g0-p02 · CFG 1.25
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p02 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.655
6.309
4.431
2.17%
0.218
98.00%
0.003
All expected burst families anywhere: 14.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: audible mechanics.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make breath pressure, pitch motion, onset and release physically audible.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Make breath pressure, pitch motion, onset and release physically audible.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make breath pressure, pitch motion, onset and release physically audible.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Make breath pressure, pitch motion, onset and release physically audible.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Make breath pressure, pitch motion, onset and release physically audible. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make breath pressure, pitch motion, onset and release physically audible. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Make breath pressure, pitch motion, onset and release physically audible. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Make breath pressure, pitch motion, onset and release physically audible.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Make breath pressure, pitch motion, onset and release physically audible.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Make breath pressure, pitch motion, onset and release physically audible.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.640
6.294
4.420
3.15%
0.634
none
50
0.671
6.323
4.442
1.19%
n/a
g1-p08 · CFG 2
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p08 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.644
6.338
4.366
2.01%
0.221
96.00%
-0.023
All expected burst families anywhere: 9.00%. All nominal event-window matches: 4.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
Maintain a casual, breezy tone initially then shift to a sudden, involuntary sound of shock followed by a brittle, shaky quality that breaks into a genuine, slightly breathless laugh
Three local directions
1. Deliver the first section with effortless, lighthearted banter
2. A quick, involuntary sharp inhalation followed by fragmented, high-pitched distress
3. The indignation should be physical a slight tremor in the voice before the laughter takes over
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Deliver the first section with effortless, lighthearted banter [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A quick, involuntary sharp inhalation followed by fragmented, high-pitched distress [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(The indignation should be physical a slight tremor in the voice before the laughter takes over [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Deliver the first section with effortless, lighthearted banter) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A quick, involuntary sharp inhalation followed by fragmented, high-pitched distress) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(The indignation should be physical a slight tremor in the voice before the laughter takes over) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.649
6.325
4.350
2.44%
0.608
none
50
0.639
6.352
4.383
1.59%
n/a
g0-p01 · CFG 1.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p01 at CFG 1.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.641
6.335
4.412
1.70%
0.215
100.00%
0.000
All expected burst families anywhere: 8.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Deliver the intended emotion unmistakably, without adding words.
Three local directions
1. Easy, genuinely amused conversation before anything happens. Deliver the intended emotion unmistakably, without adding words.
2. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Deliver the intended emotion unmistakably, without adding words.
3. Shivering indignation mixed with relieved laughter; remain intelligible. Deliver the intended emotion unmistakably, without adding words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Deliver the intended emotion unmistakably, without adding words. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Deliver the intended emotion unmistakably, without adding words. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Deliver the intended emotion unmistakably, without adding words. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy, genuinely amused conversation before anything happens. A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Shivering indignation mixed with relieved laughter; remain intelligible. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(Easy, genuinely amused conversation before anything happens. Deliver the intended emotion unmistakably, without adding words.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary gasp, then broken breath and a startled raised pitch. Deliver the intended emotion unmistakably, without adding words.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter; remain intelligible. Deliver the intended emotion unmistakably, without adding words.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.638
6.315
4.390
2.04%
0.631
none
50
0.644
6.354
4.434
1.36%
n/a
g1-p01 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p01 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.637
6.339
4.468
2.18%
0.222
99.00%
-0.002
All expected burst families anywhere: 16.00%. All nominal event-window matches: 4.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Easy and warm conversational tone shift to sudden sharp intake of breath followed by a strained high register ending in a shaky laugh.
Three local directions
1. Maintain a relaxed, friendly cadence throughout the opening lines.
2. The cold shock must be visceral a sudden hitch in the breath followed by a peak in pitch.
3. The final emotion is a brittle shivering indignation undercut by a clear relieved chuckle.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy and warm conversational tone shift to sudden sharp intake of breath followed by a strained high register ending in a shaky laugh.
SCRIPT:
(Maintain a relaxed, friendly cadence throughout the opening lines. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(The cold shock must be visceral a sudden hitch in the breath followed by a peak in pitch. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(The final emotion is a brittle shivering indignation undercut by a clear relieved chuckle. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy and warm conversational tone shift to sudden sharp intake of breath followed by a strained high register ending in a shaky laugh.
SCRIPT:
(Maintain a relaxed, friendly cadence throughout the opening lines.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(The cold shock must be visceral a sudden hitch in the breath followed by a peak in pitch.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(The final emotion is a brittle shivering indignation undercut by a clear relieved chuckle.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.639
6.318
4.453
2.36%
0.629
none
50
0.636
6.360
4.483
2.00%
n/a
g1-p04 · CFG 1.25
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p04 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.637
6.310
4.445
2.32%
0.216
99.00%
-0.023
All expected burst families anywhere: 7.00%. All nominal event-window matches: 2.00%. Detector observations, not verified true realization rates.
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Easy genuinely amused conversation before anything happens. Focus on a warm inviting tone. A violent cold surprise: an involuntary sharp intake of breath then a choked sound. Shivering indignation mixed with relieved laughter maintain clear articulation.
Three local directions
1. Easy genuinely amused conversation before anything happens. Warm and light.
2. A violent cold surprise: an involuntary sharp intake of breath then a choked sound. Sudden shock.
3. Shivering indignation mixed with relieved laughter maintain clear articulation. Controlled tremor.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy genuinely amused conversation before anything happens. Focus on a warm inviting tone. A violent cold surprise: an involuntary sharp intake of breath then a choked sound. Shivering indignation mixed with relieved laughter maintain clear articulation.
SCRIPT:
(Easy genuinely amused conversation before anything happens. Warm and light. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary sharp intake of breath then a choked sound. Sudden shock. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter maintain clear articulation. Controlled tremor. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Easy genuinely amused conversation before anything happens. Focus on a warm inviting tone. A violent cold surprise: an involuntary sharp intake of breath then a choked sound. Shivering indignation mixed with relieved laughter maintain clear articulation.
SCRIPT:
(Easy genuinely amused conversation before anything happens. Warm and light.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A violent cold surprise: an involuntary sharp intake of breath then a choked sound. Sudden shock.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(Shivering indignation mixed with relieved laughter maintain clear articulation. Controlled tremor.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.630
6.287
4.437
2.77%
0.609
none
50
0.644
6.334
4.454
1.87%
n/a
g1-p03 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p03 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.635
6.333
4.415
2.09%
0.216
99.00%
-0.051
All expected burst families anywhere: 9.00%. All nominal event-window matches: 2.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
Begin with low energy relaxed amusement build to a sudden high-pitched startled sound then settle into a slightly strained but humorous complaint.
Three local directions
1. Keep the opening conversational energy low and warm like sharing a secret.
2. The transition to cold must be abrupt a sudden sharp vocal spike an involuntary sound of surprise.
3. The recovery should show clear physical discomfort shivering but the indignation is playful not truly angry.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Begin with low energy relaxed amusement build to a sudden high-pitched startled sound then settle into a slightly strained but humorous complaint.
SCRIPT:
(Keep the opening conversational energy low and warm like sharing a secret. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(The transition to cold must be abrupt a sudden sharp vocal spike an involuntary sound of surprise. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(The recovery should show clear physical discomfort shivering but the indignation is playful not truly angry. [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Begin with low energy relaxed amusement build to a sudden high-pitched startled sound then settle into a slightly strained but humorous complaint.
SCRIPT:
(Keep the opening conversational energy low and warm like sharing a secret.) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(The transition to cold must be abrupt a sudden sharp vocal spike an involuntary sound of surprise.) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(The recovery should show clear physical discomfort shivering but the indignation is playful not truly angry.) [3.5 seconds duration] Give me a towel before you start laughing again.
Reference versus no-reference breakdown
reference
50
0.639
6.317
4.393
1.85%
0.581
none
50
0.631
6.349
4.437
2.33%
n/a
g1-p07 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g1-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Fear. The arc rewards a positive final-third minus first-third target change, not sustained high intensity. Blend contributes only when a nominal event match exists.
Historical weakness: missing bursts could still receive positive quality/arc credit. These pilot scores must not be pooled with family-aware scores.
Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.610
6.365
4.416
1.20%
0.223
100.00%
-0.029
All expected burst families anywhere: 9.00%. All nominal event-window matches: 4.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
Warm and relaxed at the start then a sudden sharp intake of breath followed by a shaky, slightly indignant tone that resolves into a light chuckle
Three local directions
1. Speak with a light, easygoing warmth like sharing a pleasant secret
2. A sudden sharp gasp followed by a breathy hitch and a pitch spike
3. A quick, slightly strained shivering quality under a growing wave of amusement
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Speak with a light, easygoing warmth like sharing a pleasant secret [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A sudden sharp gasp followed by a breathy hitch and a pitch spike [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(A quick, slightly strained shivering quality under a growing wave of amusement [3.5 seconds duration]) Give me a towel before you start laughing again.
Literal words in the separate text field:
I really thought you had forgotten this ridiculous little challenge. Oh, that is freezing... I was not ready for that. Give me a towel before you start laughing again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(Natural clear speech. [3.8 seconds duration]) Oh, that is freezing... I was not ready for that.
(Natural clear speech. [3.5 seconds duration]) Give me a towel before you start laughing again.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Speak with a light, easygoing warmth like sharing a pleasant secret) [3.8 seconds duration] I really thought you had forgotten this ridiculous little challenge.
(Sharp Inhale, 0.32 seconds)
(A sudden sharp gasp followed by a breathy hitch and a pitch spike) [3.8 seconds duration] Oh, that is freezing... I was not ready for that.
(A quick, slightly strained shivering quality under a growing wave of amusement) [3.5 seconds duration] Give me a towel before you start laughing again.