Humaneness Voice Large Prompting Manuals


g0-p01	full-v2-gen0	concise trained baseline	1	6.301	4.517	0.55%	0.201	0.338	0.444	pass
g0-p02	full-v2-gen0	audible mechanics	1.25	6.327	4.524	0.44%	0.206	0.347	0.448	pass
g0-p03	full-v2-gen0	situation and intention	1.5	6.263	4.508	0.61%	0.227	0.360	0.455	pass
g0-p04	full-v2-gen0	organic conversation	2	6.121	4.497	0.45%	0.192	0.322	0.434	pass
g0-p05	full-v2-gen0	attack, sustain and release	2.5	5.797	4.331	0.42%	0.244	0.375	0.446	diagnostic
g0-p06	full-v2-gen0	containment contrast	3	5.670	4.259	0.54%	0.236	0.383	0.434	diagnostic
g0-p07	full-v2-gen0	reacting to an unheard partner	1.5	6.229	4.501	0.47%	0.196	0.334	0.439	pass
g0-p08	full-v2-gen0	GENERAL-only delivery cues	2	6.182	4.496	0.37%	0.176	0.317	0.429	pass
g0-p09	full-v2-gen0	inline-only delivery cues	2.5	6.112	4.486	0.51%	0.215	0.341	0.445	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	2.033	0.236
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.427	0.527
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.252	0.437

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 vn_VOLT_reg. 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 − 2.684) / 0.868; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.

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

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

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


0.383	5.670	4.259	0.54%	0.434	100.00%	-0.120

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Begin with the feeling held back, then let it leak through and become unmistakable.

Three local directions

1. Begin in a very quiet confidential whisper. Begin with the feeling held back, then let it leak through and become unmistakable.
2. Suddenly project an urgent loud warning to an unheard person. Begin with the feeling held back, then let it leak through and become unmistakable.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Begin with the feeling held back, then let it leak through and become unmistakable.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.379	5.664	4.263	0.34%	0.471
none	50	0.387	5.676	4.254	0.73%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	2.057	0.244
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.406	0.510
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.185	0.415

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.375	5.797	4.331	0.42%	0.446	100.00%	-0.077

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Build an audible attack, sustain the intended quality, then release it without losing the final words.

Three local directions

1. Begin in a very quiet confidential whisper. Build an audible attack, sustain the intended quality, then release it without losing the final words.
2. Suddenly project an urgent loud warning to an unheard person. Build an audible attack, sustain the intended quality, then release it without losing the final words.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Build an audible attack, sustain the intended quality, then release it without losing the final words.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.367	5.772	4.333	0.28%	0.514
none	50	0.384	5.822	4.328	0.56%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	2.008	0.227
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.310	0.448
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.288	0.449

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.360	6.263	4.508	0.61%	0.455	100.00%	-0.021

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. You urgently need one unheard listener to understand this; speak to that person, not to an audience.

Three local directions

1. Begin in a very quiet confidential whisper. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
2. Suddenly project an urgent loud warning to an unheard person. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. You urgently need one unheard listener to understand this; speak to that person, not to an audience.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.356	6.248	4.473	0.62%	0.569
none	50	0.364	6.279	4.543	0.59%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	1.938	0.206
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.247	0.424
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.316	0.458

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.347	6.327	4.524	0.44%	0.448	100.00%	0.010

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.

Three local directions

1. Begin in a very quiet confidential whisper. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
2. Suddenly project an urgent loud warning to an unheard person. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Use physical breath pressure, pitch motion, consonant attack and release rather than naming the feeling.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.357	6.323	4.490	0.13%	0.601
none	50	0.338	6.331	4.558	0.74%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	1.979	0.215
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.328	0.471
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.075	0.377

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.341	6.112	4.486	0.51%	0.445	100.00%	-0.042

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Let the local direction carry the intended delivery beside each sentence.

Three local directions

1. Begin in a very quiet confidential whisper. Let the local direction carry the intended delivery beside each sentence.
2. Suddenly project an urgent loud warning to an unheard person. Let the local direction carry the intended delivery beside each sentence.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.338	6.086	4.475	0.62%	0.549
none	50	0.343	6.138	4.497	0.40%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	1.927	0.201
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.234	0.424
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.241	0.433

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.338	6.301	4.517	0.55%	0.444	100.00%	0.000

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Make the intended feeling unmistakably audible while preserving every word.

Three local directions

1. Begin in a very quiet confidential whisper. Make the intended feeling unmistakably audible while preserving every word.
2. Suddenly project an urgent loud warning to an unheard person. Make the intended feeling unmistakably audible while preserving every word.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Make the intended feeling unmistakably audible while preserving every word.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. Make the intended feeling unmistakably audible while preserving every word.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.333	6.277	4.485	0.07%	0.591
none	50	0.343	6.326	4.549	1.03%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	1.911	0.196
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.293	0.444
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.166	0.408

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.334	6.229	4.501	0.47%	0.439	100.00%	0.004

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. React to a silent partner standing beside you; generate only one adult voice.

Three local directions

1. Begin in a very quiet confidential whisper. React to a silent partner standing beside you; generate only one adult voice.
2. Suddenly project an urgent loud warning to an unheard person. React to a silent partner standing beside you; generate only one adult voice.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. React to a silent partner standing beside you; generate only one adult voice.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.331	6.241	4.477	0.27%	0.595
none	50	0.337	6.218	4.524	0.67%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	1.898	0.192
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.320	0.462
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.009	0.356

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.322	6.121	4.497	0.45%	0.434	100.00%	-0.023

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.

Three local directions

1. Begin in a very quiet confidential whisper. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
2. Suddenly project an urgent loud warning to an unheard person. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Think while speaking, with organic micro pauses and spontaneous emphasis, not a polished reading.
SCRIPT:
(Begin in a very quiet confidential whisper. 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.
(Suddenly project an urgent loud warning to an unheard person. 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.
(Finish with a breathless but intelligible explanation of the fictional danger. 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	0.325	6.100	4.447	0.42%	0.567
none	50	0.319	6.143	4.547	0.47%	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.


Projected loudness `vn_VOLT_reg`	40.00%	high	2.684	0.868	1.846	0.176
Fear `emo_Fear`	30.00%	high	0.322	0.612	0.288	0.444
Arousal `vn_AROU_reg`	30.00%	high	2.437	1.127	2.071	0.376

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 vn_VOLT_reg. 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 − 2.684) / 0.868; 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.


0.317	6.182	4.496	0.37%	0.429	100.00%	-0.006

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. Carry the full intended delivery through the global direction.

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

1. Begin in a very quiet confidential whisper. Carry the full intended delivery through the global direction.
2. Suddenly project an urgent loud warning to an unheard person. Carry the full intended delivery through the global direction.
3. Finish with a breathless but intelligible explanation of the fictional danger. 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. Begin in a very quiet confidential whisper. Suddenly project an urgent loud warning to an unheard person. Finish with a breathless but intelligible explanation of the fictional danger. 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	0.307	6.171	4.489	0.26%	0.584
none	50	0.326	6.193	4.504	0.48%	n/a

Whisper → sudden loud warning · 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-p03 · 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-p09 · CFG 2.5

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-p07 · CFG 1.5

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