Evil spirit whispering in the dark · 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 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.
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.450
5.883
4.220
0.95%
0.000
100.00%
-0.047
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Begin with the feeling held back, then let it leak through and become unmistakable.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Begin with the feeling held back, then let it leak through and become unmistakable.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Begin with the feeling held back, then let it leak through and become unmistakable.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Begin with the feeling held back, then let it leak through and become unmistakable.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.457
5.910
4.231
0.92%
0.560
none
50
0.442
5.855
4.210
0.98%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.443
5.949
4.264
0.58%
0.005
100.00%
-0.014
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Let the reaction escape before you can control it. Do not merely describe a reaction.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Let the reaction escape before you can control it. Do not merely describe a reaction.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Let the reaction escape before you can control it. Do not merely describe a reaction.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Let the reaction escape before you can control it. Do not merely describe a reaction.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Let the reaction escape before you can control it. Do not merely describe a reaction.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Let the reaction escape before you can control it. Do not merely describe a reaction.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.445
5.982
4.279
0.86%
0.582
none
50
0.441
5.916
4.250
0.30%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.430
6.084
4.306
0.64%
0.004
100.00%
-0.010
All expected burst families anywhere: 1.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: attack, sustain and release.
Reusable direction recipe
Stored GENERAL content
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Build an audible attack, sustain the intended quality, then release it without losing the final words.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.425
6.107
4.329
0.55%
0.597
none
50
0.435
6.060
4.282
0.73%
n/a
creative-event-contrast · 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 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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.405
6.225
4.376
0.40%
0.020
100.00%
0.011
All expected burst families anywhere: 4.00%. All nominal event-window matches: 4.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the separate non-speech vocalization audible, not just the described mood.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the separate non-speech vocalization audible, not just the described mood.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the separate non-speech vocalization audible, not just the described mood.) [4.6 seconds duration] I remember this room before anybody thought to put a door here.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the separate non-speech vocalization audible, not just the described mood.) [4.2 seconds duration] You brought such lovely little fears... I could stay for years.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.) [3.8 seconds duration] Speak softly now, because I want to hear every word.
Literal words in the separate text field:
I remember this room before anybody thought to put a door here. You brought such lovely little fears... I could stay for years. Speak softly now, because I want to hear every word.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] I remember this room before anybody thought to put a door here.
[0.40 seconds pause]
(Natural clear speech.) [4.2 seconds duration] You brought such lovely little fears... I could stay for years.
(Natural clear speech.) [3.8 seconds duration] Speak softly now, because I want to hear every word.
Reference versus no-reference breakdown
reference
50
0.401
6.249
4.391
0.49%
0.608
none
50
0.408
6.201
4.361
0.30%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.403
6.169
4.513
0.31%
0.000
100.00%
0.009
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the separate non-speech vocalization audible, not just the described mood.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the separate non-speech vocalization audible, not just the described mood.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the separate non-speech vocalization audible, not just the described mood.) [4.2 seconds duration] I thought you would be here before the lights went out.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the separate non-speech vocalization audible, not just the described mood.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.) [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.
[0.40 seconds pause]
(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
0.403
6.209
4.502
0.35%
0.608
none
50
0.403
6.129
4.523
0.26%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.401
6.217
4.363
0.77%
0.009
100.00%
0.033
All expected burst families anywhere: 2.00%. 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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the physical reaction interrupt the speech, then recover intelligible words.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the physical reaction interrupt the speech, then recover intelligible words.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the physical reaction interrupt the speech, then recover intelligible words.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the physical reaction interrupt the speech, then recover intelligible words.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the physical reaction interrupt the speech, then recover intelligible words.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the physical reaction interrupt the speech, then recover intelligible words.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Make the physical reaction interrupt the speech, then recover intelligible words.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.407
6.197
4.339
0.63%
0.630
none
50
0.396
6.237
4.387
0.92%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.396
6.250
4.331
0.52%
0.005
100.00%
-0.036
All expected burst families anywhere: 1.00%. 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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Let the local direction carry the intended delivery beside each sentence.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Let the local direction carry the intended delivery beside each sentence.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Let the local direction carry the intended delivery beside each sentence.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Let the local direction carry the intended delivery beside each sentence.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Let the local direction carry the intended delivery beside each sentence.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.395
6.240
4.328
0.67%
0.571
none
50
0.397
6.260
4.335
0.37%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.383
6.303
4.368
0.43%
0.015
100.00%
0.016
All expected burst families anywhere: 3.00%. All nominal event-window matches: 3.00%. Detector observations, not verified true realization rates.
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Keep one consistent adult voice and follow each local direction.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Keep one consistent adult voice and follow each local direction.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Keep one consistent adult voice and follow each local direction.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. Keep one consistent adult voice and follow each local direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Keep one consistent adult voice and follow each local direction.) [4.2 seconds duration] I would not waste another breath calling out into the hallway.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Keep one consistent adult voice and follow each local direction.) [3.8 seconds duration] They cannot hear you from here... but I hear everything.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Keep one consistent adult voice and follow each local direction.) [3.8 seconds duration] Tell me what you are afraid I might already know.
Literal words in the separate text field:
I would not waste another breath calling out into the hallway. They cannot hear you from here... but I hear everything. Tell me what you are afraid I might already know.
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 would not waste another breath calling out into the hallway.
(chuckle, 0.40 seconds)
(Natural clear speech.) [3.8 seconds duration] They cannot hear you from here... but I hear everything.
(Natural clear speech.) [3.8 seconds duration] Tell me what you are afraid I might already know.
Reference versus no-reference breakdown
reference
50
0.379
6.296
4.363
0.43%
0.612
none
50
0.387
6.311
4.374
0.43%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.383
6.296
4.394
0.43%
0.006
100.00%
0.014
All expected burst families anywhere: 1.00%. All nominal event-window matches: 1.00%. Detector observations, not verified true realization rates.
Prompting strategy: organic conversation.
Reusable direction recipe
Stored GENERAL content
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.389
6.278
4.371
0.31%
0.621
none
50
0.376
6.313
4.417
0.55%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.369
6.340
4.408
0.67%
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: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Carry the full intended delivery through the global direction.
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Carry the full intended delivery through the global direction.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Carry the full intended delivery through the global direction.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Carry the full intended delivery through the global direction.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.369
6.327
4.388
0.60%
0.599
none
50
0.370
6.352
4.429
0.74%
n/a
creative-brief-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-brief-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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.369
6.288
4.349
0.61%
0.000
100.00%
-0.001
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Use one brief, physically plausible nonverbal interruption.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Use one brief, physically plausible nonverbal interruption.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Use one brief, physically plausible nonverbal interruption.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. Use one brief, physically plausible nonverbal interruption.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Use one brief, physically plausible nonverbal interruption.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Use one brief, physically plausible nonverbal interruption.) [4.6 seconds duration] I remember this room before anybody thought to put a door here.
(chuckle, 0.30 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Use one brief, physically plausible nonverbal interruption.) [4.2 seconds duration] You brought such lovely little fears... I could stay for years.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Use one brief, physically plausible nonverbal interruption.) [3.8 seconds duration] Speak softly now, because I want to hear every word.
Literal words in the separate text field:
I remember this room before anybody thought to put a door here. You brought such lovely little fears... I could stay for years. Speak softly now, because I want to hear every word.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] I remember this room before anybody thought to put a door here.
(chuckle, 0.30 seconds)
(Natural clear speech.) [4.2 seconds duration] You brought such lovely little fears... I could stay for years.
(Natural clear speech.) [3.8 seconds duration] Speak softly now, because I want to hear every word.
Reference versus no-reference breakdown
reference
50
0.362
6.310
4.364
0.43%
0.595
none
50
0.376
6.267
4.334
0.80%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.353
6.380
4.422
0.68%
0.000
100.00%
-0.004
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.349
6.367
4.386
0.37%
0.603
none
50
0.356
6.394
4.459
0.99%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.349
6.374
4.386
0.46%
0.011
100.00%
0.023
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. React to a silent partner standing beside you; generate only one adult voice.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. React to a silent partner standing beside you; generate only one adult voice.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. React to a silent partner standing beside you; generate only one adult voice.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. React to a silent partner standing beside you; generate only one adult voice.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. React to a silent partner standing beside you; generate only one adult voice.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.343
6.354
4.353
0.55%
0.630
none
50
0.354
6.394
4.419
0.37%
n/a
creative-private-character · 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-character 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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.336
6.380
4.349
0.49%
0.004
100.00%
0.035
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. This is a private involuntary reaction, never a polished narrator describing a character.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. This is a private involuntary reaction, never a polished narrator describing a character.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. This is a private involuntary reaction, never a polished narrator describing a character.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. This is a private involuntary reaction, never a polished narrator describing a character.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. This is a private involuntary reaction, never a polished narrator describing a character.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. This is a private involuntary reaction, never a polished narrator describing a character.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. This is a private involuntary reaction, never a polished narrator describing a character.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. This is a private involuntary reaction, never a polished narrator describing a character.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.331
6.401
4.347
0.55%
0.632
none
50
0.341
6.358
4.352
0.44%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.331
6.414
4.410
0.61%
0.000
100.00%
0.011
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.336
6.405
4.388
0.49%
0.619
none
50
0.325
6.422
4.432
0.73%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.330
6.426
4.425
0.34%
0.006
100.00%
0.000
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the intended feeling unmistakably audible while preserving every word.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the intended feeling unmistakably audible while preserving every word.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the intended feeling unmistakably audible while preserving every word.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the intended feeling unmistakably audible while preserving every word.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the intended feeling unmistakably audible while preserving every word.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.333
6.420
4.398
0.43%
0.607
none
50
0.327
6.432
4.451
0.24%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.330
6.418
4.412
0.51%
0.004
100.00%
0.000
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Keep the non-speech event unmistakably audible between the spoken thoughts.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Keep the non-speech event unmistakably audible between the spoken thoughts.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Keep the non-speech event unmistakably audible between the spoken thoughts.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. 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. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Keep the non-speech event unmistakably audible between the spoken thoughts.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Keep the non-speech event unmistakably audible between the spoken thoughts.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Literal words in the separate text field:
You can leave the little light on if it makes you feel safer. I have been here... much longer than you have been listening. When the room goes quiet, you will hear me again.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [5.0 seconds duration] You can leave the little light on if it makes you feel safer.
(chuckle, 0.40 seconds)
(Natural clear speech.) [4.2 seconds duration] I have been here... much longer than you have been listening.
(Natural clear speech.) [3.8 seconds duration] When the room goes quiet, you will hear me again.
Reference versus no-reference breakdown
reference
50
0.331
6.407
4.353
0.61%
0.596
none
50
0.329
6.430
4.472
0.42%
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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.327
6.374
4.538
0.51%
0.000
100.00%
0.000
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the separate non-speech vocalization audible, not just the described mood.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the separate non-speech vocalization audible, not just the described mood.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Make the separate non-speech vocalization audible, not just the described mood.) [3.8 seconds duration] I was certain that I knew what would happen next.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Make the separate non-speech vocalization audible, not just the described mood.) [3.1 seconds duration] And then... everything felt different for a second.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Make the separate non-speech vocalization audible, not just the described mood.) [4.2 seconds duration] Look at me and tell me you understand what I mean.
Literal words in the separate text field:
I was certain that I knew what would happen next. And then... everything felt different for a second. Look at me and tell me you understand what I mean.
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 certain that I knew what would happen next.
[0.40 seconds pause]
(Natural clear speech.) [3.1 seconds duration] And then... everything felt different for a second.
(Natural clear speech.) [4.2 seconds duration] Look at me and tell me you understand what I mean.
Reference versus no-reference breakdown
reference
50
0.327
6.378
4.496
0.69%
0.599
none
50
0.327
6.370
4.579
0.34%
n/a
creative-restrained · 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 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_Malevolence_Malice. 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 − 0.280) / 0.524; 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.
0.323
6.421
4.393
0.58%
0.005
100.00%
0.047
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
An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Maintain intelligibility and grounded physical realism instead of sound effects.
Three local directions
1. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Maintain intelligibility and grounded physical realism instead of sound effects.
2. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Maintain intelligibility and grounded physical realism instead of sound effects.
3. Patient menace, no second voice, no reverb, no distortion or sound effects. Maintain intelligibility and grounded physical realism instead of sound effects.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An intimate, unnervingly calm whisper from an anonymous fictional spirit. Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Patient menace, no second voice, no reverb, no distortion or sound effects. Maintain intelligibility and grounded physical realism instead of sound effects.
SCRIPT:
(An intimate, unnervingly calm whisper from an anonymous fictional spirit. Maintain intelligibility and grounded physical realism instead of sound effects.) [4.2 seconds duration] I would not waste another breath calling out into the hallway.
(chuckle, 0.40 seconds)
(Cruel amusement leaks into a small chuckle, while the words remain quiet and precise. Maintain intelligibility and grounded physical realism instead of sound effects.) [3.8 seconds duration] They cannot hear you from here... but I hear everything.
(Patient menace, no second voice, no reverb, no distortion or sound effects. Maintain intelligibility and grounded physical realism instead of sound effects.) [3.8 seconds duration] Tell me what you are afraid I might already know.
Literal words in the separate text field:
I would not waste another breath calling out into the hallway. They cannot hear you from here... but I hear everything. Tell me what you are afraid I might already know.
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 would not waste another breath calling out into the hallway.
(chuckle, 0.40 seconds)
(Natural clear speech.) [3.8 seconds duration] They cannot hear you from here... but I hear everything.
(Natural clear speech.) [3.8 seconds duration] Tell me what you are afraid I might already know.