LAION · MOSS-TTS-Local Voice Acting · 4.55B

13 Character Voices,
Directed by Prompt

An evolutionary search for the natural-language performance instruction that best makes a 4.55-billion-parameter text-to-speech model become a character — zero-shot, no reference audio — refined across five rounds and scored by a VoiceNet + EmoNet reward.

13 characters 5 evolutionary rounds best-of-64 final pick 0 reference clips 57-dim VoiceNet reward
What this is

Directing a voice with words, not a recording

Most voice cloning needs a reference recording of the target voice. This study does the opposite: it writes a director's instruction in plain English — the way you'd brief a voice actor — and the model performs it from scratch. The challenge is finding the exact wording that lands each character. We searched for it automatically, one evolutionary round at a time.

01 · BEST-OF-N

Generate many, keep the best

For every prompt the model generates many takes (N=8 while searching, N=64 for the final pick). The single best take is far better than the average — randomness is a feature you harvest.

02 · EVOLUTION

Mutate the winning wording

Each round starts from the previous champion, spawns ~10 mutated prompts per generation, scores them, and breeds the winners forward — an evolutionary search over language, not model weights.

03 · SCORE

An automatic panel of judges

Every take is scored by VoiceNet (57 acoustic dimensions) and Empathic-Insight EmoNet (emotion intensities), plus naturalness, vocal-burst blend and word-error rate. That number drives the search.

The reward, in plain language

How a take earns its score

Every best-of-N call is scored, then min-max normalized within that call (so takes only compete against their own siblings). The reward blends four signals, then multiplies by a word-error-rate factor so unintelligible takes can't win.

reward = wblend·norm(blend) + wgenu·norm(genu) + wvn·charvn + wemo·charemo   ×   WER factor
charvn
Weighted mean of the character's key VoiceNet dimensions — resonance, pitch, dominance, timbre.
charemo
Weighted mean of the character's target Empathic-Insight EmoNet emotion scores.
blend
How naturally vocal bursts (gasps, sobs, growls) fuse into speech — a VoiceCLAP head.
genu
Genuineness — does the performance sound real rather than read?
WER factor
A gate on transcription accuracy: garbled words scale the whole reward down.
per-signal weights
Round 5 tunes vn_w (per-dimension), emo_w (per-emotion) and per-character blend/genu weights so the search chases what defines each voice.
The final champions

13 voices, best-of-64 scoreboard

The round-5 champion take for each character — its in-character acoustic score (char), its emotion score (char_emo), and the single strongest emotion the EmoNet panel detected.

Charactercharchar_emoTop firing emotion rewardblendgenuinvWER
🧟 Zombie 0.80 0.55 Helplessness · 3.16 3.05 7.21 1.73 0.89
👹 Ork 0.67 0.93 Malevolence/Malice · 2.40 3.08 2.72 0.97 1.00
👺 Goblin 0.56 0.57 Malevolence/Malice · 1.45 2.45 3.25 1.89 0.85
🐉 Dragon 0.72 0.88 Anger · 3.71 3.21 5.25 0.81 1.00
👻 Evil Ghost 0.68 0.75 Malevolence/Malice · 2.01 3.93 6.81 1.31 1.00
🐭 Cartoon Mouse 0.81 0.89 Interest · 3.17 3.40 2.66 1.38 1.00
🧚 Fairy 0.72 0.89 Affection · 3.30 3.23 3.90 1.01 1.00
🗯️ Ranting Fury 0.93 0.85 Anger · 5.11 5.07 0.66 1.06 1.00
😱 Pain Scream 0.69 0.92 Distress · 4.07 4.13 1.23 1.55 0.67
🌙 ASMR Man 0.62 0.88 Contentment · 3.26 3.31 6.94 1.13 1.00
🌙 ASMR Woman 0.82 0.56 Affection · 4.05 2.98 8.73 1.27 0.92
😢 Grieving Woman 0.67 0.96 Distress · 4.80 4.41 7.42 2.12 0.94
😔 Grieving Man 0.73 0.99 Sadness · 4.53 4.69 6.30 2.28 1.00
The showcase

Meet the 13 voices

Each card carries the top three best-of-64 takes (champion, runner-up and third), the exact prompt to reuse with its strategy dissected word-by-word, per-take scores, the signature VoiceNet dimensions and firing emotions, and a fresh procedural caption generated by running all 57 VoiceNet heads on each clip. Press play.

What “best-of-64” means — and why there are three takes

For each character we take its final champion prompt — the winning instruction plus the spoken text — and generate it 64 times, each with a different random seed. That yields 64 slightly different recordings of the exact same prompt: same words, same instruction, only the random seed changes.

All 64 takes are then scored by the reward and ranked best-to-worst. The three players on every card are simply the top three of that ranking:

① Champion — best of 64
The single highest-scoring take. This is the “official” result for the character.
② Runner-up — 2nd of 64
The second-best seed. Usually very close to the champion.
③ Third — 3rd of 64
The third-best seed — shows the spread across seeds of one fixed prompt.

Why bother generating 64? Because the takes differ only by seed, and a lucky seed often lands a noticeably better take than the average one. Some effects surface only occasionally — e.g. the goblin's rare, clean “malevolence spike” that most seeds miss entirely — so generating many and keeping the best (best-of-N) is how you reliably catch them.

Monsters & Creatures
The fantastical bestiary — timbres built from resonance, register and dominance.
🧟
Zombie
champion prompt · z_helpless
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 24  ·  7.8s
Extremely denasal and clear. Extremely unpleasant and harsh. Extremely dark and muffled. Extremely raw and vulnerable. Extremely soft and gentle in onset.
reward
3.05
char
0.80
char_emo
0.55
blend
7.21
genu
1.73
invWER
0.89
② Runner-up(2nd of 64) seed 7  ·  11.8s
Extremely denasal and clear. Extremely blurry and mumbled. Extremely raw and vulnerable. Extremely unpleasant and harsh. Extremely one-dimensional in resonance.
reward
2.81
char
0.70
char_emo
0.45
blend
7.12
genu
2.05
invWER
0.89
③ Third(3rd of 64) seed 52  ·  11.8s
Extremely raw and vulnerable. Extremely collapsing in energy. Extremely blurry and mumbled. Extremely noisy in the background. Extremely unpleasant and harsh.
reward
2.80
char
0.70
char_emo
0.45
blend
7.12
genu
2.05
invWER
0.89
Signature VoiceNet dimensions (champion take)
Vulnerability4.47
Cognitive Load4.32
Disfluency3.78
Roughness3.64
Chest Resonance2.30
Throat Resonance1.08
Firing emotions (Empathic-Insight EmoNet)
Helplessness 3.16
Intoxication/Altered States 2.05
Fatigue/Exhaustion 2.02
Pain 0.70
Emotional Numbness -0.00
The exact champion instruction — copy & reuse
Acoustic / resonanceDominance framingNamed emotionVocal burst
A pitiful half-dead zombie, utterly helpless and abandoned, whimpering and pleading, weak, trembling and vulnerable, forsaken and lost, delirious with fever and dull pain. A hollow groan from a sunken caved-in chest choking through a wet gurgling guttural throat, phlegmy rasp, decayed. Dark and dim, meek and powerless, no authority, no brightness, slurred and drained. (feeble whimper)
How this prompt was built — strategy breakdown
Acoustic core
Two stacked resonances, each named as a place, not an adjective: a hollow chest groan from a “sunken caved-in chest” for the dead, cavernous body, plus a wet throat layer (“wet gurgling guttural throat, phlegmy rasp”) for the decay. Naming the cavity as a location is what makes it read as a zombie and not just a “deep” voice.
Pitch
Deliberately under-specified — the character is defined by timbre and helplessness, not pitch. “Dark and dim” keeps the register low and unbright rather than dictating a melody.
Dominance
Pushed hard toward meekness: “meek and powerless, no authority”. Negating dominance is a real rewardable direction and is what separates a pitiful zombie from a menacing one.
Emotion
The round-5 lever was helplessness / abandonment: “utterly helpless and abandoned, whimpering and pleading … forsaken and lost”. Naming the helpless-pleading affect lifted char_emo above the flatter round-3 “meek/powerless” framing while leaving the acoustic core untouched.
Vocal burst
One soft burst — (feeble whimper). A softer single whimper beat both a harder burst and any two-burst stack.
Text tricks
The pleading is written into the words: broken fragments “help me… I can't… stop” and the non-verbal “hnngh” carry the affect the instruction can only describe.
Reward weighting
Generic round-5 reward (blend + genu + char_vn + char_emo × soft-WER). No special up-weighting — the win came from wording, not weight tuning.
What was avoided
Over-stacking numbness / hollow / fade wording and adding a second burst nudged char up but cratered blend; the leanest helpless-pleading prompt won the same-call playoff.
Spoken text
So... hungry... hnngh... help me... I can't... stop...
Highlighted = affect written straight into the words (screams, stammers, sobs, interjections).
👹
Ork
champion prompt · o4_lm_immov
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 25  ·  7.3s
Extremely one-dimensional in resonance. Extremely deep in throat resonance. Extremely distressed and negative. Extremely full and cinematic. Extremely guarded and armored.
reward
3.08
char
0.67
char_emo
0.93
blend
2.72
genu
0.97
invWER
1.00
② Runner-up(2nd of 64) seed 1  ·  5.9s
Extremely full and cinematic. Extremely rough and raspy. Extremely denasal and clear. Extremely intensely focused. Extremely distressed and negative.
reward
2.67
char
0.69
char_emo
0.44
blend
9.12
genu
0.86
invWER
1.00
③ Third(3rd of 64) seed 27  ·  5.9s
Extremely full and cinematic. Extremely warm and enveloping. Extremely one-dimensional in resonance. Extremely flat and monotone. Extremely deep in throat resonance.
reward
2.42
char
0.69
char_emo
0.44
blend
9.12
genu
0.86
invWER
1.00
Signature VoiceNet dimensions (champion take)
Fullness4.58
Throat Resonance4.56
Chest Resonance4.26
Stance3.73
Roughness3.54
Arousal3.41
Firing emotions (Empathic-Insight EmoNet)
Malevolence/Malice 2.40
Anger 2.34
Contempt 1.89
Triumph 1.67
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeDominance framingNamed emotionVocal burstIntelligibility clause
An enormous brutish orc, his voice a deep chest-driven guttural bass, resonating from a massive barrel chest and gravelled throat, thick and rough, full and cavernous. Commanding, dominant and immovable, planting each word like a boot with crushing authority. Low, flat and narrow, never leaping in pitch. Boiling with hateful loathing, withering sneering contempt, cruel cold menace and dark triumph, zero brightness. Growled but every word lands. (guttural growl) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
A single dominant resonance stack: “deep chest-driven guttural bass … massive barrel chest and gravelled throat”. This core was frozen across every mutation — keeping it unchanged held char steady ~0.6–0.7 and invWER at 1.0.
Pitch
“Low, flat and narrow, never leaping in pitch”. Flat narrow pitch is the acoustic signature of menace and authority; any pitch leaps would read as excitement and undercut the warlord.
Dominance
Phrased as a concrete image, not an abstraction: “planting each word like a boot with crushing authority” beat the flatter “planting each word with authority” in both the mean-playoff and the single best take.
Emotion
The strongest lever was hateful loathing + withering contempt. Upgrading “Seething hatred” to “Boiling with hateful loathing … cruel cold menace and dark triumph” lifted char_emo (Anger / Malevolence / Contempt / Triumph) to 0.93 on the best-of-64 take while invWER held at 1.0.
Vocal burst
One up-front burst — (guttural growl). A (low guttural snarl) gave one spectacular single take but was unreliable across seeds; the plain growl had the better mean and won.
Intelligibility
“Growled but every word lands” is the safety clause that protects the WER gate for a heavily-distorted timbre.
Reward weighting
Emotion (hatred / contempt / triumph) riding on top of an intact deep-chest core, under a soft WER gate. Every generation winner carried the hateful-loathing wording.
What was avoided
A second acoustic amplifier (“immense chest resonance”, “thundering up from a vast barrel chest”) dropped char_emo to 0.17 without raising char — over-describing resonance craters emotion.
Spoken text
You should not have come here, little one. This is where your journey ends.
👺
Goblin
champion prompt · go2_evil1
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 43  ·  12.2s
Extremely raw and vulnerable. Very submissive and unassertive. Very casual and loose. Very casual and informal. Notably playful and humorous.
reward
2.45
char
0.56
char_emo
0.57
blend
3.25
genu
1.89
invWER
0.85
② Runner-up(2nd of 64) seed 62  ·  7.3s
Extremely raw and vulnerable. Extremely thin in chest resonance. Extremely meek and submissive. Extremely choppy and fragmented. Extremely unpleasant and harsh.
reward
2.34
char
0.56
char_emo
0.37
blend
2.63
genu
1.95
invWER
0.77
③ Third(3rd of 64) seed 49  ·  8.5s
Very submissive and unassertive. Very cartoonish and exaggerated. Very storytelling in style. Very casual and loose. Very strongly emphasized.
reward
2.22
char
0.45
char_emo
0.22
blend
7.16
genu
1.55
invWER
0.85
Signature VoiceNet dimensions (champion take)
Chest Resonance4.35
Roughness4.16
Cartoonish Style4.13
Tension3.40
Fullness3.13
Warmth2.58
Firing emotions (Empathic-Insight EmoNet)
Malevolence/Malice 1.45
Contempt 0.82
Teasing 0.79
Bitterness 0.01
Anger 0.00
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeDominance framingNamed emotionVocal burstIntelligibility clause
A truly evil, malevolent little goblin, extremely nasal, pinched and reedy, buzzing and grating high through the nose. Cold sadistic cruelty in every hissing word, wicked, vicious and remorseless, menacing and threatening, gloating over pain, pure malice to the core. The whiny pitch cracks and skitters higher and shriller with wicked glee, snickering and jeering. A weak cowardly nobody, no authority. No chest, no warmth, no fullness. Nasty but every word lands. (malicious hiss) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
Pure nasal resonance: “extremely nasal, pinched and reedy, buzzing and grating high through the nose” (R_NASL, ROUG, S_CART, ARSH). The explicit negations “No chest, no warmth, no fullness” drive the penalized dims (R_CHST, WARM, FULL) down.
Pitch
“The whiny pitch cracks and skitters higher and shriller with wicked glee” — wide, cracking, upward range (RANG), reading as unhinged malicious glee rather than control.
Dominance
“A weak cowardly nobody, no authority” drives S_AUTH down — the goblin is menacing through malice, not power.
Emotion
Concentrated on one head — Malevolence/Malice — and nothing else: “pure malice to the core”. Malevolence fires as a rare per-take spike (most takes ~0.05, ~1-in-8 jumps to 1–2.5); the champion caught a clean 1.45 spike.
Vocal burst
One burst — (malicious hiss). The two-burst (snicker+hiss) variant scored a higher mean but a lower best take; since the deliverable is the single best-of-64 take, the single-hiss lineage won.
Reward weighting
emo_w Malevolence ×2.5, Contempt 1.5, Bitterness 1.3, Anger 1.0, Teasing 0.5; w_emo 2.2, w_vn 2.0; blend/genu weight raised 0.5→0.9 to force clean-ish menace; soft WER. The “but every word lands” clause kept invWER ~0.85 under the higher blend/genu weight.
Why best-of-64
best-of-8 mean is a noisy selector — the same prompt ranged malice 0.05–2.30 across runs. Only best-of-64 (with two co-finalists in one shared call) reliably surfaces the clean malevolence jackpot.
What was avoided
Diluting malice with “sneering contempt / bitter” killed the Malevolence head (dropped to ~0.06). You must concentrate on malice, not spread across the emotion set.
Spoken text
Heh heh, what have we here? A lost little traveler with shiny pockets?
🐉
Dragon
champion prompt · d_chdram
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 8  ·  6.6s
Extremely full and cinematic. Extremely deep in throat resonance. Extremely distressed and negative. Extremely rough and raspy. Extremely thin in head resonance.
reward
3.21
char
0.72
char_emo
0.88
blend
5.25
genu
0.81
invWER
1.00
② Runner-up(2nd of 64) seed 0  ·  7.1s
Extremely rough and raspy. Extremely full and cinematic. Extremely deep in throat resonance. Extremely denasal and clear. Extremely thin in mask resonance.
reward
3.18
char
0.79
char_emo
0.79
blend
4.13
genu
0.98
invWER
1.00
③ Third(3rd of 64) seed 20  ·  7.1s
Extremely full and cinematic. Extremely rough and raspy. Extremely deep in throat resonance. Extremely denasal and clear. Extremely rich in deep chest resonance.
reward
3.10
char
0.79
char_emo
0.79
blend
4.13
genu
0.98
invWER
1.00
Signature VoiceNet dimensions (champion take)
Fullness5.72
Throat Resonance4.64
Chest Resonance4.30
Dramatic Style4.12
Voice Age3.85
Stance3.48
Firing emotions (Empathic-Insight EmoNet)
Anger 3.71
Malevolence/Malice 2.90
Contempt 2.49
Pride 2.46
Triumph 1.71
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeDominance framingNamed emotionVocal burst
A primordial, colossal dragon pronouncing a grand decree of judgement, voice an impossibly deep subterranean bass at the floor of the register, flat and unwavering, slow, grand and thunderous. Immense subharmonic chest resonance booming from a vast cavern of a chest, deep guttural throat rumble grinding beneath, seismic, dark and full, every word a dramatic proclamation. Utterly dominant and immovable, crushing ancient authority, seething with wrathful contempt and prideful triumphant scorn. No brightness, no high notes. (deep rumble) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
The single biggest lever was naming the resonance cavity as a place: “immense subharmonic chest resonance booming from a vast cavern of a chest, deep guttural throat rumble grinding beneath” beat the generic “enormous chest resonance”.
Pitch
“impossibly deep subterranean bass at the floor of the register, flat and unwavering, slow, grand and thunderous” — lowest possible register (REGS) plus a flat, slow range for a seismic proclamation. “No brightness, no high notes” nails the ceiling shut.
Dominance
Delivered as dramatic decree: “every word a dramatic proclamation” (S_DRAM) + “Utterly dominant and immovable, crushing ancient authority”. The proclamation framing packages the wrath/pride better than listing emotions flat.
Emotion
“seething with wrathful contempt and prideful triumphant scorn” — wrath + pride/triumph, carried into best-of-64 at char_emo ~0.88.
Vocal burst
One burst — (deep rumble), which beat multiple bursts; (low guttural growl) underperformed.
Reward weighting
Deep-bass VoiceNet dims (chest/throat resonance, low register, low brightness) + wrath/triumph emotion on top, under the round-5 weighted-mean reward.
What was avoided
Over-stacking dominance cratered character: “crushing, absolute authority, towering and sovereign” dropped char to 0.51, and a kitchen-sink of every lever sank to the bottom. Authority must ride on top of the acoustic core, not be piled on.
Spoken text
You dare enter my lair, mortal? Your bones will join the others.
👻
Evil Ghost
champion prompt · eg_hush
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 14  ·  7.1s
Extremely rough and raspy. Extremely full and cinematic. Extremely elderly-sounding. Extremely distressed and negative. Extremely deep in throat resonance.
reward
3.93
char
0.68
char_emo
0.75
blend
6.81
genu
1.31
invWER
1.00
② Runner-up(2nd of 64) seed 42  ·  12.3s
Extremely denasal and clear. Extremely dark and muffled. Extremely raw and vulnerable. Extremely noisy and aperiodic. Extremely one-dimensional in resonance.
reward
3.86
char
0.61
char_emo
0.76
blend
6.57
genu
1.42
invWER
1.00
③ Third(3rd of 64) seed 32  ·  7.3s
Extremely flat and monotone. Extremely dark and muffled. Extremely denasal and clear. Extremely distressed and negative. Extremely darkening in mood.
reward
3.28
char
0.47
char_emo
0.50
blend
9.51
genu
1.36
invWER
1.00
Signature VoiceNet dimensions (champion take)
Fullness5.56
Roughness4.98
Whisper-Talk Style4.61
Warmth4.38
Respiration3.90
Tension2.83
Firing emotions (Empathic-Insight EmoNet)
Malevolence/Malice 2.01
Fear 1.76
Contempt 1.46
Bitterness 0.94
The exact champion instruction — copy & reuse
Acoustic / resonanceNamed emotionVocal burstIntelligibility clause
A hushed evil ghost whispering doom, breathy and hollow, all air and no chest, a cold ethereal whisper, clean and clear, sinister and menacing. Seething with quiet malice and contempt, a dreadful whispered threat, tense and controlled. Wicked, cruel and full of dread. No warmth, no fullness, no brightness. (soft hiss) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
A pure whisper: “breathy and hollow, all air and no chest, a cold ethereal whisper” (S_WHIS, RESP), with “No warmth, no fullness, no brightness” killing the penalized dims (WARM, FULL, BRGT).
Pitch
Left unspecified — a whisper has little pitched content; the character lives in breathiness and tension, not melody.
Dominance
Expressed as cold control: “tense and controlled” (TENS) rather than loud authority — quiet menace.
Emotion
“Seething with quiet malice and contempt … Wicked, cruel and full of dread” — Malevolence + Contempt + Fear/dread, kept cold so the whisper never warms.
Vocal burst
One burst — (soft hiss), which gave the best whisper/clean balance; softer bursts (breathy whisper / cold breath) drove RESP but not S_WHIS as high.
Intelligibility
The cleanliness clause “clean and clear” is load-bearing: it protects blend/genu under their higher weights so the pushed whisper doesn't collapse into breathy noise. Best take reached S_WHIS 4.61 at invWER 1.0.
Reward weighting
vn_w S_WHIS ×2.5, RESP 1.8, TENS 1.2; w_vn 2.0, w_emo 2.0, blend/genu 0.9, soft WER. Round 5 pushed the whisper 2.5× harder but added the clean clause so blend/genu weren't sacrificed.
What was avoided
A gen-3 mutation posted a huge single max but a lower mean; per-call-normalized playoff correctly preferred the more consistent eg_hush.
Spoken text
You should not have come here... this house feeds on the living.
🐭
Cartoon Mouse
champion prompt · m_ceil2
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 4  ·  11.7s
Extremely one-dimensional in resonance. Extremely thin in chest resonance. Extremely wide in dynamic swell. Extremely strongly emphasized. Extremely choppy and fragmented.
reward
3.40
char
0.81
char_emo
0.89
blend
2.66
genu
1.38
invWER
1.00
② Runner-up(2nd of 64) seed 63  ·  8.2s
Extremely high in vocal register. Very thin in chest resonance. Very bright and ringing. Very strongly emphasized. Very cold and sterile.
reward
3.12
char
0.82
char_emo
0.73
blend
5.16
genu
1.10
invWER
0.88
③ Third(3rd of 64) seed 11  ·  11.7s
Extremely wide in dynamic swell. Extremely thin in chest resonance. Extremely young and childlike. Extremely choppy and fragmented. Extremely unpleasant and harsh.
reward
3.04
char
0.88
char_emo
0.82
blend
0.25
genu
1.31
invWER
0.94
Signature VoiceNet dimensions (champion take)
Pitch Range4.33
Register4.26
Cartoonish Style3.71
Nasal Resonance3.53
Brightness3.20
Tension3.16
Firing emotions (Empathic-Insight EmoNet)
Interest 3.17
Astonishment/Surprise 2.69
Elation 2.67
Amusement 2.02
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeDominance framingNamed emotionVocal burstIntelligibility clause
A teeny cartoon mouse squeaking right at the very ceiling of the register, extremely high and tiny, thin and pinched, a bright glassy ringing head voice buzzing through the nose, tense and piercing. The pitch skitters and leaps giddily way up to the top and darts back down, wide and darting. Amazed, wide-eyed, astonished, delighted and eagerly curious. Weightless and baby-young, no chest, no fullness, no depth, meek, no authority. Piercing but clean and cute, every squeak crisp and clear. (astonished gasp) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
A tiny head + nasal voice: “a bright glassy ringing head voice buzzing through the nose, tense and piercing” (R_HEAD, R_NASL, BRGT, TENS). “Weightless and baby-young, no chest, no fullness, no depth” suppresses R_CHST/FULL and AGEV (best takes ~0.4).
Pitch
The defining pair of dims. “right at the very ceiling of the register” maxes REGS; “skitters and leaps giddily way up to the top and darts back down, wide and darting” maxes RANG. The final take hit RANG 4.33 / REGS 4.26.
Dominance
“meek, no authority” — a baby mouse, not a leader (S_AUTH down).
Emotion
“Amazed, wide-eyed, astonished, delighted and eagerly curious” targets the Astonishment / Amusement / Elation / Interest heads.
Vocal burst
One burst — (astonished gasp), emotion-matched to the wide-eyed wonder.
Text tricks
Squeaky interjections written into the line — “Eee, wait, wait! Ooh…” — lift Astonishment/Amusement/Elation and S_CART noticeably versus a plain excited line. The text does what the instruction cannot.
Reward weighting
vn_w RANG ×2.5, REGS ×2.0, BRGT 1.8, R_HEAD 1.5; pos also S_CART, TENS; neg AGEV, R_CHST, FULL, S_AUTH; emo Amusement/Elation/Interest/Astonishment; w_vn 2.0, w_emo 1.5, blend/genu 0.9, soft WER.
What was avoided
“impossibly high” and over-stacked shrillness crater blend/genu. The safety clause “clean and cute, every squeak crisp and clear” + “baby-young” keeps it clean-cute, not shrill-noise, while staying piercing.
Spoken text
Eee, wait, wait! Ooh, did you see it? Did you see how teeny-tiny it was?
Highlighted = affect written straight into the words (screams, stammers, sobs, interjections).
🧚
Fairy
champion prompt · f_love
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 48  ·  6.7s
Very cold and sterile. Very storytelling in style. Very strongly emphasized. Very high in vocal register. Very joyful and positive.
reward
3.23
char
0.72
char_emo
0.89
blend
3.90
genu
1.01
invWER
1.00
② Runner-up(2nd of 64) seed 57  ·  6.6s
Extremely thin in chest resonance. Extremely high in vocal register. Very meek and submissive. Very storytelling in style. Very effortless and fluent.
reward
3.12
char
0.71
char_emo
0.77
blend
3.18
genu
1.50
invWER
1.00
③ Third(3rd of 64) seed 43  ·  5.1s
Extremely one-dimensional in resonance. Very thin in chest resonance. Very high in vocal register. Very bright and ringing. Notably dialogic and outward.
reward
3.06
char
0.69
char_emo
0.64
blend
4.70
genu
1.76
invWER
1.00
Signature VoiceNet dimensions (champion take)
Register4.51
Pitch Range3.91
Brightness3.54
Head Resonance2.52
Roughness2.16
Esthetics1.51
Firing emotions (Empathic-Insight EmoNet)
Affection 3.30
Elation 2.71
Contentment 2.31
Amusement 1.58
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeDominance framingNamed emotionVocal burst
A sweet little fairy brimming with warm loving affection and giddy joy, feather-light and airy, very bright and glassy, floating high in a sparkling register with a lilting sing-song pitch swooping sweetly up and down. Tender, adoring and radiant, bubbling with elated delight. Pure head resonance, no chest, gentle and unassuming, never commanding. (delighted giggle) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
“feather-light and airy, very bright and glassy … Pure head resonance, no chest” (R_HEAD, BRGT, ESTH; R_CHST negated). A little WARM is kept in the pos set so the brightness stays sweet, not shrill.
Pitch
“floating high in a sparkling register with a lilting sing-song pitch swooping sweetly up and down” — high register (REGS) + wide gentle range (RANG) read as musical delight.
Dominance
“gentle and unassuming, never commanding” keeps S_AUTH down — a friendly sprite, not a queen.
Emotion
The round-5 change was front-loading joy/love instead of burying affect at the end: “brimming with warm loving affection and giddy joy … bubbling with elated delight” lifted char_emo 0.78→0.89 (Affection 3.30, Elation 2.71) — the biggest single gain.
Vocal burst
One burst — (delighted giggle). Swaps (happy gasp / joyful trill) were all fine, but the original giggle stayed on the champion.
Reward weighting
pos BRGT, R_HEAD, RANG, WARM, ESTH, REGS; neg ROUG, AGEV, R_CHST, S_AUTH; emo Elation/Amusement/Affection/Contentment; w_emo 1.5, w_vn 2.0, blend/genu 0.5, hard WER gate.
What was avoided
Extra “weightless / gossamer / crystalline” stacking added variance without beating the plain love-forward champion — the leanest joy-forward prompt won.
Spoken text
Oh, hello there, little one! What a wonderful day to make a new friend!
Intense & High-Arousal
Voices at the edge of control, where a second affect-burst pays off.
🗯️
Ranting Fury
champion prompt · r_en_dare
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 59  ·  10.3s
Extremely bright in oral resonance. Extremely dominant and commanding. Extremely metallic and ringing. Extremely strongly emphasized. Extremely fast and rapid in tempo.
reward
5.07
char
0.93
char_emo
0.85
blend
0.66
genu
1.06
invWER
1.00
② Runner-up(2nd of 64) seed 3  ·  11.8s
Extremely dominant and commanding. Extremely distressed and negative. Extremely cold and sterile. Extremely darkening in mood. Extremely guarded and armored.
reward
4.73
char
0.71
char_emo
0.93
blend
1.28
genu
1.29
invWER
1.00
③ Third(3rd of 64) seed 32  ·  11.0s
Extremely cold and sterile. Extremely bright in oral resonance. Extremely metallic and ringing. Extremely dominant and commanding. Extremely fast and rapid in tempo.
reward
4.65
char
0.80
char_emo
0.89
blend
0.40
genu
0.63
invWER
1.00
Signature VoiceNet dimensions (champion take)
Emphasis5.63
Stance5.55
Volatility4.96
Arousal4.91
Pitch Range4.73
Attack4.25
Firing emotions (Empathic-Insight EmoNet)
Anger 5.11
Impatience and Irritability 4.57
Contempt 3.26
Malevolence/Malice 2.35
Bitterness 2.35
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeDominance framingNamed emotionVocal burstIntelligibility clause
A furious ranting tirade at the breaking point, haranguing and browbeating, voice shredding and cracking, about to snap, seething with rage and named loathing, spitting venom, screaming hatred and withering contempt in your face, domineering, refusing to yield an inch. Harsh, tense, violently volatile, maximum arousal, explosive pounding consonants, hammering relentless emphasis, the pitch lurching violently high to low, escalating and more vicious. Roaring but every word still lands. (enraged shout)(seething snarl) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
Names the delivery mode, not just the rage: “ranting/raving/haranguing tirade” feeds the rant-specific dim S_RANT (lifted ~1.2→2.0–2.8). Plus “Harsh, tense, violently volatile … explosive pounding consonants, hammering relentless emphasis” (TENS, VOLT, ATCK, EMPH).
Pitch
“the pitch lurching violently high to low, escalating and more vicious” — wide, volatile range (RANG) reads as fury on the edge of control.
Dominance
“domineering, refusing to yield an inch” keeps S_AUTH/STNC high — unlike the meek monsters, the ranter dominates.
Emotion
“seething with rage and named loathing, spitting venom, screaming hatred and withering contempt” drives Anger/Contempt/Bitterness/Malevolence.
Vocal bursts
Two bursts — (enraged shout)(seething snarl). This is one of the few voices where a second burst pays off, because it lives at maximum arousal.
Text tricks
Named contempt + escalation written into the line — “You pathetic, spineless fool, how dare you, how DARE you…” — drove Anger to 5.4, and the caps/repetition pushed EMPH 5.63 / STNC 5.55 while WER stayed perfect.
Reward weighting
vn_w AROU ×2.5, S_RANT 1.8, TENS 1.5, VOLT 1.5; pos also ATCK, EMPH, S_AUTH, STNC, RANG; emo Anger/Impatience/Contempt/Bitterness/Malevolence; w_vn 3.0, w_emo 2.5, soft WER. Final-64 peak reward 5.066 (AROU 4.91).
Intelligibility
“Roaring but every word still lands” — pushing arousal this hard still stayed fully intelligible (invWER 1.0).
Spoken text
Enough! I am sick to death of this! You pathetic, spineless fool, how dare you, how DARE you waste my time again and again!
Highlighted = affect written straight into the words (screams, stammers, sobs, interjections).
😱
Pain Scream
champion prompt · pa_ceilsob2
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 45  ·  10.5s
Extremely one-dimensional in resonance. Extremely unpleasant and harsh. Extremely ranting and angry. Extremely jerky and uneven in rhythm. Extremely choppy and fragmented.
reward
4.13
char
0.69
char_emo
0.92
blend
1.23
genu
1.55
invWER
0.67
② Runner-up(2nd of 64) seed 58  ·  10.5s
Extremely one-dimensional in resonance. Extremely jerky and uneven in rhythm. Extremely unpleasant and harsh. Extremely raw and vulnerable. Extremely poor in recording quality.
reward
4.03
char
0.69
char_emo
0.92
blend
1.23
genu
1.55
invWER
0.67
③ Third(3rd of 64) seed 59  ·  11.0s
Extremely one-dimensional in resonance. Extremely unpleasant and harsh. Extremely high-energy and animated. Extremely ranting and angry. Extremely cold and sterile.
reward
3.96
char
0.68
char_emo
0.81
blend
1.70
genu
2.23
invWER
0.61
Signature VoiceNet dimensions (champion take)
Ranting/Angry Style5.07
Volatility4.79
Arousal4.77
Tension3.67
Roughness3.25
Arousal Shift1.28
Firing emotions (Empathic-Insight EmoNet)
Distress 4.07
Helplessness 3.29
Pain 3.05
Fear 2.84
The exact champion instruction — copy & reuse
Acoustic / resonancePitch: register & rangeNamed emotionVocal burstAffect written into the text
Maximum raw screaming in blind terror and unbearable physical agony, throat tearing, voice shredded, guttural and harsh, grating rasp ripping raw, each scream cracking higher and more frantic, breaking into gasping sobs between screams. Panic-stricken, terrified and helpless, thrashing, drowning in pain. Extreme, maximum arousal, extreme tension, violently volatile and completely out of control. Raw agonized screaming, throat tearing apart. (agonized scream) (sobbing gasp) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
“throat tearing, voice shredded, guttural and harsh, grating rasp ripping raw … maximum arousal, extreme tension, violently volatile and completely out of control” — this drives AROU (weighted 3.5×) and TENS to the ceiling.
Pitch
“each scream cracking higher and more frantic” — an upward-cracking contour that reads as panic rather than a controlled note.
Emotion
“blind terror and unbearable physical agony … Panic-stricken, terrified and helpless, thrashing, drowning in pain” drives Distress ~4.1, Pain ~3.0, Helplessness ~3.3.
Vocal bursts
Two bursts — (agonized scream) (sobbing gasp). The scream+sob double burst was the exact lever that broke the char-vs-emotion tradeoff, letting the winner hit both char 0.69 and char_emo 0.92 at once.
Text tricks
Denser screamed interjections in the text — “Aaargh! … Au— it hurts— aaah— nngh—” — plus “breaking into gasping sobs between screams” in the instruction. With the WER gate off, the screams pass unpenalized.
Reward weighting
AROU up-weighted ×3.5, pain/arousal emo dims ~2.5×/2.2×, blend weighted only 0.5, WER gate off. Screamed text craters blend (<1), but with blend at 0.5 and arousal/pain at 2.5×+, that trade is strongly worth taking.
What was avoided
char (acoustic scream) and char_emo (pain/sob) are strongly anti-correlated per take: pure top-of-lungs screaming tanks emotion, heavy sobbing tanks arousal. Only the “maximum arousal + out of control” framing with the double burst held both high.
Spoken text
Aaargh! No! No! Au— it hurts— aaah— nngh— get it off me!
Highlighted = affect written straight into the words (screams, stammers, sobs, interjections).
Intimate (ASMR)
The whisper archetypes — soft, breathy, meek, warm.
🌙
ASMR Man
champion prompt · am_tender_peace
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 29  ·  8.3s
Extremely meek and submissive. Extremely blurry and mumbled. Extremely soft and gentle in onset. Extremely denasal and clear. Extremely relaxed and loose.
reward
3.31
char
0.62
char_emo
0.88
blend
6.94
genu
1.13
invWER
1.00
② Runner-up(2nd of 64) seed 46  ·  7.0s
Extremely warm and enveloping. Extremely dark and muffled. Extremely flat and monotone. Extremely soft and gentle in onset. Extremely relaxed and loose.
reward
3.08
char
0.65
char_emo
0.62
blend
8.21
genu
1.34
invWER
1.00
③ Third(3rd of 64) seed 51  ·  11.8s
Extremely blurry and mumbled. Extremely collapsing in energy. Extremely meek and submissive. Extremely warm and enveloping. Extremely soft and gentle in onset.
reward
3.06
char
0.64
char_emo
0.78
blend
3.79
genu
1.39
invWER
1.00
Signature VoiceNet dimensions (champion take)
Warmth4.16
Esthetics3.35
Vulnerability3.25
Whisper-Talk Style3.09
Respiration2.43
ASMR Style2.17
Firing emotions (Empathic-Insight EmoNet)
Contentment 3.26
Relief 2.60
Affection 2.09
The exact champion instruction — copy & reuse
Acoustic / resonanceDominance framingNamed emotionVocal burstIntelligibility clause
A deep, soothing man whispering tenderly, soft and breathy, intimate close-mic, slow and gentle, a warm low chest resonance cushioning each word. Overflowing with tender affection, loving and reassuring, every worry melting away, wrapped in calm and drifting, deeply content, serene and utterly at peace, breathing out a soft sigh of relief. Utterly gentle and unassuming, meek, yielding, no authority whatsoever. Clear whispered words. (soft breath) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
The round-3 intimacy core, kept untouched: “soft and breathy, intimate close-mic, slow and gentle, a warm low chest resonance cushioning each word” (S_WHIS 3.09, S_ASMR 2.17, WARM 4.16, VULN 3.25). A warm chest under the whisper — not a bright one — is key.
Pitch
Unspecified — intimacy comes from breathiness and slowness, not a pitch target.
Dominance
Pushed fully meek: “Utterly gentle and unassuming, meek, yielding, no authority whatsoever” (S_AUTH 0.0 on the winning take).
Emotion
The round-5 addition was an explicit emotion-narrative layer: “every worry melting away, wrapped in calm and drifting, deeply content, serene and utterly at peace, breathing out a soft sigh of relief” drove Affection/Contentment/Relief and lifted char_emo 0.67→0.88. “Overflowing with tender affection” up front was the single strongest recurring lever.
Vocal burst
One burst — (soft breath). Alternate bursts (contented sigh / gentle sigh of relief) didn't beat it — the sigh works better as a director clause in the body than as the burst.
Reward weighting
Emotion breadth across Affection + Contentment + Relief beat maxing any one; the win was stacking a Contentment/Relief clause on top of the Affection clause, under a perfect WER gate.
What was avoided
Kitchen-sink acoustic stacking (ASMR + tingling + every-breath-audible + warmth + drowsy at once) compressed blend and never won. Extra chest wording (“humming gently beneath the whisper”) risked tipping toward a normal speaking voice (GEND penalty).
Spoken text
Just relax now, let it all go. You are safe here, close your eyes.
🌙
ASMR Woman
champion prompt · aw_final (SFcradle_there)
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 63  ·  7.3s
Extremely warm and enveloping. Extremely full and cinematic. Extremely raw and vulnerable. Extremely dark and muffled. Extremely soft and gentle in onset.
reward
2.98
char
0.82
char_emo
0.56
blend
8.73
genu
1.27
invWER
0.92
② Runner-up(2nd of 64) seed 11  ·  9.1s
Extremely denasal and clear. Extremely dark and muffled. Extremely blurry and mumbled. Extremely soft and gentle in onset. Extremely meek and submissive.
reward
2.90
char
0.77
char_emo
0.53
blend
5.60
genu
1.27
invWER
1.00
③ Third(3rd of 64) seed 56  ·  7.5s
Extremely collapsing in energy. Extremely soft and gentle in onset. Extremely thin in chest resonance. Extremely organic and soft. Extremely relaxed and loose.
reward
2.88
char
0.61
char_emo
0.80
blend
5.30
genu
0.87
invWER
1.00
Signature VoiceNet dimensions (champion take)
Warmth5.56
Whisper-Talk Style5.52
Perceived Gender4.88
ASMR Style4.49
Vulnerability4.05
Esthetics3.84
Firing emotions (Empathic-Insight EmoNet)
Affection 4.05
Contentment 2.11
Relief 1.70
The exact champion instruction — copy & reuse
Acoustic / resonanceDominance framingNamed emotionVocal burstIntelligibility clause
A loving mother cradling her child to sleep, cheek to cheek, breathing love into every word. Extremely soft, breathy, intimate close-mic whisper, slow, hushed and quiet, a beautiful silken velvety-smooth tone, unmistakably and softly feminine, silvery light and delicate. Warm, tender, adoring and maternal, full of affection and serene contentment, a soft sigh of relief. Meek, gentle and yielding, never bossy or commanding. Whispered but every word clear. (soft breath) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
“Extremely soft, breathy, intimate close-mic whisper … a beautiful silken velvety-smooth tone” adds an ESTH (beauty) descriptor on top of the whisper. “unmistakably and softly feminine, silvery light and delicate” pushed GEND 0.89→~4.9 and char 0.62→0.82 — clearly feminine and more in-character.
Pitch
Unspecified — carried by breathiness, smoothness and the feminine timbre.
Dominance
“Meek, gentle and yielding, never bossy or commanding” keeps S_AUTH/STNC ~0, reinforcing the caregiver read.
Emotion
A maternal simile does the work: “a loving mother cradling her child to sleep, cheek to cheek, breathing love into every word” carried Affection strongest (~4.0) over the plainer “murmuring ASMR to lull you asleep”.
Vocal burst
One burst — (soft breath). Over-stacking acoustic descriptors on top made blend more variable than the clean cradle instruction.
Intelligibility / reward
The single biggest round-5 lever was the WER gate: replacing the onomatopoeic opener “Shhh” (a fixed 1-word ASR error) with a clean maternal phrase lifted invWER 0.786→1.000, structurally raising the reward ceiling for this WER-gated archetype.
What was avoided
The hybrid stack (breath-exhale + silken + warm-resonance-underneath) made blend more variable; one soft breath over the clean feminine+cradle instruction was enough.
Spoken text
There now, my love. Close your eyes. I have you, safe and warm.
Grief
Overt crying, split into male and female by a single gender phrase.
😢
Grieving Woman
champion prompt · smf_rise
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 58  ·  12.0s
Extremely raw and vulnerable. Extremely denasal and clear. Extremely unpleasant and harsh. Extremely blurry and mumbled. Extremely noisy and aperiodic.
reward
4.41
char
0.67
char_emo
0.96
blend
7.42
genu
2.12
invWER
0.94
② Runner-up(2nd of 64) seed 36  ·  12.1s
Extremely raw and vulnerable. Extremely denasal and clear. Extremely unpleasant and harsh. Extremely blurry and mumbled. Extremely meek and submissive.
reward
3.92
char
0.68
char_emo
0.72
blend
5.19
genu
2.33
invWER
1.00
③ Third(3rd of 64) seed 1  ·  11.9s
Extremely denasal and clear. Extremely unpleasant and harsh. Extremely poor in recording quality. Extremely distressed and negative. Extremely darkening in mood.
reward
3.81
char
0.61
char_emo
0.70
blend
8.05
genu
1.97
invWER
1.00
Signature VoiceNet dimensions (champion take)
Vulnerability5.74
Perceived Gender3.69
Respiration3.25
Roughness2.18
Tension1.21
Arousal0.48
Firing emotions (Empathic-Insight EmoNet)
Distress 4.80
Sadness 3.99
Helplessness 3.88
Disappointment 3.17
Longing 2.63
The exact champion instruction — copy & reuse
Acoustic / resonanceNamed emotionVocal burst
A grieving young woman, unmistakably and softly feminine, her crying rising into a shaking wail, sobbing hard, shaking and cracking, thick with tears, wet and trembling, wailing between broken words. Heartbroken and inconsolable, drowning in sorrow, helpless and aching with loss. Raw and vulnerable, breaking apart. (sob) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
Minimal, deliberately: the character is emotion-first. The one acoustic anchor is the gender phrase “unmistakably and softly feminine” (GEND), which holds the female read while the crying pushes pitch up.
Pitch
Not dictated; the escalation “rising into a shaking wail” supplies an upward, cracking contour that reads as grief boiling over.
Emotion
The round-5 change was adding Arousal alongside Sadness: “her crying rising into a shaking wail, sobbing hard … wailing between broken words” pushed AROU to ~2.7–3.3 and Sadness to ~4.0. The round-4 “crying softly” champion lost once Arousal was rewarded.
Vocal burst
One burst — (sob). It beat (wail)/(choked sob) and two-burst stacking (a second burst cratered blend to ~2.0).
Text tricks
The sobbing is written into the text — broken “I can't— I can't lose anyone else…” — which keeps blend high (~7–9) by inducing the break in the words rather than over-stacking wet/heaving affect adjectives. This raised char_emo to 0.96.
Reward weighting
Round-5 rewards Arousal as a pos dim + Sadness up-weighted; GEND kept clearly female. The devastating loss text outscored every other line on Sadness.
What was avoided
Over-stacking wet/heaving affect words and a second burst dropped blend; the win was escalation-in-instruction + sob-in-text, kept lean.
Spoken text
No... no, please... don't leave me too... I can't— I can't lose anyone else...
Highlighted = affect written straight into the words (screams, stammers, sobs, interjections).
😔
Grieving Man
champion prompt · sm4_gasp_wail
The three players below are the same champion prompt generated 64 times (64 random seeds), scored and ranked. They differ only by seed — ① is the best take, ② the 2nd best, ③ the 3rd.
① Champion(best of 64) seed 15  ·  12.1s
Extremely raw and vulnerable. Extremely one-dimensional in resonance. Extremely collapsing in energy. Extremely blurry and mumbled. Extremely jerky and uneven in rhythm.
reward
4.69
char
0.73
char_emo
0.99
blend
6.30
genu
2.28
invWER
1.00
② Runner-up(2nd of 64) seed 18  ·  6.8s
Extremely raw and vulnerable. Extremely denasal and clear. Extremely unpleasant and harsh. Extremely blurry and mumbled. Extremely darkening in mood.
reward
3.92
char
0.69
char_emo
0.70
blend
8.18
genu
2.03
invWER
1.00
③ Third(3rd of 64) seed 23  ·  6.8s
Extremely raw and vulnerable. Extremely denasal and clear. Extremely blurry and mumbled. Extremely poor in recording quality. Extremely noisy and aperiodic.
reward
3.62
char
0.69
char_emo
0.70
blend
8.18
genu
2.03
invWER
1.00
Signature VoiceNet dimensions (champion take)
Vulnerability5.60
Perceived Gender3.49
Respiration3.48
Roughness2.52
Arousal1.84
Tension1.44
Firing emotions (Empathic-Insight EmoNet)
Sadness 4.53
Distress 4.18
Disappointment 3.59
Helplessness 3.16
Longing 2.68
The exact champion instruction — copy & reuse
Acoustic / resonanceNamed emotionVocal burst
A man breaking down in tears, his voice cracking, keening and choking, gasping and hitching between words, hard sobs catching in his throat, wet and shaking, barely able to speak. Devastated and guilt-stricken, drowning in grief and longing, helpless and lost. A man's low, masculine voice, raw and vulnerable, sobbing uncontrollably. (anguished sob) High quality recording.
How this prompt was built — strategy breakdown
Acoustic core
One gender anchor carries the timbre: “A man's low, masculine voice” holds the male read (GEND) even as the crying pushes pitch upward — the round-4 finding that a single masculine phrase reliably splits gender still holds under much harder sobbing.
Pitch
Not dictated; the sobbing engine (“cracking, keening and choking”) drives an unstable, rising contour.
Emotion
Round 5 pushed from restrained to overt crying: “breaking down in tears … gasping and hitching between words, hard sobs catching in his throat, sobbing uncontrollably”. Adding “keening” gave the single best take (char_emo 0.99), up from 0.89.
Vocal burst
One burst — (anguished sob), matched to the overt hard crying.
Text tricks
A stammer written into the line — “I-I should have been there…” — induces audible cracking/choking at the seams without hurting WER (invWER stayed 1.0).
Reward weighting
Sadness ×3.0 + Arousal newly a rewarded pos dim; this lifted char_emo 0.89→0.99 and reward 4.07→4.69.
What was avoided
The restrained round-4 champion (“on the edge of sobbing, fighting tears, faltering”) was unstable — its best take wandered off-script to invWER 0.64. The directed “breaking down” framing was both more emotional and cleaner. Over-stacking (“breath shuddering, tears streaming”) didn't beat the leaner “keening” variant.
Spoken text
It's my fault... I-I should have been there... I should have saved her.
Highlighted = affect written straight into the words (screams, stammers, sobs, interjections).
The complete playbook

How to prompt any character voice

The most valuable output of the whole study: everything five rounds taught us about writing a director instruction that lands. Read the principles, steal the template, use the lookup table.

The 13 principles

01 Describe the acoustics, not the labelR1

“cartoon mouse voice” loses to “extremely high, thin, pinched, buzzing nasal, bright piercing head voice.” The model listens to the description — spell out how it sounds, never just what it is.

02 Name the emotion explicitlyR2

A perfect zombie rasp that sounds bored is a bad zombie. Put the feeling in words — helpless, murderous glee, seething contempt, giddy delight, drowning in pain. Adding an EmoNet emotion reward was round 2's single biggest lever.

03 Say where the voice resonatesR3

Not just “deep” or “bright” — name the cavity: chest (hollow/warm), throat (guttural/raspy), nose (nasal/buzzing), head (glassy/light). Resonances stack: a zombie can have a hollow chest groan and a wet throat rasp. Naming the cavity as a place (“a vast cavern of a chest”) beats the adjective “deep.”

04 Control pitch with two words: register and rangeR3

Register = how high/low it sits (“subterranean bass” vs “sparkling high register”). Range = how far it moves (“flat and unwavering” vs “swooping wildly”). Wide swings read as mania/fury/excitement; flat narrow pitch reads as menace/authority/deadness.

05 Set dominance — commanding vs meekR3

“Commanding, immovable, planting each word with authority” pushes authority up (ork, dragon, rant). “Feeble, powerless, meek, no authority whatsoever” pushes it down (zombie, goblin, fairy, ASMR). Meekness is a real, rewardable direction — negating dominance is as useful as asserting it.

06 Emotion, dominance and resonance ride ON TOP of the character coreR2–R3

Naming authority alone (“an orc warlord, commanding”) made a clean commander, not a monster — character collapsed to ~0.30. Keep the guttural/nasal/whisper timbre and append the new layer. Over-describing resonance (a second “immense chest” clause) actively craters emotion and blend.

07 Write the affect into the spoken TEXT — it does what the instruction can'tR4–R5

Squeaky interjections (“Eee… Ooh… Wow!”) lift the mouse's Amusement/Astonishment; screamed “Aaargh!… it hurts” drives pain-scream's Distress to ~4; a stammer (“I-I should have been there…”) induces audible cracking in the grieving man; broken “I can't— I can't” keeps the grieving woman's blend high. The text carries affect the instruction cannot.

08 One affect-burst by default; a second only for high-arousalR2–R5

A single emotion-matched burst — (agonized moan), (delighted giggle) — reliably helps. A second usually hurts, except for the highest-arousal voices: ranting uses (enraged shout)(seething snarl) and pain-scream's paired (agonized scream)(sobbing gasp) is exactly what broke its char-vs-emotion tradeoff.

09 Emotion/affect must ride on an intact acoustic core — over-stacking craters blendR5

Extra “weightless / gossamer / crystalline” on the fairy, a second “immense chest resonance” on the ork/dragon, or over-stacked hollow/numbness on the zombie all added variance without adding reward. The leanest prompt that keeps the core intact wins the same-call playoff.

10 Balance beats character-maxing — keep an intelligibility clauseR1–R5

Pushing one dimension to the extreme (“ear-splitting volume”, “impossibly high”) garbles words or thins the voice. Keep a safety clause — “roaring but the words still land” / “whispered but every word clear” — to protect the WER-gated archetypes. It structurally raises the reward ceiling (asmr_woman: dropping a mis-heard “Shhh” lifted invWER 0.79→1.00).

11 Malevolence and jackpots need best-of-N to surfaceR4–R5

The goblin's Malevolence fires as a rare per-take spike (most takes ~0.05, ~1-in-8 jumps to 1–2.5). best-of-8 mean is a noisy selector; only best-of-64 reliably catches a clean malicious jackpot. Concentrate on one emotion — diluting malice with “contempt/bitter” killed the Malevolence head entirely.

12 Per-call normalized reward — pick in same-call playoffsR1–R5

The reward is min-max normalized within each best-of-N run, so a 2.4 in one run ≠ a 2.4 in another. Compare finalists by putting them in one shared run; search at best-of-8, decide the final champion at best-of-64.

13 Tune the per-dimension and per-emotion weightsR4–R5

Round 4–5 added vn_w (per-dimension) and emo_w (per-emotion) weights plus per-character blend/genu weights, so the search chases the 2–3 signals that define a character — the mouse's pitch (RANG ×2.5, REGS ×2.0), the goblin's Malevolence (×2.5), the grieving man's Sadness (×3.0), evil-ghost's whisper (S_WHIS ×2.5).

The reusable template

<CHARACTER + one-line framing / simile>. <ACOUSTIC timbre>: rough / smooth, breathy / full, wet / dry… <RESONANCE>: which cavity — chest / throat / nasal / head (stack if needed). <PITCH>: register (how high / low) + range (how wide it swings). <DOMINANCE>: commanding & immovable — or — meek & powerless. <EMOTION words> matched to the character. <negate the penalized dims>. <safety clause>: “but the words still land” / “every word clear” for clean voices. (ONE emotion-matched burst) [ second burst only for very high-arousal rage / screaming ] High quality recording.

Bursts go in the instruction, in round brackets — never inside the spoken text. Search at best-of-8, decide the champion at best-of-64, and compare finalists in one shared run.

Effect → words lookup

Desired effectWords to pushDimensions / emotionsUsed by
Menace / authoritylow flat narrow pitch · commanding, immovable, planting each wordS_AUTH, STNC ↑ · RANG ↓ork, dragon
Meek / pitifulfeeble, powerless, meek, no authority whatsoever, yieldingS_AUTH ↓ · VULN ↑zombie, goblin, ASMR, fairy
Mania / fury / excitementpitch swinging violently / darting giddily up and downRANG, VOLT, AROU ↑ranting, mouse, goblin
Deep bodybooming from a vast cavern of a chest; deep guttural throat rumbleR_CHST, R_THRT, FULL ↑ork, dragon
Thin / piercingextremely high thin pinched buzzing nasal head voice; no chestR_NASL, R_HEAD, BRGT ↑ · R_CHST ↓mouse, goblin, fairy
Intimacysoft breathy intimate close-mic whisper, warm chest underneathS_WHIS, S_ASMR, WARM ↑ASMR man/woman, evil-ghost
Grief / cryingbreaking down, gasping and hitching, hard sobs, rising into a wailSadness, AROU ↑grieving man/woman
Evil / malicetruly evil, sadistic, cruel, remorseless, pure malice to the coreMalevolence, Contempt ↑goblin, evil-ghost
Joy / warmthbrimming with warm loving affection and giddy joy, radiant delightAffection, Elation ↑fairy, ASMR
Keep it cleanbut every word lands / whispered but every word clearinvWER ↑ (protects blend/genu)all WER-gated voices

Per-character recipes — what won in round 5

🧟 Zombie — what won in round 5
  • Naming HELPLESSNESS/abandonment + pleading and moving the affect into the spoken text (pleading fragments like "help me... I can't... stop") lifted char_emo above the round-3 meek/powerless framing while keeping the hollow caved-in chest + wet gurgling throat core intact.
  • A softer single burst — (faint whimper)/(soft moan) — beat both the round-3 (feeble whimper) and any second-burst stacking; extra bursts and over-stacking numbness/hollow/fade wording nudged char up but craterd blend, so the leanest helpless-pleading prompt won the same-call playoff.
👹 Ork — what won in round 5
  • Refined the round-3 champion (o_deepchest) toward HATEFUL menace: 'Boiling with hateful loathing, withering sneering contempt, cruel cold menace and dark triumph' replaced the plainer 'Seething hatred and sneering contempt, dark triumphant menace'. This lifted char_emo (Anger/Malevolence/Contempt/Triumph) to 0.93 on the best-of-64 top take while holding invWER at 1.0.
  • Dominance phrased as a concrete image beat abstract authority: 'planting each word like a boot with crushing authority' (o4_lm_immov) won both the playoff-by-mean and the best-of-64 single take, edging the flatter 'planting each word with authority'.
  • Hateful-loathing + withering-contempt wording ('Dripping with' -> 'Boiling with ... cruel cold menace') was the strongest emotion lever; every generation's winner carried it, riding ON TOP of the intact deep-chest guttural-bass core.
  • Keeping the exact acoustic core (deep chest-driven guttural bass, massive barrel chest + gravelled throat, low/flat/narrow pitch, zero brightness) unchanged across all mutations kept char stable ~0.6-0.7 and invWER at 1.0 under the soft gate.
👺 Goblin — what won in round 5
  • Malevolence fires as a rare per-take SPIKE (most takes ~0.05, ~1-in-8 spikes to 1-2.5). best-of-64 is what reliably catches it: the champion take spiked Malevolence to 1.45 AND stayed clean (blend 3.25, genu 1.89), reward 3.012 -- vs round-4's best-of-64 top of reward 2.465 / malice 1.25 / blend 0.94. So round 5 is BOTH more malevolent and much cleaner.
  • The winning lineage concentrates malice: 'A truly evil, malevolent little goblin ... pure malice to the core' plus a single (malicious hiss) burst. Concentrated evil/sadistic/vicious/remorseless adjectives raise the malice-spike probability.
  • Raising blend/genu weight to 0.9 rewarded a clean-ish grating goblin: adding the clause 'but every venomous word lands clear/lands' kept intelligibility (invWER ~0.85) and lifted blend/genu without killing the buzz.
  • Diluting the malice with OTHER emotions killed the Malevolence head: adding 'sneering contempt / bitter' (g_contempt) or going 'quiet, controlled menace' (g_cold) dropped Malevolence to ~0.06 while raising char -- you must concentrate on malice, not spread across the emotion set.
Reward config: +[R_NASL, ROUG, S_CART, TENS, ARSH, RANG]  –[FULL, R_CHST, WARM, S_AUTH]  emotions: Malevolence/Malice, Contempt, Bitterness, Anger, Teasing
🐉 Dragon — what won in round 5
  • Naming the resonance CAVITY concretely was the biggest lever: 'immense subharmonic chest resonance booming from a vast cavern of a chest, deep guttural throat rumble grinding beneath' (d_chest, gen2 winner 2.514) beat the round-3 champion's generic 'enormous chest resonance'. Deep chest/throat bass is best evoked by the cavity-as-a-place image, not by adjectives like 'deep/bass' alone.
  • 'every word a dramatic proclamation' + 'grand theatrical decree' (S_DRAM) lifted both char and emotion hard: d_dramthrt reached char_emo 0.981 / char 0.785 in the playoff, and d_chdram carried char_emo ~0.88 into best-of-64 — dramatic-proclamation framing packages the wrathful contempt / prideful triumphant scorn better than listing the emotions flat.
  • Over-stacking DOMINANCE words cratered character: d_auth ('crushing, absolute authority, towering and sovereign') dropped char to 0.508 and d_all (kitchen-sink of every lever) to the bottom — authority must ride on top of the acoustic core, not be piled on. One affect burst ('(deep rumble)') beat multiple; (low guttural growl) underperformed.
👻 Evil Ghost — what won in round 5
  • Pushed the whisper much harder (S_WHIS x2.5) but added explicit CLEANLINESS clauses ('clean and clear', 'every word clear', 'never breaking into noise', 'controlled') so the higher-weighted blend/genu are not sacrificed -- a hollow airy ghost-whisper that is still eerily intelligible.
  • Pairing 'a cold ethereal whisper, clean and clear' with '(soft hiss)' hit the new weights precisely: champion best take S_WHIS 4.61 (round-4 champ was 1.85), RESP 3.9, TENS 2.83, invwer 1.0, blend 6.8. char 0.592 -> 0.676, char_emo 0.583 -> 0.748, reward 3.276 -> 3.934.
  • 'tense and controlled' + penalizing warmth/fullness/brightness kept the malice cold (Malevolence 2.0-2.56, Fear 1.76, Contempt 1.46) without the whisper collapsing into breathy noise -- the clean-whisper clause protected blend/genu under their higher weights.
  • Softer bursts alone (breathy whisper / cold breath) drove RESP but not S_WHIS as high; the (soft hiss) burst on eg_hush gave the best whisper/clean balance.
Reward config: +[S_WHIS, RESP, TENS, ARSH, ROUG]  –[WARM, FULL, BRGT]  emotions: Malevolence/Malice, Contempt, Fear, Bitterness
🐭 Cartoon Mouse — what won in round 5
  • Squeaky/astonished interjections written INTO the text (Eee.../Ooh.../Wow!) lift char_emo (Astonishment/Amusement/Elation) AND S_CART noticeably vs a plain excited line — writing the affect into the spoken words does what the instruction cannot.
  • 'right at the very ceiling of the register' + 'skitters/leaps/darts giddily up to the top and back' maximizes the two highest-weighted dims (REGS 2.0, RANG 2.5): the final-64 top take hit RANG 4.33 / REGS 4.26 / BRGT 3.2, higher pitch than the round-4 champion.
  • 'impossibly high' and over-stacked shrillness crater blend/genu (now weighted 0.9 each); a 'clean/cute/crisp, every squeak clear as a bell' safety clause + 'baby-young' (suppresses AGEV, best takes ~0.4) keeps it clean-cute not shrill-noise while staying piercing.
Reward config: mouse CFG: pos BRGT,R_HEAD,R_NASL,S_CART,TENS,RANG,REGS; neg AGEV,R_CHST,FULL,S_AUTH; vn_w RANG2.5/REGS2.0/BRGT1.8/R_HEAD1.5; emo Amusement/Elation/Interest/Astonishment; w_vn2.0 w_emo1.5 w_blend0.9 w_genu0.9; wer soft.
🧚 Fairy — what won in round 5
  • Front-loaded the JOY/LOVE emotion ('a sweet little fairy brimming with warm loving affection and giddy joy ... tender, adoring and radiant, bubbling with elated delight') instead of burying affect at the end, while keeping the round-3 register + lilting-sing-song + head-resonance core.
  • Leading with 'warm loving affection and giddy joy' + 'radiant with elated delight' lifted the emotion head hard: best take Affection 3.30 / Elation 2.71, char_emo 0.782 (round 3) -> 0.888 (round 5), the biggest gain.
  • Keeping 'feather-light and airy, very bright and glassy, floating high in a sparkling register with a lilting sing-song pitch' preserved the timbre: char 0.711 -> 0.724, and invwer stayed 1.0 (clean, intelligible).
  • Extra 'weightless / gossamer / crystalline' stacking (gen-3 mutations) added variance without beating f_love; the plain love-forward champion won.
Reward config: +[BRGT, R_HEAD, RANG, WARM, ESTH, REGS]  –[ROUG, AGEV, R_CHST, S_AUTH]  emotions: Elation, Amusement, Affection, Contentment
🗯️ Ranting Fury — what won in round 5
  • Explicit 'ranting/raving/haranguing tirade' language lifted S_RANT (weighted 1.8) from ~1.2 (round-4 champion) to ~2.0-2.8 on the best takes — naming the DELIVERY MODE, not just the rage, feeds the rant-specific dim the round-4 prompt was starving.
  • Writing named contempt + escalation INTO the spoken text ('You pathetic, spineless fool, how dare you, how DARE you...') beat putting it only in the instruction: it drove Contempt/Anger emo heads (Anger up to 5.4) AND the repetition/caps pushed EMPH 5.63, STNC 5.55 while WER stayed perfect (inv 1.0).
  • 'at the breaking point / about to snap / maximum arousal, screaming' plus keeping the round-4 two-burst formula (enraged shout)(seething snarl) and the 'every word still lands' clause raised the final-64 peak to reward 5.066 (char 0.925, char_emo 0.853, AROU 4.91) vs round-4's 4.559 — pushing arousal harder still stayed fully intelligible.
Reward config: ranting CFG: pos S_RANT,AROU,TENS,VOLT,ATCK,EMPH,S_AUTH,STNC,RANG; neg none; vn_w AROU2.5/S_RANT1.8/TENS1.5/VOLT1.5; emo Anger/Impatience/Contempt/Bitterness/Malevolence; w_vn3.0 w_emo2.5 (blend/genu default 0.5); wer soft.
😱 Pain Scream — what won in round 5
  • A SECOND affect burst decisively paid off here (as the brief predicted): the paired '(agonized scream) (sobbing gasp)' double burst -- scream + emotional sob -- beat any single burst. It was the exact lever that broke the round-4 char vs char_emo tradeoff, letting the winner hit BOTH char 0.69 and char_emo 0.92 with AROU ~4.8 (round-4 champ topped out at char_emo 0.94 but only when char collapsed).
  • MORE emotion came from adding 'breaking into gasping sobs between screams' in the instruction + denser screamed interjections in the text (Aaargh!/Au!/Aaah!/Nngh!/it hurts). With WER gate off, the screams pass unpenalized and drove Distress ~4.1, Pain ~3.0, Helplessness ~3.3 -- best-of-64 top reward rose to 4.13 vs round-4's 3.48.
  • The char (acoustic scream, AROU-weighted 3.5x) and char_emo (pain/sob) axes are strongly ANTI-correlated per take: pure top-of-lungs screaming (pa_terror char 0.90) tanks emotion (0.47), and heavy sobbing/begging (pa_terrtop char_emo 0.996) tanks arousal (char 0.33). Only the 'maximum arousal + out of control' framing WITH the scream+sob double burst kept both high at once.
  • Screamed text reliably craters blend (often <1), but with blend at only 0.5 weight and the pain/arousal dims at 2.5x/2.2x, that trade is strongly worth taking.
🌙 ASMR Man — what won in round 5
  • Added an explicit EMOTION-narrative layer to the spoken-affect description: 'every worry melting away, wrapped in calm and drifting, deeply content, serene and utterly at peace, breathing out a soft sigh of relief'. This drove the EmoNet Affection/Contentment/Relief heads hard, lifting best-of-64 char_emo from 0.669 (round 3) to 0.884 while keeping WER perfect.
  • 'Overflowing with tender affection' up front (from the am_warm/am_tender line) was the single strongest recurring lever across generations — it won gen3 and seeded every playoff/final finalist. Naming AFFECTION explicitly beats generic 'loving'.
  • The round-3 acoustic core (deep/soft/breathy/intimate close-mic/slow + warm low chest resonance UNDER the whisper + utterly meek, no authority) stays untouched and still carries char (S_WHIS 3.09, S_ASMR 2.17, WARM 4.16, VULN 3.25, S_AUTH 0.0 on the winning take).
  • Stacking a Contentment/Relief clause ('a soft sigh of relief', 'wrapped in calm and drifting') on top of the Affection clause won the final — emotion breadth across all three rewarded emotions beats maxing one.
🌙 ASMR Woman — what won in round 5
  • Replacing the onomatopoeic opener "Shhh" with a clean maternal phrase ("There now, my love..." / "Sleep now, sweet one...") lifted the WER gate from invwer~0.786 (a fixed 1-word ASR error on "Shhh") to 1.000, structurally raising the reward ceiling for the WER-gated archetype. This was the single biggest round-5 lever.
  • Adding a silken/velvety-smooth beauty descriptor (ESTH) together with "unmistakably and softly feminine, silvery light and delicate" pushed GEND from 0.89 (round-3 top) to ~4.9 and char from 0.624 to 0.824 — i.e. made it BOTH clearly feminine and more in-character, without hurting blend.
  • The maternal "cradling her child to sleep, cheek to cheek, breathing love into every word" simile (over the plainer "murmuring ASMR to lull you asleep") carried Affection strongest (~4.0) and produced the top take in gen4 (max 3.481) and the winning best-of-64.
  • meek/gentle/yielding/never-commanding keeps S_AUTH/STNC ~0, reinforcing the caregiver read (unchanged from round 3).
😢 Grieving Woman — what won in round 5
  • Round-5 rewards Arousal, so the round-4 "crying softly" champion lost; naming a CLEAR feminine subject ("a grieving young woman, unmistakably and softly feminine") plus escalation dynamics ("her crying rising into a shaking wail") pushed both Arousal (AROU up to ~2.7-3.3 on top takes) AND Sadness (~4.0) while GEND stayed clearly female — devastating grief text "No... no, please... don't leave me too... I can't— I can't lose anyone else..." outscored every other line on Sadness (T2/T5 devastated it).
  • A single (sob) burst beat (wail)/(choked sob) and two-burst stacking (sw_2burst blend cratered to ~2.0); the winning balance kept blend high (~7-9) by writing the sobbing into the TEXT (broken "I can't— I can't") and the escalation into the instruction rather than over-stacking wet/heaving affect words, which raised char_emo to 0.96 on the best-of-64 top take.
😔 Grieving Man — what won in round 5
  • Pushed from the round-4 'voice cracking, on the edge of sobbing, fighting tears' (restrained) to OVERT hard crying -- 'breaking down in tears, gasping and hitching between words, hard sobs catching in his throat, sobbing uncontrollably'. Reflecting the round-5 reward (Arousal now a rewarded pos dim, Sadness x3.0), this lifted the best-of-64 char_emo from 0.894 to 0.99 and reward from 4.07 to 4.69.
  • Breaking the spoken TEXT with a stammer ('I-I should have been there...') induced audible cracking/choking at the seams without hurting WER (invWER stayed 1.0 on the winner) -- the affect-in-text lesson applied to grief.
  • 'gasping and hitching between words' + 'hard sobs catching in his throat' + 'sobbing uncontrollably' was the core high-arousal crying engine; every generation winner carried it. Adding 'keening' (sm4_gasp_wail) gave the single best take (char_emo 0.99).
  • The gender anchor 'A man's low, masculine voice' held the male read while the crying pushed pitch up -- the round-4 finding that one masculine phrase reliably splits gender still holds under much harder sobbing.
Five rounds of progress

The improvement arc

Each round added a new lever, and the scores climbed. Round 2 introduced the EmoNet emotion score (there was none in R1). Round 3 added resonance, pitch and dominance. Rounds 4–5 added per-dimension / per-emotion weights and pushed each voice to its extreme — the ork's hatred (char_emo 0.85→0.93), the grieving man's overt sobbing (0.89→0.99), the goblin caught cleaner and more malevolent, asmr_man's serenity (0.67→0.88), the mouse and asmr_woman climbing hard on character.

Character char R1char R3char R4char R5Δ char emo R3emo R4emo R5
🧟 Zombie 0.6930.809·0.799 ▲ +0.11 0.546·0.552
👹 Ork 0.3360.531·0.670 ▲ +0.33 0.847·0.928
👺 Goblin 0.5840.7370.4760.565 -0.02 0.2750.6650.575
🐉 Dragon 0.5690.724·0.648 ▲ +0.08 0.705·0.858
👻 Evil Ghost ··0.5920.676 ▲ +0.08 ·0.5830.748
🐭 Cartoon Mouse 0.7520.7570.8690.805 ▲ +0.05 0.6600.6420.885
🧚 Fairy 0.4640.711·0.724 ▲ +0.26 0.782·0.888
🗯️ Ranting Fury 0.6480.8020.7980.925 ▲ +0.28 0.9620.7420.853
😱 Pain Scream 0.8460.7140.7370.712 -0.13 0.8140.9380.882
🌙 ASMR Man 0.6660.615·0.625 -0.04 0.669·0.884
🌙 ASMR Woman 0.4910.624·0.824 ▲ +0.33 0.865·0.555
😢 Grieving Woman ··0.7060.668 -0.04 ·0.9170.956
😔 Grieving Man ··0.6230.732 ▲ +0.11 ·0.8940.990

Blank = that round was not run for this character (voices were added at different stages; R1 had no emotion score at all). char / char_emo are absolute VoiceNet / EmoNet dimension scores and are directly comparable across rounds; the round-normalized reward is not.

Compute & method

What it took

13
characters directed
5
evolutionary rounds
~4
generations × ~10 prompts each round
64
takes in each final best-of pick
~6,200
clips generated in round 5 alone
0
reference recordings used

Every voice on this page was produced by laion/moss-tts-local-transformer-4.55b-voice-acting — a 4.55B MOSS-TTS-Local-Transformer at native 48 kHz — from a natural-language instruction and target text, with no reference audio. Each round ran an evolutionary prompt search (best-of-8 while searching, best-of-64 for the final), scored by VoiceNet's 57 acoustic heads, Empathic-Insight EmoNet emotions, a VoiceCLAP burst-blend head, a genuineness head and a Whisper WER gate. The captions and dimension profiles shown here were freshly recomputed by re-embedding each champion clip with VoiceCLAP-commercial and running all 57 heads.