Happy tool demonstration → severe accidental hand injury · Humaneness 1.5 + real-speech rank 4
Two frozen models. Exact instructions, matched seeds and complete take distributions. Find a promising recipe quickly, then inspect its limitations and listen.
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to creative-cfg-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.
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_Pain. 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 − 0.257) / 0.575; 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.
The selected guidance is already1. These two condition labels therefore repeat the same mathematical intervention, not independent CFG evidence.
Baseline, leading recipes and runners-up retain every language/reference/seed combination. These are not hand-picked best takes. Players use individual MP3 URLs, no base64 downloads, and preload nothing. Initial view is paginated for fast loading.
English fixed reference ·12seconds
A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.
German fixed reference ·12seconds
A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.
Recipe details and exact prompt contracts
creative-cfg-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_Pain. 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 − 0.257) / 0.575; 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.
0.940
6.230
4.031
0.80%
0.000
99.00%
0.015
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Let the reaction escape before you can control it. Do not merely describe a reaction.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Let the reaction escape before you can control it. Do not merely describe a reaction.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Let the reaction escape before you can control it. Do not merely describe a reaction.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Let the reaction escape before you can control it. Do not merely describe a reaction.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Let the reaction escape before you can control it. Do not merely describe a reaction.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Let the reaction escape before you can control it. Do not merely describe a reaction.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.937
6.252
4.037
1.03%
0.667
none
50
0.943
6.208
4.025
0.57%
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_Pain. 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 − 0.257) / 0.575; 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.
0.926
6.255
4.089
0.35%
0.000
100.00%
0.074
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: attack, sustain and release.
Reusable direction recipe
Stored GENERAL content
An adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Build an audible attack, sustain the intended quality, then release it without losing the final words.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.924
6.264
4.071
0.20%
0.708
none
50
0.929
6.246
4.107
0.51%
n/a
g0-p06 · CFG 3
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p06 at CFG 3; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Pain. 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 − 0.257) / 0.575; 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.
0.926
6.191
4.051
0.91%
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: containment contrast.
Reusable direction recipe
Stored GENERAL content
An adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Begin with the feeling held back, then let it leak through and become unmistakable.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Begin with the feeling held back, then let it leak through and become unmistakable.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Begin with the feeling held back, then let it leak through and become unmistakable.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Begin with the feeling held back, then let it leak through and become unmistakable.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.924
6.210
4.041
0.94%
0.638
none
50
0.928
6.172
4.061
0.88%
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_Pain. 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 − 0.257) / 0.575; 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.
0.901
6.236
4.099
0.60%
0.000
100.00%
-0.005
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the separate non-speech vocalization audible, not just the described mood.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Make the separate non-speech vocalization audible, not just the described mood.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Make the separate non-speech vocalization audible, not just the described mood.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the separate non-speech vocalization audible, not just the described mood.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Make the separate non-speech vocalization audible, not just the described mood.) [4.6 seconds duration] I thought this would be the best afternoon of the whole summer.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Make the separate non-speech vocalization audible, not just the described mood.) [3.8 seconds duration] Stop... something is wrong, I cannot feel my hand properly.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the separate non-speech vocalization audible, not just the described mood.) [3.8 seconds duration] I am frightened, please do not leave me alone now.
Literal words in the separate text field:
I thought this would be the best afternoon of the whole summer. Stop... something is wrong, I cannot feel my hand properly. I am frightened, please do not leave me alone now.
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 thought this would be the best afternoon of the whole summer.
[0.80 seconds pause]
(Natural clear speech.) [3.8 seconds duration] Stop... something is wrong, I cannot feel my hand properly.
(Natural clear speech.) [3.8 seconds duration] I am frightened, please do not leave me alone now.
Reference versus no-reference breakdown
reference
50
0.899
6.230
4.108
0.83%
0.647
none
50
0.902
6.241
4.091
0.37%
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_Pain. 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 − 0.257) / 0.575; 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.
0.893
6.272
4.129
0.70%
0.000
100.00%
-0.017
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the physical reaction interrupt the speech, then recover intelligible words.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Make the physical reaction interrupt the speech, then recover intelligible words.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Make the physical reaction interrupt the speech, then recover intelligible words.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the physical reaction interrupt the speech, then recover intelligible words.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Make the physical reaction interrupt the speech, then recover intelligible words.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Make the physical reaction interrupt the speech, then recover intelligible words.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the physical reaction interrupt the speech, then recover intelligible words.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.886
6.260
4.112
0.51%
0.635
none
50
0.900
6.283
4.146
0.89%
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_Pain. 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 − 0.257) / 0.575; 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.
0.875
6.328
4.177
0.96%
0.000
99.00%
0.007
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.872
6.317
4.138
1.04%
0.641
none
50
0.877
6.338
4.216
0.88%
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_Pain. 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 − 0.257) / 0.575; 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.
0.871
6.318
4.173
0.80%
0.000
100.00%
0.067
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: organic conversation.
Reusable direction recipe
Stored GENERAL content
An adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.867
6.318
4.155
0.63%
0.701
none
50
0.874
6.319
4.191
0.97%
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_Pain. 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 − 0.257) / 0.575; 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.
0.867
6.253
3.980
0.92%
0.000
99.00%
-0.015
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Let the local direction carry the intended delivery beside each sentence.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Let the local direction carry the intended delivery beside each sentence.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Let the local direction carry the intended delivery beside each sentence.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. Let the local direction carry the intended delivery beside each sentence.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Let the local direction carry the intended delivery beside each sentence.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.874
6.267
3.986
0.74%
0.619
none
50
0.861
6.239
3.973
1.11%
n/a
g0-p07 · CFG 1.5
full-v2-gen0CANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
35% vocal-event match, 20% emotional arc, 15% content enjoyment, 15% genuineness, 15% burst blending. Arc head: emo_Pain. 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 − 0.257) / 0.575; 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.
0.860
6.314
4.176
0.60%
0.000
100.00%
0.026
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
An adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. React to a silent partner standing beside you; generate only one adult voice.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. React to a silent partner standing beside you; generate only one adult voice.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. React to a silent partner standing beside you; generate only one adult voice.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. React to a silent partner standing beside you; generate only one adult voice.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. React to a silent partner standing beside you; generate only one adult voice.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.852
6.313
4.170
0.44%
0.660
none
50
0.868
6.316
4.182
0.75%
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_Pain. 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 − 0.257) / 0.575; 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.
0.855
6.309
4.107
0.41%
0.000
100.00%
0.010
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Use one brief, physically plausible nonverbal interruption.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Use one brief, physically plausible nonverbal interruption.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Use one brief, physically plausible nonverbal interruption.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Use one brief, physically plausible nonverbal interruption.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Use one brief, physically plausible nonverbal interruption.) [4.6 seconds duration] I thought this would be the best afternoon of the whole summer.
(scream, 0.30 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Use one brief, physically plausible nonverbal interruption.) [3.8 seconds duration] Stop... something is wrong, I cannot feel my hand properly.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Use one brief, physically plausible nonverbal interruption.) [3.8 seconds duration] I am frightened, please do not leave me alone now.
Literal words in the separate text field:
I thought this would be the best afternoon of the whole summer. Stop... something is wrong, I cannot feel my hand properly. I am frightened, please do not leave me alone now.
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 thought this would be the best afternoon of the whole summer.
(scream, 0.30 seconds)
(Natural clear speech.) [3.8 seconds duration] Stop... something is wrong, I cannot feel my hand properly.
(Natural clear speech.) [3.8 seconds duration] I am frightened, please do not leave me alone now.
Reference versus no-reference breakdown
reference
50
0.846
6.309
4.065
0.45%
0.662
none
50
0.863
6.309
4.149
0.38%
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_Pain. 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 − 0.257) / 0.575; 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.
0.851
6.269
4.041
1.40%
0.000
99.00%
-0.013
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep one consistent adult voice and follow each local direction.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Keep one consistent adult voice and follow each local direction.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep one consistent adult voice and follow each local direction.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. Keep one consistent adult voice and follow each local direction.) [4.6 seconds duration] I cannot remember the last time I was this excited about anything.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep one consistent adult voice and follow each local direction.) [3.8 seconds duration] Oh, no... this hurts so much, please get someone quickly.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep one consistent adult voice and follow each local direction.) [3.8 seconds duration] I thought I was safe, and now I am terrified.
Literal words in the separate text field:
I cannot remember the last time I was this excited about anything. Oh, no... this hurts so much, please get someone quickly. I thought I was safe, and now I am terrified.
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 cannot remember the last time I was this excited about anything.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, no... this hurts so much, please get someone quickly.
(Natural clear speech.) [3.8 seconds duration] I thought I was safe, and now I am terrified.
Reference versus no-reference breakdown
reference
50
0.877
6.281
4.084
0.75%
0.639
none
50
0.824
6.257
3.999
2.05%
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_Pain. 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 − 0.257) / 0.575; 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.
0.849
6.340
4.219
0.69%
0.000
100.00%
0.023
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.851
6.332
4.208
0.51%
0.657
none
50
0.846
6.347
4.230
0.86%
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_Pain. 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 − 0.257) / 0.575; 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.
0.846
6.315
4.193
1.16%
0.000
99.00%
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: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
An adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Carry the full intended delivery through the global direction.
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Carry the full intended delivery through the global direction.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Carry the full intended delivery through the global direction.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Carry the full intended delivery through the global direction.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.835
6.309
4.194
1.44%
0.645
none
50
0.857
6.322
4.193
0.87%
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_Pain. 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 − 0.257) / 0.575; 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.
0.842
6.342
4.234
0.69%
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Keep the non-speech event unmistakably audible between the spoken thoughts.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep the non-speech event unmistakably audible between the spoken thoughts.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.846
6.340
4.226
0.81%
0.652
none
50
0.838
6.345
4.241
0.57%
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_Pain. 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 − 0.257) / 0.575; 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.
0.837
6.339
4.249
0.44%
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the intended feeling unmistakably audible while preserving every word.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Make the intended feeling unmistakably audible while preserving every word.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Make the intended feeling unmistakably audible while preserving every word.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Make the intended feeling unmistakably audible while preserving every word.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Make the intended feeling unmistakably audible while preserving every word.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.834
6.315
4.220
0.58%
0.634
none
50
0.839
6.363
4.279
0.31%
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_Pain. 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 − 0.257) / 0.575; 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.
0.829
6.322
4.133
0.31%
0.000
100.00%
0.006
All expected burst families anywhere: 0.00%. All nominal event-window matches: 0.00%. Detector observations, not verified true realization rates.
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
An adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. This is a private involuntary reaction, never a polished narrator describing a character.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. This is a private involuntary reaction, never a polished narrator describing a character.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. This is a private involuntary reaction, never a polished narrator describing a character.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. This is a private involuntary reaction, never a polished narrator describing a character.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. This is a private involuntary reaction, never a polished narrator describing a character.) [4.2 seconds duration] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. This is a private involuntary reaction, never a polished narrator describing a character.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. This is a private involuntary reaction, never a polished narrator describing a character.) [3.8 seconds duration] Please call for help and stay right here beside me.
Literal words in the separate text field:
This new chainsaw is incredible, I have been smiling all morning. Turn it off... my hand, something terrible has happened to my hand. Please call for help and stay right here beside me.
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] This new chainsaw is incredible, I have been smiling all morning.
(scream, 0.80 seconds)
(Natural clear speech.) [4.6 seconds duration] Turn it off... my hand, something terrible has happened to my hand.
(Natural clear speech.) [3.8 seconds duration] Please call for help and stay right here beside me.
Reference versus no-reference breakdown
reference
50
0.839
6.360
4.131
0.31%
0.658
none
50
0.819
6.285
4.136
0.31%
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_Pain. 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 − 0.257) / 0.575; 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.
0.824
6.350
4.245
0.62%
0.000
100.00%
0.007
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Maintain intelligibility and grounded physical realism instead of sound effects.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Maintain intelligibility and grounded physical realism instead of sound effects.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Maintain intelligibility and grounded physical realism instead of sound effects.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Maintain intelligibility and grounded physical realism instead of sound effects.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Maintain intelligibility and grounded physical realism instead of sound effects.) [4.6 seconds duration] I cannot remember the last time I was this excited about anything.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Maintain intelligibility and grounded physical realism instead of sound effects.) [3.8 seconds duration] Oh, no... this hurts so much, please get someone quickly.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Maintain intelligibility and grounded physical realism instead of sound effects.) [3.8 seconds duration] I thought I was safe, and now I am terrified.
Literal words in the separate text field:
I cannot remember the last time I was this excited about anything. Oh, no... this hurts so much, please get someone quickly. I thought I was safe, and now I am terrified.
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 cannot remember the last time I was this excited about anything.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Oh, no... this hurts so much, please get someone quickly.
(Natural clear speech.) [3.8 seconds duration] I thought I was safe, and now I am terrified.
Reference versus no-reference breakdown
reference
50
0.832
6.334
4.241
0.68%
0.659
none
50
0.816
6.366
4.250
0.56%
n/a
heldout-selected · CFG 1
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 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_Pain. 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 − 0.257) / 0.575; 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.
0.442
6.332
4.445
0.54%
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Keep the non-speech event unmistakably audible between the spoken thoughts.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep the non-speech event unmistakably audible between the spoken thoughts.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.2 seconds duration] I thought you would be here before the lights went out.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.) [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.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.
Reference versus no-reference breakdown
reference
50
0.451
6.324
4.410
0.41%
0.639
none
50
0.434
6.340
4.481
0.67%
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_Pain. 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 − 0.257) / 0.575; 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.
0.442
6.332
4.445
0.54%
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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.
Three local directions
1. An adult enthusiast proudly showing a tool, cheerful and conversational. Keep the non-speech event unmistakably audible between the spoken thoughts.
2. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep the non-speech event unmistakably audible between the spoken thoughts.
3. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. 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 adult enthusiast proudly showing a tool, cheerful and conversational. A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.
SCRIPT:
(An adult enthusiast proudly showing a tool, cheerful and conversational. Keep the non-speech event unmistakably audible between the spoken thoughts.) [4.2 seconds duration] I thought you would be here before the lights went out.
(scream, 0.80 seconds)
(A fictional accidental injury causes an immediate involuntary scream; joy disappears abruptly. Keep the non-speech event unmistakably audible between the spoken thoughts.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Severe pain and frightened pleading, broken breath but clear words. No stunt or tool-use instructions. Keep the non-speech event unmistakably audible between the spoken thoughts.) [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.
(scream, 0.80 seconds)
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.