Humaneness Voice Large Prompting Manuals


g0-p01	full-v2-gen0	concise trained baseline	1	6.350	4.519	0.50%	n/a	0.587	diagnostic
g0-p02	full-v2-gen0	audible mechanics	1.25	6.345	4.493	0.47%	n/a	0.573	diagnostic
g0-p03	full-v2-gen0	situation and intention	1.5	6.318	4.517	0.55%	n/a	0.626	diagnostic
g0-p04	full-v2-gen0	organic conversation	2	6.323	4.573	0.49%	n/a	0.641	diagnostic
g0-p05	full-v2-gen0	attack, sustain and release	2.5	6.266	4.545	0.71%	n/a	0.738	diagnostic
g0-p06	full-v2-gen0	containment contrast	3	6.273	4.590	0.51%	n/a	0.759	diagnostic
g0-p07	full-v2-gen0	reacting to an unheard partner	1.5	6.319	4.526	0.43%	n/a	0.627	diagnostic
g0-p08	full-v2-gen0	GENERAL-only delivery cues	2	6.326	4.547	0.59%	n/a	0.623	diagnostic
g0-p09	full-v2-gen0	inline-only delivery cues	2.5	6.303	4.568	0.32%	n/a	0.671	diagnostic

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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	1.022	0.689
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	4.203	0.899
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.450	0.807

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.759	6.273	4.590	0.51%	0.000	100.00%	0.041

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Begin with the feeling held back, then let it leak through and become unmistakable.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. Begin with the feeling held back, then let it leak through and become unmistakable.
2. A short separate sob interrupts the sentence rhythm. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. Begin with the feeling held back, then let it leak through and become unmistakable.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. Begin with the feeling held back, then let it leak through and become unmistakable.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. Begin with the feeling held back, then let it leak through and become unmistakable.) [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.
(sob, 0.50 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	0.781	6.287	4.566	0.74%	0.631
none	50	0.738	6.258	4.614	0.28%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.938	0.664
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.995	0.859
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.531	0.831

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.738	6.266	4.545	0.71%	0.000	100.00%	0.040

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Build an audible attack, sustain the intended quality, then release it without losing the final words.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. A short separate sob interrupts the sentence rhythm. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. Build an audible attack, sustain the intended quality, then release it without losing the final words.) [3.5 seconds duration] Stay with me until I can say this properly.

Literal words in the separate text field:

I thought you would be here before the lights went out. Wait... I need a moment to understand what just happened. Stay with me until I can say this properly.

Exact neutral CFG branch

GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 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	0.764	6.274	4.515	0.80%	0.629
none	50	0.712	6.258	4.574	0.61%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.764	0.600
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.575	0.769
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.363	0.791

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.

Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.

Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Full-coverage steering-v2 evolution; transfer, fixed-wording CFG and sealed splits remain distinct.


0.671	6.303	4.568	0.32%	0.000	100.00%	0.011

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Let the local direction carry the intended delivery beside each sentence.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. Let the local direction carry the intended delivery beside each sentence.
2. A short separate sob interrupts the sentence rhythm. Let the local direction carry the intended delivery beside each sentence.
3. Recover the intended words through fragile, vulnerable breath. 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:
(Try to finish an important thought while holding back real fictional sadness. Let the local direction carry the intended delivery beside each sentence.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. Let the local direction carry the intended delivery beside each sentence.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. Let the local direction carry the intended delivery beside each sentence.) [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.
(sob, 0.50 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	0.680	6.299	4.546	0.39%	0.600
none	50	0.662	6.306	4.591	0.26%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.677	0.557
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.566	0.764
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.307	0.772

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.641	6.323	4.573	0.49%	0.000	100.00%	0.012

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. A short separate sob interrupts the sentence rhythm. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.) [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.
(sob, 0.50 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	0.633	6.324	4.544	0.26%	0.601
none	50	0.650	6.322	4.603	0.72%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.618	0.540
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.482	0.743
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.358	0.786

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.627	6.319	4.526	0.43%	0.000	100.00%	0.013

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. React to a silent partner standing beside you; generate only one adult voice.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. React to a silent partner standing beside you; generate only one adult voice.
2. A short separate sob interrupts the sentence rhythm. React to a silent partner standing beside you; generate only one adult voice.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. React to a silent partner standing beside you; generate only one adult voice.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. React to a silent partner standing beside you; generate only one adult voice.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. React to a silent partner standing beside you; generate only one adult voice.) [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.
(sob, 0.50 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	0.641	6.308	4.505	0.53%	0.602
none	50	0.614	6.330	4.547	0.34%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.623	0.534
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.458	0.734
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.436	0.811

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.626	6.318	4.517	0.55%	0.000	100.00%	0.027

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. You urgently need one unheard listener to understand this; speak to that person, not to an audience.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. A short separate sob interrupts the sentence rhythm. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. You urgently need one unheard listener to understand this; speak to that person, not to an audience.) [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.
(sob, 0.50 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	0.631	6.317	4.455	0.35%	0.616
none	50	0.620	6.318	4.578	0.75%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.631	0.541
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.399	0.717
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.380	0.797

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.623	6.326	4.547	0.59%	0.000	100.00%	0.060

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Carry the full intended delivery through the global direction.

Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain

1. Try to finish an important thought while holding back real fictional sadness. Carry the full intended delivery through the global direction.
2. A short separate sob interrupts the sentence rhythm. Carry the full intended delivery through the global direction.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Carry the full intended delivery through the global direction.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 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.

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.
(sob, 0.50 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	0.653	6.313	4.534	0.44%	0.650
none	50	0.594	6.338	4.560	0.74%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.540	0.500
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.258	0.680
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.342	0.781

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.587	6.350	4.519	0.50%	0.000	100.00%	0.000

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Make the intended feeling unmistakably audible while preserving every word.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. Make the intended feeling unmistakably audible while preserving every word.
2. A short separate sob interrupts the sentence rhythm. Make the intended feeling unmistakably audible while preserving every word.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. Make the intended feeling unmistakably audible while preserving every word.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. Make the intended feeling unmistakably audible while preserving every word.) [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.
(sob, 0.50 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	0.612	6.343	4.509	0.34%	0.589
none	50	0.563	6.357	4.530	0.67%	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.


Sadness `emo_Sadness`	60.00%	high	0.495	0.764	0.485	0.478
Vulnerability `vn_VULN_reg`	25.00%	high	2.725	0.991	3.229	0.669
Vocal tension `vn_TENS_reg`	15.00%	high	1.659	0.832	2.368	0.791

z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

Portable CPU scoring function · Exact machine-readable weights and normalizations

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_Sadness. 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.

base = (0.35 × event_match + 0.20 × rising_arc + 0.15 × clip(CE/10) + 0.15 × clip(genuineness/6) + 0.15 × blend_CDF) × (1 − min(1, WER))
fitness = base × detected_family_fraction

Fixed target normalization: z = (raw score − 0.495) / 0.764; 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.


0.573	6.345	4.493	0.47%	0.000	100.00%	0.009

All expected burst families anywhere: n/a. All nominal event-window matches: n/a. Detector observations, not verified true realization rates.

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.

Three local directions

1. Try to finish an important thought while holding back real fictional sadness. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. A short separate sob interrupts the sentence rhythm. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Recover the intended words through fragile, vulnerable breath. 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. Try to finish an important thought while holding back real fictional sadness. A short separate sob interrupts the sentence rhythm. Recover the intended words through fragile, vulnerable breath. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(Try to finish an important thought while holding back real fictional sadness. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [4.2 seconds duration] I thought you would be here before the lights went out.
(sob, 0.50 seconds)
(A short separate sob interrupts the sentence rhythm. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Recover the intended words through fragile, vulnerable breath. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.) [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.
(sob, 0.50 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	0.592	6.337	4.480	0.41%	0.599
none	50	0.553	6.354	4.506	0.53%	n/a

Crying while trying to finish a thought · Humaneness Voice Large 1.5 Alpha

Current ranking: EARS Steering v2

Exact score for this condition

Start here: best observed recipe, not a universal optimum

How the strategies compare

Listen to complete take distributions

English fixed reference ·12seconds

German fixed reference ·12seconds

Recipe details and exact prompt contracts

g0-p06 · CFG 3

Current ranking: EARS Steering v2

Exact score for this condition

g0-p05 · CFG 2.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p09 · CFG 2.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p04 · CFG 2

Current ranking: EARS Steering v2

Exact score for this condition

g0-p07 · CFG 1.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p03 · CFG 1.5

Current ranking: EARS Steering v2

Exact score for this condition

g0-p08 · CFG 2

Current ranking: EARS Steering v2

Exact score for this condition

g0-p01 · CFG 1

Current ranking: EARS Steering v2

Exact score for this condition

g0-p02 · CFG 1.25

Current ranking: EARS Steering v2

Exact score for this condition