Anger and controlled ranting · Humaneness Voice Large 1.5 Alpha

Two frozen models. Exact instructions, matched seeds and complete take distributions. Find a promising recipe quickly, then inspect its limitations and listen.

Model, weights and inference code Pinned revision

← Model overview · Complete scenario take index

Listen to matched examples

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 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5803.4540.998
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.7560.832
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

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 emo_Anger. VoiceCLAP and absolute speaker cosine are diagnostics here, not extra weighted rewards.

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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

repair-canonical_both · CFG 1.5

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.

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.

heldout-selected1.50.5636.2934.0710.26%n/a0.612pass
heldout-g11.00.4236.3594.2550.30%n/a0.553pass
repair-canonical_bothcanonical-v1-explorationreacting to an unheard partner1.56.3354.0010.18%0.9980.9150.823pass
repair-historical_controlcanonical-v1-explorationreacting to an unheard partner1.56.3444.1050.41%0.9970.8800.825pass
heldout-g1canonical-v1-validationreacting to an unheard partner1.06.3594.2550.30%0.4370.4230.553pass
heldout-selectedcanonical-v1-validationreacting to an unheard partner1.56.2934.0710.26%0.5730.5630.612pass
g0-p01pilot-v1concise trained baseline1.06.3854.1941.11%0.9710.8170.811pass
g0-p02pilot-v1audible mechanics1.256.3474.1481.04%0.9810.8480.813pass
g0-p03pilot-v1situation and intention1.56.3054.1221.86%0.9820.8740.805pass
g0-p04pilot-v1organic conversation2.06.2483.9821.23%1.0000.9320.812pass
g0-p05pilot-v1attack, sustain and release2.56.1763.8681.30%1.0000.9630.805pass
g0-p06pilot-v1containment contrast3.06.0853.7851.32%1.0000.9790.799pass
g0-p07pilot-v1reacting to an unheard partner1.56.3444.1050.41%0.9970.8800.825pass
g0-p08pilot-v1GENERAL-only delivery cues2.06.3014.0403.16%0.9930.9050.796pass
g0-p09pilot-v1inline-only delivery cues2.56.1463.8983.25%1.0000.9530.790pass
g1-p01pilot-v1reacting to an unheard partner2.06.2834.1702.23%0.9950.8750.810pass
g1-p02pilot-v1reacting to an unheard partner2.56.2634.0430.38%1.0000.9270.822pass
g1-p03pilot-v1reacting to an unheard partner1.56.3734.2141.74%0.9820.8460.812pass
g1-p04pilot-v1audible mechanics1.56.3424.3221.40%0.9460.7700.800pass
g1-p05pilot-v1audible mechanics26.3384.1700.55%0.9920.8820.824pass
g1-p06pilot-v1audible mechanics1.256.3934.3651.59%0.9030.6880.780pass
g1-p07pilot-v1organic conversation2.56.2574.1776.81%0.9830.8480.767diagnostic
g1-p08pilot-v1organic conversation1.56.3764.2161.49%0.9730.8140.810pass
g1-p09pilot-v1organic conversation1.256.3924.2740.63%0.8910.6930.778pass

Listen to complete take distributions

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 1.5

canonical-v1-explorationCANONICAL

Current ranking: EARS Steering v2

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

Observed head means below refer to repair-canonical_both at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5803.4540.998
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.7560.832
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9156.3354.0010.18%0.823100.00%-0.014

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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] You promised me you would tell the truth this time.
(Natural clear speech.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(Natural clear speech.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.9176.3153.9780.23%0.625
none500.9136.3554.0230.12%n/a

heldout-selected · CFG 1.5

canonical-v1-validationCANONICAL

Current ranking: EARS Steering v2

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

Observed head means below refer to heldout-selected at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5800.6330.573
Ranting
vn_S_RANT_reg
50.00%high1.4341.2761.6230.553
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.5636.2934.0710.26%0.612100.00%-0.003

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] I was certain that I knew what would happen next.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.) [3.1 seconds duration] And then... everything felt different for a second.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] Look at me and tell me you understand what I mean.

Literal words in the separate text field:

I was certain that I knew what would happen next. And then... everything felt different for a second. Look at me and tell me you understand what I mean.
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] I was certain that I knew what would happen next.
(Natural clear speech.) [3.1 seconds duration] And then... everything felt different for a second.
(Natural clear speech.) [4.2 seconds duration] Look at me and tell me you understand what I mean.
Reference versus no-reference breakdown
reference500.5776.2714.0420.33%0.619
none500.5486.3144.1000.19%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.

Anger
emo_Anger
50.00%high0.3610.5800.3100.437
Ranting
vn_S_RANT_reg
50.00%high1.4341.2761.1280.410
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.4236.3594.2550.30%0.553100.00%0.000

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] I thought you would be here before the lights went out.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.) [3.5 seconds duration] Stay with me until I can say this properly.

Literal words in the separate text field:

I thought you would be here before the lights went out. Wait... I need a moment to understand what just happened. Stay with me until I can say this properly.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech.) [4.2 seconds duration] I thought you would be here before the lights went out.
(Natural clear speech.) [3.8 seconds duration] Wait... I need a moment to understand what just happened.
(Natural clear speech.) [3.5 seconds duration] Stay with me until I can say this properly.
Reference versus no-reference breakdown
reference500.4466.3434.2050.27%0.622
none500.4016.3764.3040.33%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.

Anger
emo_Anger
50.00%high0.3610.5804.1231.000
Ranting
vn_S_RANT_reg
50.00%high1.4341.2763.8330.959
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9796.0853.7851.32%0.79998.00%-0.027

Prompting strategy: containment contrast.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Start with the feeling held back, then let it escape through the voice.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Start with the feeling held back, then let it escape through the voice.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Start with the feeling held back, then let it escape through the voice.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Start with the feeling held back, then let it escape through the voice. [3.8 seconds duration]) I have spent the entire morning fixing what you broke.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Start with the feeling held back, then let it escape through the voice. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Bitter indignation remains audible as you insist on being taken seriously. Start with the feeling held back, then let it escape through the voice. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.

Literal words in the separate text field:

I have spent the entire morning fixing what you broke. You knew exactly what you were doing... and you did it anyway. For once in your life, take responsibility for the mess you made.
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]) I have spent the entire morning fixing what you broke.
(Natural clear speech. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Natural clear speech. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Start with the feeling held back, then let it escape through the voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Start with the feeling held back, then let it escape through the voice.) [3.8 seconds duration] I have spent the entire morning fixing what you broke.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Start with the feeling held back, then let it escape through the voice.) [4.6 seconds duration] You knew exactly what you were doing... and you did it anyway.
(Bitter indignation remains audible as you insist on being taken seriously. Start with the feeling held back, then let it escape through the voice.) [4.6 seconds duration] For once in your life, take responsibility for the mess you made.
Reference versus no-reference breakdown
reference500.9816.0553.7530.82%0.613
none500.9786.1153.8181.82%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.

Anger
emo_Anger
50.00%high0.3610.5803.9591.000
Ranting
vn_S_RANT_reg
50.00%high1.4341.2763.4270.926
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9636.1763.8681.30%0.80598.00%0.016

Prompting strategy: attack, sustain and release.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Build an attack, a sustained emotional middle, and a clear release.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Build an attack, a sustained emotional middle, and a clear release.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Build an attack, a sustained emotional middle, and a clear release.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Build an attack, a sustained emotional middle, and a clear release. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Build an attack, a sustained emotional middle, and a clear release. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Bitter indignation remains audible as you insist on being taken seriously. Build an attack, a sustained emotional middle, and a clear release. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Build an attack, a sustained emotional middle, and a clear release.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Build an attack, a sustained emotional middle, and a clear release.) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Build an attack, a sustained emotional middle, and a clear release.) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(Bitter indignation remains audible as you insist on being taken seriously. Build an attack, a sustained emotional middle, and a clear release.) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.9676.1503.8200.88%0.655
none500.9596.2013.9161.73%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.

Anger
emo_Anger
50.00%high0.3610.5803.8061.000
Ranting
vn_S_RANT_reg
50.00%high1.4341.2763.2630.905
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9536.1463.8983.25%0.79094.00%-0.036

Prompting strategy: inline-only delivery cues.

Reusable direction recipe

Stored GENERAL content — NOT forwarded in inline-onlyP09

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Let the local directions alone carry the changing emotion.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Let the local directions alone carry the changing emotion.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Let the local directions alone carry the changing emotion.
3. Bitter indignation remains audible as you insist on being taken seriously. 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:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Let the local directions alone carry the changing emotion. [3.8 seconds duration]) I have spent the entire morning fixing what you broke.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Let the local directions alone carry the changing emotion. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Bitter indignation remains audible as you insist on being taken seriously. Let the local directions alone carry the changing emotion. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.

Literal words in the separate text field:

I have spent the entire morning fixing what you broke. You knew exactly what you were doing... and you did it anyway. For once in your life, take responsibility for the mess you made.
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]) I have spent the entire morning fixing what you broke.
(Natural clear speech. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Natural clear speech. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.
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:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Let the local directions alone carry the changing emotion.) [3.8 seconds duration] I have spent the entire morning fixing what you broke.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Let the local directions alone carry the changing emotion.) [4.6 seconds duration] You knew exactly what you were doing... and you did it anyway.
(Bitter indignation remains audible as you insist on being taken seriously. Let the local directions alone carry the changing emotion.) [4.6 seconds duration] For once in your life, take responsibility for the mess you made.
Reference versus no-reference breakdown
reference500.9556.1323.8692.71%0.604
none500.9506.1613.9263.79%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.

Anger
emo_Anger
50.00%high0.3610.5803.6771.000
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.9730.865
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9326.2483.9821.23%0.81298.00%-0.024

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Speak organically and spontaneously, thinking while speaking, never reading a list.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Speak organically and spontaneously, thinking while speaking, never reading a list.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Speak organically and spontaneously, thinking while speaking, never reading a list.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Speak organically and spontaneously, thinking while speaking, never reading a list. [3.8 seconds duration]) You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Speak organically and spontaneously, thinking while speaking, never reading a list. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. Speak organically and spontaneously, thinking while speaking, never reading a list. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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]) You promised me you would tell the truth this time.
(Natural clear speech. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Natural clear speech. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Speak organically and spontaneously, thinking while speaking, never reading a list.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Speak organically and spontaneously, thinking while speaking, never reading a list.) [3.8 seconds duration] You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Speak organically and spontaneously, thinking while speaking, never reading a list.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. Speak organically and spontaneously, thinking while speaking, never reading a list.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.9416.2163.9340.22%0.615
none500.9246.2804.0292.23%n/a

g1-p02 · 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-p02 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5803.8361.000
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.9880.853
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9276.2634.0430.38%0.822100.00%-0.031

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Start with tightly controlled, clipped anger. Each word is measured but carries immense force. The escalation is internal and building.

Three local directions

1. Start with tight, controlled vocal fry conveying suppressed rage. Speak through clenched teeth.
2. Allow the pitch to rise slightly as the control cracks. The fury is sharp, precise, and aimed.
3. End with a low, gravelly tone. The anger settles into a deep, unwavering sense of being wronged.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Start with tightly controlled, clipped anger. Each word is measured but carries immense force. The escalation is internal and building.
SCRIPT:
(Start with tight, controlled vocal fry conveying suppressed rage. Speak through clenched teeth. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Allow the pitch to rise slightly as the control cracks. The fury is sharp, precise, and aimed. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(End with a low, gravelly tone. The anger settles into a deep, unwavering sense of being wronged. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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 tightly controlled, clipped anger. Each word is measured but carries immense force. The escalation is internal and building.
SCRIPT:
(Start with tight, controlled vocal fry conveying suppressed rage. Speak through clenched teeth.) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(Allow the pitch to rise slightly as the control cracks. The fury is sharp, precise, and aimed.) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(End with a low, gravelly tone. The anger settles into a deep, unwavering sense of being wronged.) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.9476.2433.9720.30%0.608
none500.9066.2824.1140.47%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.

Anger
emo_Anger
50.00%high0.3610.5803.4010.993
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.7280.816
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.9056.3014.0403.16%0.79693.00%-0.000

Prompting strategy: GENERAL-only delivery cues.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Carry the complete intended performance through the global direction.

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

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Carry the complete intended performance through the global direction.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Carry the complete intended performance through the global direction.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Carry the complete intended performance through the global direction.
SCRIPT:
(Natural clear speech.) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(Natural clear speech.) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(Natural clear speech.) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.9186.2713.9853.54%0.639
none500.8916.3324.0952.78%n/a

g1-p05 · 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-p05 at CFG 2; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5803.3560.992
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.5050.772
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8826.3384.1700.55%0.82499.00%-0.016

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Build the anger through relentless, tightly controlled vocal intensity. Every syllable must feel like a deliberate, forceful jab.

Three local directions

1. Deliver lines with sharp, clipped articulation. Keep the pitch oscillating with barely contained rage.
2. Focus on the resonance in the chest to convey deep, simmering fury.
3. End with a vocal texture that suggests deep, painful stubbornness.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Build the anger through relentless, tightly controlled vocal intensity. Every syllable must feel like a deliberate, forceful jab.
SCRIPT:
(Deliver lines with sharp, clipped articulation. Keep the pitch oscillating with barely contained rage. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Focus on the resonance in the chest to convey deep, simmering fury. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(End with a vocal texture that suggests deep, painful stubbornness. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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. Build the anger through relentless, tightly controlled vocal intensity. Every syllable must feel like a deliberate, forceful jab.
SCRIPT:
(Deliver lines with sharp, clipped articulation. Keep the pitch oscillating with barely contained rage.) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(Focus on the resonance in the chest to convey deep, simmering fury.) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(End with a vocal texture that suggests deep, painful stubbornness.) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.8806.3024.1300.63%0.623
none500.8846.3734.2110.47%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.

Anger
emo_Anger
50.00%high0.3610.5803.2340.997
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.4120.763
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8806.3444.1050.41%0.825100.00%0.000

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice. [3.8 seconds duration]) You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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]) You promised me you would tell the truth this time.
(Natural clear speech. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Natural clear speech. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.) [3.8 seconds duration] You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.8896.3294.0650.46%0.639
none500.8726.3594.1450.35%n/a

repair-historical_control · CFG 1.5

canonical-v1-explorationHISTORICAL_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 repair-historical_control at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5803.2340.997
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.4120.763
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8806.3444.1050.41%0.825100.00%0.000

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. React to a silent person right beside you; generate only one voice.) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(Bitter indignation remains audible as you insist on being taken seriously. React to a silent person right beside you; generate only one voice.) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.8896.3294.0650.46%0.639
none500.8726.3594.1450.35%n/a

g1-p01 · 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-p01 at CFG 2.0; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5803.2570.995
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.3980.755
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8756.2834.1702.23%0.81095.00%-0.031

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Initial outburst full of sharp physical tension and ragged breath. Focus on the explosive release of frustration. Maintain high volume.

Three local directions

1. A sudden burst of raw, high-energy anger. Project this from your diaphragm. Every word is a punch.
2. Slightly lower the volume but increase the internal pressure. The anger becomes a focused, vibrating threat.
3. A final, sustained wave of weary resentment. The volume drops, but the intensity of the disapproval remains heavy.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Initial outburst full of sharp physical tension and ragged breath. Focus on the explosive release of frustration. Maintain high volume.
SCRIPT:
(A sudden burst of raw, high-energy anger. Project this from your diaphragm. Every word is a punch. [3.8 seconds duration]) You promised me you would tell the truth this time.
(Slightly lower the volume but increase the internal pressure. The anger becomes a focused, vibrating threat. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(A final, sustained wave of weary resentment. The volume drops, but the intensity of the disapproval remains heavy. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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]) You promised me you would tell the truth this time.
(Natural clear speech. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Natural clear speech. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.
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. Initial outburst full of sharp physical tension and ragged breath. Focus on the explosive release of frustration. Maintain high volume.
SCRIPT:
(A sudden burst of raw, high-energy anger. Project this from your diaphragm. Every word is a punch.) [3.8 seconds duration] You promised me you would tell the truth this time.
(Slightly lower the volume but increase the internal pressure. The anger becomes a focused, vibrating threat.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(A final, sustained wave of weary resentment. The volume drops, but the intensity of the disapproval remains heavy.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.8746.2544.1102.05%0.609
none500.8766.3124.2292.40%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.

Anger
emo_Anger
50.00%high0.3610.5803.1770.982
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.4600.766
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8746.3054.1221.86%0.80595.00%0.012

Prompting strategy: situation and intention.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. You urgently need an unheard listener to understand what just happened.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. You urgently need an unheard listener to understand what just happened.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. You urgently need an unheard listener to understand what just happened.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. You urgently need an unheard listener to understand what just happened. [3.8 seconds duration]) I have spent the entire morning fixing what you broke.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. You urgently need an unheard listener to understand what just happened. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Bitter indignation remains audible as you insist on being taken seriously. You urgently need an unheard listener to understand what just happened. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.

Literal words in the separate text field:

I have spent the entire morning fixing what you broke. You knew exactly what you were doing... and you did it anyway. For once in your life, take responsibility for the mess you made.
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]) I have spent the entire morning fixing what you broke.
(Natural clear speech. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Natural clear speech. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. You urgently need an unheard listener to understand what just happened.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. You urgently need an unheard listener to understand what just happened.) [3.8 seconds duration] I have spent the entire morning fixing what you broke.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. You urgently need an unheard listener to understand what just happened.) [4.6 seconds duration] You knew exactly what you were doing... and you did it anyway.
(Bitter indignation remains audible as you insist on being taken seriously. You urgently need an unheard listener to understand what just happened.) [4.6 seconds duration] For once in your life, take responsibility for the mess you made.
Reference versus no-reference breakdown
reference500.8886.2784.0522.15%0.651
none500.8606.3334.1911.57%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.

Anger
emo_Anger
50.00%high0.3610.5803.0060.981
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.2370.715
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8486.3474.1481.04%0.81397.00%0.007

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Make breath pressure, pitch motion, onset and release physically audible.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Make breath pressure, pitch motion, onset and release physically audible.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Make breath pressure, pitch motion, onset and release physically audible.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Make breath pressure, pitch motion, onset and release physically audible. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Make breath pressure, pitch motion, onset and release physically audible. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Bitter indignation remains audible as you insist on being taken seriously. Make breath pressure, pitch motion, onset and release physically audible. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Make breath pressure, pitch motion, onset and release physically audible.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Make breath pressure, pitch motion, onset and release physically audible.) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Make breath pressure, pitch motion, onset and release physically audible.) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(Bitter indignation remains audible as you insist on being taken seriously. Make breath pressure, pitch motion, onset and release physically audible.) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.8566.3334.0951.57%0.646
none500.8416.3604.2010.51%n/a

g1-p07 · 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-p07 at CFG 2.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5802.9230.983
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.2310.712
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8486.2574.1776.81%0.76785.00%-0.019

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Explode with raw, sputtering frustration like a steam valve failing Maintain sharp, percussive consonants on stressed syllables and drive the breath from deep in the diaphragm

Three local directions

1. Start with a sudden burst of pressurized frustration driving the vocal attack
2. Maintain a tight, strained quality in the mid-range to convey acute stress
3. End with a sharp, forceful exhale punctuating the final thought
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Explode with raw, sputtering frustration like a steam valve failing Maintain sharp, percussive consonants on stressed syllables and drive the breath from deep in the diaphragm
SCRIPT:
(Start with a sudden burst of pressurized frustration driving the vocal attack [3.8 seconds duration]) You promised me you would tell the truth this time.
(Maintain a tight, strained quality in the mid-range to convey acute stress [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(End with a sharp, forceful exhale punctuating the final thought [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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]) You promised me you would tell the truth this time.
(Natural clear speech. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Natural clear speech. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.
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. Explode with raw, sputtering frustration like a steam valve failing Maintain sharp, percussive consonants on stressed syllables and drive the breath from deep in the diaphragm
SCRIPT:
(Start with a sudden burst of pressurized frustration driving the vocal attack) [3.8 seconds duration] You promised me you would tell the truth this time.
(Maintain a tight, strained quality in the mid-range to convey acute stress) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(End with a sharp, forceful exhale punctuating the final thought) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.8536.2574.1426.92%0.620
none500.8426.2584.2116.71%n/a

g1-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 g1-p03 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5802.9260.982
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.2100.710
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8466.3734.2141.74%0.81295.00%-0.022

Prompting strategy: reacting to an unheard partner.

Reusable direction recipe

Stored GENERAL content

Begin with exasperated, loud exasperation. The energy is frantic and slightly unhinged at the start. Focus on physical projection.

Three local directions

1. A frantic, almost breathless delivery of rage. Sound like you are physically struggling to keep it in.
2. Transition to a commanding, unwavering tone. The fury is directed and authoritative, not chaotic.
3. Hold the indignation in a steady, resonant tone. You are stating a fundamental truth, however angry you are.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Begin with exasperated, loud exasperation. The energy is frantic and slightly unhinged at the start. Focus on physical projection.
SCRIPT:
(A frantic, almost breathless delivery of rage. Sound like you are physically struggling to keep it in. [3.8 seconds duration]) I have spent the entire morning fixing what you broke.
(Transition to a commanding, unwavering tone. The fury is directed and authoritative, not chaotic. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Hold the indignation in a steady, resonant tone. You are stating a fundamental truth, however angry you are. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.

Literal words in the separate text field:

I have spent the entire morning fixing what you broke. You knew exactly what you were doing... and you did it anyway. For once in your life, take responsibility for the mess you made.
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]) I have spent the entire morning fixing what you broke.
(Natural clear speech. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Natural clear speech. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.
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. Begin with exasperated, loud exasperation. The energy is frantic and slightly unhinged at the start. Focus on physical projection.
SCRIPT:
(A frantic, almost breathless delivery of rage. Sound like you are physically struggling to keep it in.) [3.8 seconds duration] I have spent the entire morning fixing what you broke.
(Transition to a commanding, unwavering tone. The fury is directed and authoritative, not chaotic.) [4.6 seconds duration] You knew exactly what you were doing... and you did it anyway.
(Hold the indignation in a steady, resonant tone. You are stating a fundamental truth, however angry you are.) [4.6 seconds duration] For once in your life, take responsibility for the mess you made.
Reference versus no-reference breakdown
reference500.8526.3654.1612.66%0.617
none500.8406.3814.2660.83%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.

Anger
emo_Anger
50.00%high0.3610.5802.6780.971
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.0180.663
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8176.3854.1941.11%0.81198.00%0.000

Prompting strategy: concise trained baseline.

Reusable direction recipe

Stored GENERAL content

Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Deliver the intended emotion unmistakably, without adding words.

Three local directions

1. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Deliver the intended emotion unmistakably, without adding words.
2. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Deliver the intended emotion unmistakably, without adding words.
3. Bitter indignation remains audible as you insist on being taken seriously. 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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Deliver the intended emotion unmistakably, without adding words. [3.8 seconds duration]) You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Deliver the intended emotion unmistakably, without adding words. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. Deliver the intended emotion unmistakably, without adding words. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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]) You promised me you would tell the truth this time.
(Natural clear speech. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Natural clear speech. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.
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. Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. The anger escalates into a controlled furious confrontation, not an unintelligible scream. Bitter indignation remains audible as you insist on being taken seriously. Deliver the intended emotion unmistakably, without adding words.
SCRIPT:
(Intense anger, loud frustrated ranting, sharply stressed words and forceful breath. Deliver the intended emotion unmistakably, without adding words.) [3.8 seconds duration] You promised me you would tell the truth this time.
(The anger escalates into a controlled furious confrontation, not an unintelligible scream. Deliver the intended emotion unmistakably, without adding words.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(Bitter indignation remains audible as you insist on being taken seriously. Deliver the intended emotion unmistakably, without adding words.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.8356.3654.1361.66%0.639
none500.7996.4054.2520.55%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.

Anger
emo_Anger
50.00%high0.3610.5802.8520.973
Ranting
vn_S_RANT_reg
50.00%high1.4341.2762.0080.655
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.8146.3764.2161.49%0.81098.00%-0.021

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Shift to controlled, dangerous fury channeling intense controlled rage Focus on vocal texture becoming rough and gravelly not just volume Maintain a steady, relentless rhythmic pace

Three local directions

1. Build the intensity through vocal grit not sheer loudness
2. Maintain a tight focus on the emotional target never letting the anger dissolve into shouting
3. Project authority through deliberate pacing and measured force
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Shift to controlled, dangerous fury channeling intense controlled rage Focus on vocal texture becoming rough and gravelly not just volume Maintain a steady, relentless rhythmic pace
SCRIPT:
(Build the intensity through vocal grit not sheer loudness [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Maintain a tight focus on the emotional target never letting the anger dissolve into shouting [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Project authority through deliberate pacing and measured force [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.

Literal words in the separate text field:

I asked for one simple thing, and you laughed in my face. No... this is not another little misunderstanding you can wave away. Look at me and explain why you thought that was acceptable.
Exact neutral CFG branch
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects.
SCRIPT:
(Natural clear speech. [4.6 seconds duration]) I asked for one simple thing, and you laughed in my face.
(Natural clear speech. [4.2 seconds duration]) No... this is not another little misunderstanding you can wave away.
(Natural clear speech. [4.2 seconds duration]) Look at me and explain why you thought that was acceptable.
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. Shift to controlled, dangerous fury channeling intense controlled rage Focus on vocal texture becoming rough and gravelly not just volume Maintain a steady, relentless rhythmic pace
SCRIPT:
(Build the intensity through vocal grit not sheer loudness) [4.6 seconds duration] I asked for one simple thing, and you laughed in my face.
(Maintain a tight focus on the emotional target never letting the anger dissolve into shouting) [4.2 seconds duration] No... this is not another little misunderstanding you can wave away.
(Project authority through deliberate pacing and measured force) [4.2 seconds duration] Look at me and explain why you thought that was acceptable.
Reference versus no-reference breakdown
reference500.8436.3474.1402.09%0.618
none500.7866.4054.2920.89%n/a

g1-p04 · 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-p04 at CFG 1.5; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5802.3160.946
Ranting
vn_S_RANT_reg
50.00%high1.4341.2761.7700.594
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.7706.3424.3221.40%0.80097.00%-0.007

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Start with a sudden explosive burst of frustration pushing air hard from the diaphragm. Let the initial words crack with strained volume.

Three local directions

1. Forceful onset like a sudden physical shove. Exaggerate the initial intake of breath.
2. Maintain high tension in the throat but channel it into controlled vocal projection.
3. The final delivery should sound weary but utterly uncompromising.
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 sudden explosive burst of frustration pushing air hard from the diaphragm. Let the initial words crack with strained volume.
SCRIPT:
(Forceful onset like a sudden physical shove. Exaggerate the initial intake of breath. [3.8 seconds duration]) You promised me you would tell the truth this time.
(Maintain high tension in the throat but channel it into controlled vocal projection. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(The final delivery should sound weary but utterly uncompromising. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.

Literal words in the separate text field:

You promised me you would tell the truth this time. And now... you are standing here pretending none of it happened. Do not ask me to calm down when you will not even listen.
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]) You promised me you would tell the truth this time.
(Natural clear speech. [4.2 seconds duration]) And now... you are standing here pretending none of it happened.
(Natural clear speech. [5.0 seconds duration]) Do not ask me to calm down when you will not even listen.
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 sudden explosive burst of frustration pushing air hard from the diaphragm. Let the initial words crack with strained volume.
SCRIPT:
(Forceful onset like a sudden physical shove. Exaggerate the initial intake of breath.) [3.8 seconds duration] You promised me you would tell the truth this time.
(Maintain high tension in the throat but channel it into controlled vocal projection.) [4.2 seconds duration] And now... you are standing here pretending none of it happened.
(The final delivery should sound weary but utterly uncompromising.) [5.0 seconds duration] Do not ask me to calm down when you will not even listen.
Reference versus no-reference breakdown
reference500.7686.3314.2802.08%0.632
none500.7726.3534.3650.72%n/a

g1-p09 · 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-p09 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5801.9160.891
Ranting
vn_S_RANT_reg
50.00%high1.4341.2761.4210.495
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.6936.3924.2740.63%0.778100.00%-0.016

Prompting strategy: organic conversation.

Reusable direction recipe

Stored GENERAL content

Lower the pitch slightly allowing the anger to settle into a deep, simmering resentment The delivery must sound weighted with past disappointment Insist on every word carrying undeniable moral weight

Three local directions

1. Deliver the lines with heavy, deliberate weight as if every syllable is a burden
2. Infuse the tone with a deep, wounded indignation rather than pure explosion
3. End with a firm, unwavering pitch that demands respect
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Lower the pitch slightly allowing the anger to settle into a deep, simmering resentment The delivery must sound weighted with past disappointment Insist on every word carrying undeniable moral weight
SCRIPT:
(Deliver the lines with heavy, deliberate weight as if every syllable is a burden [3.8 seconds duration]) I have spent the entire morning fixing what you broke.
(Infuse the tone with a deep, wounded indignation rather than pure explosion [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(End with a firm, unwavering pitch that demands respect [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.

Literal words in the separate text field:

I have spent the entire morning fixing what you broke. You knew exactly what you were doing... and you did it anyway. For once in your life, take responsibility for the mess you made.
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]) I have spent the entire morning fixing what you broke.
(Natural clear speech. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Natural clear speech. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.
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. Lower the pitch slightly allowing the anger to settle into a deep, simmering resentment The delivery must sound weighted with past disappointment Insist on every word carrying undeniable moral weight
SCRIPT:
(Deliver the lines with heavy, deliberate weight as if every syllable is a burden) [3.8 seconds duration] I have spent the entire morning fixing what you broke.
(Infuse the tone with a deep, wounded indignation rather than pure explosion) [4.6 seconds duration] You knew exactly what you were doing... and you did it anyway.
(End with a firm, unwavering pitch that demands respect) [4.6 seconds duration] For once in your life, take responsibility for the mess you made.
Reference versus no-reference breakdown
reference500.7176.4034.2460.97%0.623
none500.6696.3804.3010.29%n/a

g1-p06 · 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-p06 at CFG 1.25; every recipe also shows its own values below. These are recipe means, not one hand-picked audio.

Anger
emo_Anger
50.00%high0.3610.5801.9100.903
Ranting
vn_S_RANT_reg
50.00%high1.4341.2761.3440.473
z = (raw_EARS_head − frozen_training_mean) / frozen_training_std
high target = Φ(z); low target = 1 − Φ(z)
Steering v2 = Σ(weight × normalized_target)

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

Φ is a Gaussian CDF using the frozen EARS training statistics. This makes mixed emotion/style heads comparable; it is not an emotion probability. It compresses extreme scores, so raw means remain visible too. A pure-emotion condition has only its named emotion head, at 100%. A mixed anger/ranting scene uses 50% Anger + 50% Ranting. Each recipe score is the mean over all its takes, including failed takes.

Unchanged gates: ≥90% valid terminated takes; per-take WER ≤15% to count as valid; candidate mean WER ≤10%; paired reference cosine decline no worse than −0.05. Existing family-aware burst conditions still need all requested families detected in ≥60% of takes to qualify as reliable burst recipes. A high emotion/style score alone does not establish a successful scream or a correctly timed transition.

Historical quality-fitness diagnostic — not the current steering ranking

Exact score for this condition

This is the prompt-search fitness, not the training RL reward. Higher means better under this particular proxy combination, not automatically more natural or more enjoyable to a listener.

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

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

Fixed target normalization: z = (raw score − 0.361) / 0.580; 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.

0.6886.3934.3651.59%0.78098.00%-0.007

Prompting strategy: audible mechanics.

Reusable direction recipe

Stored GENERAL content

Infuse the anger with a layer of wounded disbelief. The volume should be high but the delivery should feel intellectually superior and scathing.

Three local directions

1. Use a lower, more gravelly texture to convey deep contempt. Slow the pace slightly for impact.
2. Ensure the breath sounds are heavy and deliberate, like a measured sigh of exasperation.
3. The final tone must be one of absolute, dignified refusal to yield.
Exact tested prompt — fixed EN example, seed1701, no reference
GENERAL: One adult speaker, clean high-quality recording, no music or sound effects. Infuse the anger with a layer of wounded disbelief. The volume should be high but the delivery should feel intellectually superior and scathing.
SCRIPT:
(Use a lower, more gravelly texture to convey deep contempt. Slow the pace slightly for impact. [3.8 seconds duration]) I have spent the entire morning fixing what you broke.
(Ensure the breath sounds are heavy and deliberate, like a measured sigh of exasperation. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(The final tone must be one of absolute, dignified refusal to yield. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.

Literal words in the separate text field:

I have spent the entire morning fixing what you broke. You knew exactly what you were doing... and you did it anyway. For once in your life, take responsibility for the mess you made.
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]) I have spent the entire morning fixing what you broke.
(Natural clear speech. [4.6 seconds duration]) You knew exactly what you were doing... and you did it anyway.
(Natural clear speech. [4.6 seconds duration]) For once in your life, take responsibility for the mess you made.
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. Infuse the anger with a layer of wounded disbelief. The volume should be high but the delivery should feel intellectually superior and scathing.
SCRIPT:
(Use a lower, more gravelly texture to convey deep contempt. Slow the pace slightly for impact.) [3.8 seconds duration] I have spent the entire morning fixing what you broke.
(Ensure the breath sounds are heavy and deliberate, like a measured sigh of exasperation.) [4.6 seconds duration] You knew exactly what you were doing... and you did it anyway.
(The final tone must be one of absolute, dignified refusal to yield.) [4.6 seconds duration] For once in your life, take responsibility for the mess you made.
Reference versus no-reference breakdown
reference500.6976.3834.3270.74%0.632
none500.6786.4024.4042.44%n/a