Two frozen models. Exact instructions, matched seeds and complete take distributions. Find a promising recipe quickly, then inspect its limitations and listen.
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to repair-canonical_both at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Re-scored, not regenerated: the original raw EARS predictions were reused on CPU. Old fitness, audio, prompts and scientific gates are retained. Old held-out sets tested only the previously selected wording; a newly re-ranked search winner does not inherit that validation. No universal best weighting or human-preference optimum is claimed.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Start here: best observed recipe, not a universal optimum
The current starting point uses EARS Steering v2 across the available canonical search conditions, with unchanged qualification gates taking priority. Historical candidates are used only where no canonical search exists. Held-out results below refer to the earlier selected wording and are not automatically inherited by a newly re-ranked winner.
No recipe satisfies every realization/quality gate. The highlighted recipe is the best diagnostic candidate, not a reliable recommendation.
How the strategies compare
Click any column header to sort. Do not compare scalar fitness across different missions or between legacy and event-aware scoring contracts.
Transfer to unseen sentences — earlier selection, now re-scored
The previously selected wording was repeated on five new lexically neutral bilingual scripts, at selected guidance and guidance 1. These outputs are now also re-scored with Steering v2. They do not validate other newly favored wording and are not a human listening validation.
Baseline, leading recipes and runners-up retain every language/reference/seed combination. These are not hand-picked best takes. Players use individual MP3 URLs, no base64 downloads, and preload nothing. Initial view is paginated for fast loading.
English fixed reference ·12seconds
A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.
German fixed reference ·12seconds
A clean fixed codec-reconstructed reference, a different utterance from every search scene. The same recording is used by both models.
Recipe details and exact prompt contracts
repair-canonical_both · CFG 2
canonical-v1-explorationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to repair-canonical_both at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; target_CDF = Φ(z). This is a percentile-like training-distribution transform, not an emotion probability. CE and genuineness are clipped to [0, 1] after division. For burst tasks, the same target transform is used for the two arc endpoints; blend has its own frozen normalization.
WER example: WER 5% multiplies the score by 0.95; WER 20% by 0.80; WER ≥ 100% by zero. Reported raw WER remains uncapped. The displayed fitness is the mean of per-take products, not the product of the displayed means.
Separate qualification gates: at least 90% valid, terminated takes; per-take WER ≤ 15% to count as valid; candidate mean WER ≤ 10%; paired reference cosine must not fall by more than 0.05 over all 50 reference comparisons. Corrected burst conditions additionally need every requested family detected in at least 60% of takes. Passing a gate is not a weighted score bonus.
Does not pass all study gates; diagnostic only. N=100 across both languages, both reference arms and all five seeds. Predeclared intervention; see exact genome and changed fields.
0.579
6.097
4.427
0.93%
0.621
100.00%
-0.050
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
Three local directions
1. A warm, low register whisper, inviting the listener in with gentle intimacy.
2. The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.
3. Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy.) [3.8 seconds duration] We should probably leave before they notice we are missing.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.) [3.5 seconds duration] Although... I am quite happy hiding here with you.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.) [5.0 seconds duration] Give me one more quiet minute, and then I promise we will go.
Literal words in the separate text field:
We should probably leave before they notice we are missing. Although... I am quite happy hiding here with you. Give me one more quiet minute, and then I promise we will go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.8 seconds duration] We should probably leave before they notice we are missing.
(Natural clear speech.) [3.5 seconds duration] Although... I am quite happy hiding here with you.
(Natural clear speech.) [5.0 seconds duration] Give me one more quiet minute, and then I promise we will go.
Reference versus no-reference breakdown
reference
50
0.577
6.100
4.418
1.06%
0.553
none
50
0.582
6.094
4.436
0.80%
n/a
heldout-selected · CFG 2
canonical-v1-validationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to heldout-selected at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Same frozen wording, selected CFG versus g1 on unseen texts.
0.543
6.114
4.512
0.60%
0.609
100.00%
-0.033
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
Three local directions
1. A warm, low register whisper, inviting the listener in with gentle intimacy.
2. The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.
3. Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy.) [3.5 seconds duration] I remember how ordinary everything looked a moment ago.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.) [3.8 seconds duration] Now... I am not quite sure where I should begin.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.) [4.6 seconds duration] Give me a little time and I will explain the whole thing.
Literal words in the separate text field:
I remember how ordinary everything looked a moment ago. Now... I am not quite sure where I should begin. Give me a little time and I will explain the whole thing.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.5 seconds duration] I remember how ordinary everything looked a moment ago.
(Natural clear speech.) [3.8 seconds duration] Now... I am not quite sure where I should begin.
(Natural clear speech.) [4.6 seconds duration] Give me a little time and I will explain the whole thing.
Reference versus no-reference breakdown
reference
50
0.567
6.076
4.492
0.40%
0.557
none
50
0.520
6.153
4.531
0.80%
n/a
heldout-g1 · CFG 1.0
canonical-v1-validationCANONICAL
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to heldout-g1 at CFG 1.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Same frozen wording, selected CFG versus g1 on unseen texts.
0.320
6.367
4.537
0.24%
0.507
100.00%
0.000
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
Three local directions
1. A warm, low register whisper, inviting the listener in with gentle intimacy.
2. The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.
3. Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy.) [3.5 seconds duration] I remember how ordinary everything looked a moment ago.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.) [3.8 seconds duration] Now... I am not quite sure where I should begin.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.) [4.6 seconds duration] Give me a little time and I will explain the whole thing.
Literal words in the separate text field:
I remember how ordinary everything looked a moment ago. Now... I am not quite sure where I should begin. Give me a little time and I will explain the whole thing.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [3.5 seconds duration] I remember how ordinary everything looked a moment ago.
(Natural clear speech.) [3.8 seconds duration] Now... I am not quite sure where I should begin.
(Natural clear speech.) [4.6 seconds duration] Give me a little time and I will explain the whole thing.
Reference versus no-reference breakdown
reference
50
0.315
6.358
4.492
0.27%
0.590
none
50
0.325
6.376
4.581
0.20%
n/a
g1-p03 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p03 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.676
5.527
4.285
3.22%
0.634
97.00%
-0.103
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
Remain a near-silent whisper conveying deep emotional vulnerability The whisper should feel like a fragile secret shared in darkness
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. Remain a near-silent whisper infuse the sound with deep vulnerability like a fragile disclosure
2. Continue whispering with an emotionally present quality letting the feeling guide the breath control
3. Remain a near-silent whisper maintaining intelligibility while emphasizing the emotional weight over clarity
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Remain a near-silent whisper conveying deep emotional vulnerability The whisper should feel like a fragile secret shared in darkness
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Remain a near-silent whisper conveying deep emotional vulnerability The whisper should feel like a fragile secret shared in darkness
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Natural clear speech.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Natural clear speech.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.670
5.495
4.262
3.28%
0.501
none
50
0.681
5.560
4.307
3.16%
n/a
g0-p06 · CFG 3.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 3.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.643
5.550
4.228
1.91%
0.625
98.00%
-0.201
Prompting strategy: containment contrast.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Start with the feeling held back, then let it escape through the voice.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Start with the feeling held back, then let it escape through the voice.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Start with the feeling held back, then let it escape through the voice.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Start with the feeling held back, then let it escape through the voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Start with the feeling held back, then let it escape through the voice. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Start with the feeling held back, then let it escape through the voice. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Start with the feeling held back, then let it escape through the voice. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Start with the feeling held back, then let it escape through the voice.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Start with the feeling held back, then let it escape through the voice.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Start with the feeling held back, then let it escape through the voice.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.639
5.581
4.247
1.79%
0.403
none
50
0.646
5.519
4.210
2.03%
n/a
g0-p05 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p05 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.634
5.702
4.291
1.25%
0.631
99.00%
-0.118
Prompting strategy: attack, sustain and release.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Build an attack, a sustained emotional middle, and a clear release.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Build an attack, a sustained emotional middle, and a clear release.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Build an attack, a sustained emotional middle, and a clear release.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Build an attack, a sustained emotional middle, and a clear release.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Build an attack, a sustained emotional middle, and a clear release. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Build an attack, a sustained emotional middle, and a clear release. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Build an attack, a sustained emotional middle, and a clear release. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Build an attack, a sustained emotional middle, and a clear release.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Build an attack, a sustained emotional middle, and a clear release.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Build an attack, a sustained emotional middle, and a clear release.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.639
5.739
4.294
1.38%
0.485
none
50
0.629
5.664
4.288
1.13%
n/a
g0-p09 · CFG 2.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p09 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.633
5.815
4.347
1.29%
0.635
100.00%
-0.093
Prompting strategy: inline-only delivery cues.
Reusable direction recipe
Stored GENERAL content — NOT forwarded in inline-onlyP09
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Let the local directions alone carry the changing emotion.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Let the local directions alone carry the changing emotion.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Let the local directions alone carry the changing emotion.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Let the local directions alone carry the changing emotion.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Let the local directions alone carry the changing emotion. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Let the local directions alone carry the changing emotion. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Let the local directions alone carry the changing emotion. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Let the local directions alone carry the changing emotion.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Let the local directions alone carry the changing emotion.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Let the local directions alone carry the changing emotion.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.642
5.858
4.348
1.15%
0.510
none
50
0.625
5.772
4.347
1.43%
n/a
g0-p08 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.626
5.857
4.366
3.23%
0.621
97.00%
-0.038
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Carry the complete intended performance through the global direction.
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Carry the complete intended performance through the global direction.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Carry the complete intended performance through the global direction.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Carry the complete intended performance through the global direction.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Natural clear speech.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Natural clear speech.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.619
5.843
4.344
3.02%
0.565
none
50
0.632
5.872
4.388
3.45%
n/a
g1-p09 · CFG 2
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p09 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.620
5.974
4.427
2.22%
0.630
95.00%
-0.039
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
Three local directions
1. A warm, low register whisper, inviting the listener in with gentle intimacy.
2. The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.
3. Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
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 repair-historical_control at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Predeclared intervention; see exact genome and changed fields.
0.620
5.974
4.427
2.22%
0.630
95.00%
-0.039
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
Three local directions
1. A warm, low register whisper, inviting the listener in with gentle intimacy.
2. The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.
3. Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy. [3.8 seconds duration]) We should probably leave before they notice we are missing.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence. [3.5 seconds duration]) Although... I am quite happy hiding here with you.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence. [5.0 seconds duration]) Give me one more quiet minute, and then I promise we will go.
Literal words in the separate text field:
We should probably leave before they notice we are missing. Although... I am quite happy hiding here with you. Give me one more quiet minute, and then I promise we will go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [3.8 seconds duration]) We should probably leave before they notice we are missing.
(Natural clear speech. [3.5 seconds duration]) Although... I am quite happy hiding here with you.
(Natural clear speech. [5.0 seconds duration]) Give me one more quiet minute, and then I promise we will go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Adopt a warm, conspiratorial whisper, as if you are sharing a beautiful, fragile truth with a trusted confidant.
SCRIPT:
(A warm, low register whisper, inviting the listener in with gentle intimacy.) [3.8 seconds duration] We should probably leave before they notice we are missing.
(The whisper should be smooth and flowing, like silk, with a gentle, confidential cadence.) [3.5 seconds duration] Although... I am quite happy hiding here with you.
(Sustain the warmth while keeping the volume at the absolute minimum, a hushed confidence.) [5.0 seconds duration] Give me one more quiet minute, and then I promise we will go.
Reference versus no-reference breakdown
reference
50
0.625
5.976
4.425
2.77%
0.565
none
50
0.614
5.973
4.429
1.66%
n/a
g1-p04 · CFG 1.25
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p04 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.617
5.924
4.420
2.82%
0.623
98.00%
-0.030
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper focusing on the physical closeness to the listener. The breath should feel like it is barely escaping the mouth.
Three local directions
1. Focus on the breath sound, almost a sigh of secrecy.
2. Maintain that physical closeness in the whisper texture.
3. Let the whisper feel physically close to the microphone.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper focusing on the physical closeness to the listener. The breath should feel like it is barely escaping the mouth.
SCRIPT:
(Focus on the breath sound, almost a sigh of secrecy. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Maintain that physical closeness in the whisper texture. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Let the whisper feel physically close to the microphone. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper focusing on the physical closeness to the listener. The breath should feel like it is barely escaping the mouth.
SCRIPT:
(Focus on the breath sound, almost a sigh of secrecy.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Maintain that physical closeness in the whisper texture.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Let the whisper feel physically close to the microphone.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.624
5.939
4.387
2.75%
0.573
none
50
0.610
5.908
4.452
2.90%
n/a
g0-p04 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.613
5.932
4.367
1.68%
0.627
98.00%
-0.077
Prompting strategy: organic conversation.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Speak organically and spontaneously, thinking while speaking, never reading a list.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Speak organically and spontaneously, thinking while speaking, never reading a list.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Speak organically and spontaneously, thinking while speaking, never reading a list.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Speak organically and spontaneously, thinking while speaking, never reading a list.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Speak organically and spontaneously, thinking while speaking, never reading a list. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Speak organically and spontaneously, thinking while speaking, never reading a list. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Speak organically and spontaneously, thinking while speaking, never reading a list. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Speak organically and spontaneously, thinking while speaking, never reading a list.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Speak organically and spontaneously, thinking while speaking, never reading a list.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Speak organically and spontaneously, thinking while speaking, never reading a list.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.611
5.949
4.369
1.62%
0.527
none
50
0.614
5.915
4.365
1.73%
n/a
g0-p03 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.582
6.013
4.432
1.97%
0.614
98.00%
-0.045
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. You urgently need an unheard listener to understand what just happened.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. You urgently need an unheard listener to understand what just happened.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. You urgently need an unheard listener to understand what just happened.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. You urgently need an unheard listener to understand what just happened.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. You urgently need an unheard listener to understand what just happened. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. You urgently need an unheard listener to understand what just happened. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. You urgently need an unheard listener to understand what just happened. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. You urgently need an unheard listener to understand what just happened.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. You urgently need an unheard listener to understand what just happened.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. You urgently need an unheard listener to understand what just happened.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.592
6.024
4.431
2.17%
0.559
none
50
0.573
6.002
4.433
1.77%
n/a
g1-p06 · CFG 2.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p06 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.570
6.097
4.503
3.06%
0.606
98.00%
0.002
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
Settle into a sustained, barely audible murmur. The emotion should be present but contained, like a secret held tightly in the chest.
Three local directions
1. Maintain near silence, focusing on the emotional weight.
2. Keep the intelligibility present but submerged in quiet.
3. The whisper must never rise above a breathy murmur.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Settle into a sustained, barely audible murmur. The emotion should be present but contained, like a secret held tightly in the chest.
SCRIPT:
(Maintain near silence, focusing on the emotional weight. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Keep the intelligibility present but submerged in quiet. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(The whisper must never rise above a breathy murmur. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Settle into a sustained, barely audible murmur. The emotion should be present but contained, like a secret held tightly in the chest.
SCRIPT:
(Maintain near silence, focusing on the emotional weight.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Keep the intelligibility present but submerged in quiet.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(The whisper must never rise above a breathy murmur.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.586
6.088
4.486
2.55%
0.605
none
50
0.553
6.106
4.520
3.57%
n/a
g0-p07 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p07 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.559
6.111
4.475
1.29%
0.611
100.00%
-0.041
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. React to a silent person right beside you; generate only one voice.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. React to a silent person right beside you; generate only one voice.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. React to a silent person right beside you; generate only one voice.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. React to a silent person right beside you; generate only one voice.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. React to a silent person right beside you; generate only one voice.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. React to a silent person right beside you; generate only one voice. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. React to a silent person right beside you; generate only one voice. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. React to a silent person right beside you; generate only one voice. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. React to a silent person right beside you; generate only one voice.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. React to a silent person right beside you; generate only one voice.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. React to a silent person right beside you; generate only one voice.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.566
6.093
4.440
1.28%
0.562
none
50
0.551
6.128
4.510
1.30%
n/a
g1-p01 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p01 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.548
6.078
4.495
2.06%
0.600
98.00%
-0.011
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper focus on the breath sound as the primary vocal texture Maintain extreme closeness to the microphone for intimacy
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. An extremely quiet intimate whisper prioritize the audible intake and exhale of breath
2. Continue whispering focusing on a breathy texture keeping the vocal projection minimal and internal
3. Remain a near-silent whisper ensuring the breath sounds support the minimal spoken words
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper focus on the breath sound as the primary vocal texture Maintain extreme closeness to the microphone for intimacy
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper focus on the breath sound as the primary vocal texture Maintain extreme closeness to the microphone for intimacy
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Natural clear speech.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Natural clear speech.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.546
6.059
4.468
1.73%
0.592
none
50
0.549
6.098
4.523
2.38%
n/a
g1-p08 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p08 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.546
6.113
4.489
1.60%
0.603
100.00%
-0.002
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Deliver the lines with a hesitant, almost nervous softness. Think of speaking when you are afraid someone might overhear you.
Three local directions
1. A nervous, shaky whisper, as if testing the air before speaking. Very close proximity.
2. Keep the whisper slightly breathy but focus on articulating the consonants with extreme care and softness.
3. Maintain the hesitant quality, letting the emotion of the secret hang in the near-silent delivery.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Deliver the lines with a hesitant, almost nervous softness. Think of speaking when you are afraid someone might overhear you.
SCRIPT:
(A nervous, shaky whisper, as if testing the air before speaking. Very close proximity. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Keep the whisper slightly breathy but focus on articulating the consonants with extreme care and softness. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Maintain the hesitant quality, letting the emotion of the secret hang in the near-silent delivery. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Deliver the lines with a hesitant, almost nervous softness. Think of speaking when you are afraid someone might overhear you.
SCRIPT:
(A nervous, shaky whisper, as if testing the air before speaking. Very close proximity.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Keep the whisper slightly breathy but focus on articulating the consonants with extreme care and softness.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Maintain the hesitant quality, letting the emotion of the secret hang in the near-silent delivery.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.527
6.116
4.472
1.83%
0.602
none
50
0.565
6.111
4.506
1.37%
n/a
g1-p07 · CFG 1.25
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p07 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.532
6.155
4.528
1.94%
0.596
99.00%
0.011
Prompting strategy: reacting to an unheard partner.
Reusable direction recipe
Stored GENERAL content
Start with a breathy, almost inaudible murmur, like sharing a secret into a closed ear. Maintain extreme intimacy throughout.
Three local directions
1. A barely there breathy murmur like sharing a secret into a closed ear. No vocal projection.
2. Shift to a very soft, almost sighing whisper, letting the words drift rather than land.
3. Fade the whisper down to the absolute edge of hearing, maintaining the intimate closeness.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Start with a breathy, almost inaudible murmur, like sharing a secret into a closed ear. Maintain extreme intimacy throughout.
SCRIPT:
(A barely there breathy murmur like sharing a secret into a closed ear. No vocal projection. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Shift to a very soft, almost sighing whisper, letting the words drift rather than land. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Fade the whisper down to the absolute edge of hearing, maintaining the intimate closeness. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Start with a breathy, almost inaudible murmur, like sharing a secret into a closed ear. Maintain extreme intimacy throughout.
SCRIPT:
(A barely there breathy murmur like sharing a secret into a closed ear. No vocal projection.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Shift to a very soft, almost sighing whisper, letting the words drift rather than land.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Fade the whisper down to the absolute edge of hearing, maintaining the intimate closeness.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.527
6.147
4.511
1.96%
0.614
none
50
0.536
6.164
4.546
1.91%
n/a
g0-p02 · CFG 1.25
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
Current ranking: EARS Steering v2
Emotion/style only. Content Enjoyment: 0%. Genuineness: 0%. These remain visible diagnostics, not reward terms. WER and speaker identity are separate qualification gates, not terms in this score.
Observed head means below refer to g0-p02 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.528
6.120
4.442
2.22%
0.589
99.00%
-0.026
Prompting strategy: audible mechanics.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Make breath pressure, pitch motion, onset and release physically audible.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Make breath pressure, pitch motion, onset and release physically audible.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Make breath pressure, pitch motion, onset and release physically audible.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Make breath pressure, pitch motion, onset and release physically audible.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Make breath pressure, pitch motion, onset and release physically audible. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Make breath pressure, pitch motion, onset and release physically audible. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Make breath pressure, pitch motion, onset and release physically audible. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Make breath pressure, pitch motion, onset and release physically audible.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Make breath pressure, pitch motion, onset and release physically audible.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Make breath pressure, pitch motion, onset and release physically audible.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.520
6.128
4.424
1.82%
0.577
none
50
0.536
6.111
4.461
2.62%
n/a
g0-p01 · CFG 1.0
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 1.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.439
6.276
4.519
2.13%
0.553
98.00%
0.000
Prompting strategy: concise trained baseline.
Reusable direction recipe
Stored GENERAL content
An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Deliver the intended emotion unmistakably, without adding words.
Three local directions
1. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Deliver the intended emotion unmistakably, without adding words.
2. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Deliver the intended emotion unmistakably, without adding words.
3. Remain a near-silent whisper, intelligible and emotionally present, never louder. Deliver the intended emotion unmistakably, without adding words.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Deliver the intended emotion unmistakably, without adding words. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Deliver the intended emotion unmistakably, without adding words. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Deliver the intended emotion unmistakably, without adding words. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. An extremely quiet intimate whisper, audible breath and no projected speaking voice. Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Remain a near-silent whisper, intelligible and emotionally present, never louder. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(An extremely quiet intimate whisper, audible breath and no projected speaking voice. Deliver the intended emotion unmistakably, without adding words.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Continue whispering clearly, with soft consonants and a confidential spontaneous rhythm. Deliver the intended emotion unmistakably, without adding words.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Remain a near-silent whisper, intelligible and emotionally present, never louder. Deliver the intended emotion unmistakably, without adding words.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.465
6.258
4.488
2.61%
0.603
none
50
0.414
6.295
4.550
1.65%
n/a
g1-p05 · CFG 1.5
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p05 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.389
6.319
4.555
2.06%
0.531
99.00%
0.016
Prompting strategy: situation and intention.
Reusable direction recipe
Stored GENERAL content
Continue whispering with an almost conspiratorial quality. Emphasize the softness of the 's' and 'th' sounds while maintaining the confidential pace.
Three local directions
1. Soften all consonants, making them almost porous.
2. Let the rhythm feel like a shared, private secret.
3. Deliver the words with utmost discretion and softness.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Continue whispering with an almost conspiratorial quality. Emphasize the softness of the 's' and 'th' sounds while maintaining the confidential pace.
SCRIPT:
(Soften all consonants, making them almost porous. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Let the rhythm feel like a shared, private secret. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Deliver the words with utmost discretion and softness. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Continue whispering with an almost conspiratorial quality. Emphasize the softness of the 's' and 'th' sounds while maintaining the confidential pace.
SCRIPT:
(Soften all consonants, making them almost porous.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Let the rhythm feel like a shared, private secret.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Deliver the words with utmost discretion and softness.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.
Reference versus no-reference breakdown
reference
50
0.393
6.318
4.517
1.14%
0.619
none
50
0.384
6.320
4.593
2.98%
n/a
g1-p02 · CFG 1.25
pilot-v1HISTORICAL_NONCANONICAL_TIMING_SYNTAX
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 g1-p02 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.
Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.
Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.
Historical quality-fitness diagnostic — not the current steering ranking
Exact score for this condition
This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.
50% target emotion/style, 25% content enjoyment, 25% genuineness. The target is vn_S_WHIS_reg. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.
Fixed target normalization: z = (raw score − 1.767) / 1.599; 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.
Passes legacy speech/identity gates only; burst reliability is not established. N=100 across both languages, both reference arms and all five seeds. Joint prompt+CFG candidate, not an isolated optimal-CFG estimate.
0.333
6.340
4.516
1.98%
0.503
99.00%
0.008
Prompting strategy: GENERAL-only delivery cues.
Reusable direction recipe
Stored GENERAL content
Continue whispering with a conspiratorial tone emphasizing soft consonant articulation Use a private conversational pace
Three local directions — NOT forwarded in GENERAL-onlyP08; neutral speech cues remain
1. Continue whispering adopting a tone of shared secret the consonants must feel delicate and hushed
2. Maintain the confidential rhythm letting small pauses build suspense in the delivery
3. Remain a near-silent whisper ensuring the soft consonants carry the weight of the message
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Continue whispering with a conspiratorial tone emphasizing soft consonant articulation Use a private conversational pace
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Literal words in the separate text field:
Come a little closer, because I do not want to wake anyone. I think... I finally understand what you were trying to tell me. Let us sit here quietly for another minute before we go.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) Come a little closer, because I do not want to wake anyone.
(Natural clear speech. [4.6 seconds duration]) I think... I finally understand what you were trying to tell me.
(Natural clear speech. [4.2 seconds duration]) Let us sit here quietly for another minute before we go.
Canonical duration-layout rewrite — not itself an observed legacy result
This fixes the bracket placement only. It is not evidence that the corrected prompt achieves the legacy scores. See the separately scored schema-repair experiment.
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Continue whispering with a conspiratorial tone emphasizing soft consonant articulation Use a private conversational pace
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] Come a little closer, because I do not want to wake anyone.
(Natural clear speech.) [4.6 seconds duration] I think... I finally understand what you were trying to tell me.
(Natural clear speech.) [4.2 seconds duration] Let us sit here quietly for another minute before we go.