Humaneness Voice Large Prompting Manuals


g0-p01	full-v2-gen0	concise trained baseline	1	6.352	4.525	0.47%	0.422	0.564	0.556	pass
g0-p02	full-v2-gen0	audible mechanics	1.25	6.336	4.524	0.37%	0.434	0.588	0.562	pass
g0-p03	full-v2-gen0	situation and intention	1.5	6.324	4.505	0.44%	0.449	0.611	0.568	pass
g0-p04	full-v2-gen0	organic conversation	2	6.293	4.569	0.62%	0.465	0.664	0.577	pass
g0-p05	full-v2-gen0	attack, sustain and release	2.5	6.196	4.502	0.51%	0.588	0.767	0.633	pass
g0-p06	full-v2-gen0	containment contrast	3	6.160	4.514	0.49%	0.614	0.792	0.646	pass
g0-p07	full-v2-gen0	reacting to an unheard partner	1.5	6.334	4.529	0.61%	0.443	0.606	0.564	pass
g0-p08	full-v2-gen0	GENERAL-only delivery cues	2	6.295	4.529	0.48%	0.448	0.620	0.567	pass
g0-p09	full-v2-gen0	inline-only delivery cues	2.5	6.305	4.537	0.85%	0.480	0.665	0.581	pass

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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.670	0.614
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	1.746	0.890
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	4.695	0.957
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	1.553	0.792

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.792	6.160	4.514	0.49%	0.646	100.00%	0.010

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Begin with the feeling held back, then let it leak through and become unmistakable.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. Begin with the feeling held back, then let it leak through and become unmistakable.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.825	6.181	4.488	0.36%	0.639
none	50	0.759	6.139	4.540	0.61%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.592	0.588
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	1.605	0.865
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	4.516	0.940
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	1.714	0.756

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.767	6.196	4.502	0.51%	0.633	100.00%	0.017

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Build an audible attack, sustain the intended quality, then release it without losing the final words.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.783	6.217	4.490	0.56%	0.646
none	50	0.750	6.176	4.515	0.47%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.365	0.480
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	1.179	0.756
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.941	0.854
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.059	0.661

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.665	6.305	4.537	0.85%	0.581	100.00%	0.003

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Let the local direction carry the intended delivery beside each sentence.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. Let the local direction carry the intended delivery beside each sentence.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Let the local direction carry the intended delivery beside each sentence.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.676	6.311	4.512	1.10%	0.633
none	50	0.654	6.299	4.561	0.61%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.327	0.465
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	1.157	0.757
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.922	0.857
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.000	0.681

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.664	6.293	4.569	0.62%	0.577	100.00%	-0.020

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.666	6.312	4.546	0.68%	0.609
none	50	0.662	6.274	4.591	0.55%	n/a

g0-p08 · CFG 2

full-v2-gen0CANONICAL

Current ranking: EARS Steering v2

Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.

Observed head means below refer to g0-p08 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.296	0.448
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	1.035	0.710
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.659	0.782
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.162	0.628

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.620	6.295	4.529	0.48%	0.567	100.00%	0.004

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Carry the full intended delivery through the global direction.

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

1. Try to sound composed while asking a powerful fictional character to listen. Carry the full intended delivery through the global direction.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Carry the full intended delivery through the global direction.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. 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.
(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.
(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	6.286	4.476	0.43%	0.633
none	50	0.607	6.303	4.581	0.53%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.293	0.449
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	0.990	0.698
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.632	0.781
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.291	0.590

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.611	6.324	4.505	0.44%	0.568	100.00%	-0.009

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. You urgently need one unheard listener to understand this; speak to that person, not to an audience.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. Fear breaks through into vulnerable, submissive pleading for mercy. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.627	6.320	4.468	0.54%	0.621
none	50	0.596	6.329	4.541	0.33%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.299	0.443
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	0.971	0.680
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.628	0.783
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.250	0.603

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.606	6.334	4.529	0.61%	0.564	99.00%	-0.031

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. React to a silent partner standing beside you; generate only one adult voice.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. React to a silent partner standing beside you; generate only one adult voice.
2. Fear breaks through into vulnerable, submissive pleading for mercy. React to a silent partner standing beside you; generate only one adult voice.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.618	6.335	4.501	0.28%	0.599
none	50	0.594	6.332	4.558	0.94%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.281	0.434
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	0.899	0.661
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.558	0.761
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.340	0.574

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.588	6.336	4.524	0.37%	0.562	100.00%	-0.039

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. 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.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.590	6.345	4.482	0.20%	0.590
none	50	0.586	6.327	4.566	0.54%	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.


Fear `emo_Fear`	35.00%	high	0.322	0.612	0.221	0.422
Helplessness `emo_Helplessness`	30.00%	high	0.431	0.708	0.808	0.626
Vulnerability `vn_VULN_reg`	20.00%	high	2.725	0.991	3.413	0.719
Dominance / stance `vn_STNC_reg`	15.00%	low	2.574	1.139	2.383	0.561

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.

50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is emo_Fear. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

fitness = (0.50 × target_CDF + 0.25 × clip(CE/10) + 0.25 × clip(genuineness/6)) × (1 − min(1, WER))

Fixed target normalization: z = (raw score − 0.322) / 0.612; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

Passing study gates. 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.


0.564	6.352	4.525	0.47%	0.556	100.00%	0.000

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

Try to sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Make the intended feeling unmistakably audible while preserving every word.

Three local directions

1. Try to sound composed while asking a powerful fictional character to listen. Make the intended feeling unmistakably audible while preserving every word.
2. Fear breaks through into vulnerable, submissive pleading for mercy. Make the intended feeling unmistakably audible while preserving every word.
3. Remain frightened and helpless but finish the words clearly. 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 sound composed while asking a powerful fictional character to listen. Fear breaks through into vulnerable, submissive pleading for mercy. Remain frightened and helpless but finish the words clearly. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(Try to sound composed while asking a powerful fictional character to listen. 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.
(Fear breaks through into vulnerable, submissive pleading for mercy. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Remain frightened and helpless but finish the words clearly. 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.
(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	0.567	6.355	4.480	0.34%	0.629
none	50	0.560	6.350	4.570	0.61%	n/a

Begging for mercy · Humaneness 1.5 + real-speech rank 4

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-p08 · CFG 2

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-p07 · CFG 1.5

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

g0-p01 · CFG 1

Current ranking: EARS Steering v2

Exact score for this condition