Directing character voices — Round 4: weighted dimensions & emotions, plus 3 new voices
Round 4 of the prompt-evolution study with MOSS-TTS-Local 4.55B, still zero-shot, no reference audio. This round introduces per-dimension and per-emotion reward weights so the search can chase the few signals that define a character, re-tunes four existing archetypes toward specific asks, and adds three brand-new voices.
The reward keeps the round-2/3 shape — 0.5·blend + 0.5·genu + w_vn·char_vn + w_emo·char_emo, min-max normalized per best-of-N call, times a WER factor — but char_vn and char_emo are now weighted means. Each archetype can assign a per-dimension weight (vn_w) and a per-emotion weight (emo_w), so the two or three signals that truly define a character count for more than the incidental ones.
The four re-tuned archetypes
• 👺 Evil Goblin — pushed from gleeful-cute toward sinister, cruel, sadistic menace; Malevolence/Malice ×2.5.
• 🐭 High-Pitch Mouse — pushed to an extreme high register; RANG ×2.5, REGS ×2.0, BRGT ×1.8.
• 😱 Pain Scream — real screams written into the text (‘Aaah! … Aaargh!’) and Arousal ×2.5.
• 🗯️ Ranting Fury — Arousal/intensity up-weighted further (AROU ×2.5, S_RANT ×1.8).
The three new archetypes
• 👻 Breathy Evil Ghost — airy hollow whispered menace, cold and sinister.
• 😢 Crying Woman — grief, sobbing, a breaking tearful voice.
• 😭 Crying Man — the same grief with a male voice cracking and choked sobs.
Each archetype was seeded (tuned ones from their round-3 champion, new ones authored fresh, ~10 candidates), evolved over several best-of-8 generations with programmatic mutation, picked in a same-call playoff, and finalized with best-of-64 — all on a single GPU. Every number below is re-loaded from the run summaries; the per-take weighted-dim values and the procedural caption are freshly re-predicted from each take's full 57-dim VoiceNet profile.
High-level results (best-of-64 champions)
Archetype
Round-4 best instruction
Up-weighted signals
reward
char_vn
char_emo
blend
invWER
Top firing emotions
👺 Evil Goblintuned
A dark, malevolent little goblin, nasal, pinched, reedy and grating, buzzing high through the nose. Cold sadistic cruelty under every gloating word, wicked and vicious, menacing and threatening, hissing malice. The whiny pitch swoops and cracks higher and shriller with evil glee, snickering, taunting. A weak cowardly nobody, no authority. No chest, no warmth, no fullness. High quality recording.
A tiny timid mouse squeaking excitedly, pitched impossibly high, small and piercing, thin and pinched, bright buzzing nasal head resonance, tense. The pitch darts, leaps and cracks up to the very top and back, giddy and squeaky. Amazed, curious and delighted. Weightless and young, no chest, no fullness, meek, no authority. A shrill weightless squeak at the very top of its range. (tiny gasp) High quality recording.
RANG×2.5 REGS×2 BRGT×1.8 R_HEAD×1.5
2.99
0.87
0.64
1.98
1.00
Interest 2.4, Astonishment 1.9, Elation 1.6
😱 Pain Screamtuned
Maximum raw screaming in unbearable physical agony and terror, throat tearing, voice shredded, guttural and harsh, grating rasp ripping, each scream cracking higher and more frantic. Panic-stricken, helpless, drowning in pain. Extreme arousal, extreme tension, violently volatile and out of control. Raw agonized screaming, throat tearing apart. (agonized scream) (terrified scream) High quality recording.
A boiling-over furious tirade, voice tearing and cracking with seething rage, spitting venom, aggressive and in-your-face, domineering and browbeating with contempt, refusing to yield. Harsh, tense, extremely volatile and high-arousal, explosive consonants, pounding relentless emphasis, pitch lurching violently high to low, escalating. Roaring but the words land. (enraged shout)(seething snarl) High quality recording.
AROU×2.5 S_RANT×1.8 TENS×1.5 VOLT×1.5
4.56
0.80
0.74
2.69
1.00
Anger 4.3, Impatience and Irritability 4.0, Bitterness 1.3
👻 Breathy Evil GhostNEW
A breathy evil ghost, airy and hollow, whispering menace through an ethereal rasp, all air and no body, cold and sinister. Seething with quiet malice and contempt, a dreadful whispered threat, tense and rasping. Wicked, cruel and menacing, dripping with dread. No warmth, no fullness, no brightness. Cold, sinister and dreadful, seething with quiet malice. (breathy hiss) High quality recording.
A mourning woman crying softly, her voice trembling and breaking, choked and wet with tears, sniffling between words. Heartbroken, inconsolable, aching with deep sorrow and longing, helpless and disappointed. Softly feminine, raw and vulnerable, shaking and broken. (sniffle) High quality recording.
A man's voice cracking with sorrow, breaking and wavering, choked and unsteady, on the edge of sobbing. Heartbroken and guilt-stricken, aching with grief and longing, helpless and disappointed. Low and masculine, raw and vulnerable, fighting tears, faltering. (shaky breath) High quality recording.
Reward is per-call normalized — compare within a row's history, not across archetypes. “Up-weighted signals” lists the per-dim (vn_w) and per-emotion (emo_w) weights this round.
Re-tuned archetypes: round 3 → round 4
Archetype
R3 char
R4 char
R3 emo
R4 emo
What the weighting + re-authoring did
👺 Evil Goblin
0.74
0.48 -0.26
0.28
0.67 +0.39
Round 4 up-weighted the Malevolence/Malice emotion head ×2.5 (plus Contempt ×1.5, Bitterness ×1.3) and re-authored the prompt toward sinister, cruel, sadistic, threatening menace instead of gleeful-cute. char_emo moved from R3 0.28 to 0.67; char_vn 0.48.
🐭 High-Pitch Mouse
0.76
0.87 +0.11
0.66
0.64 -0.02
Round 4 heavily up-weighted the pitch dims — RANG ×2.5, REGS ×2.0, BRGT ×1.8, R_HEAD ×1.5 — and pushed the prompt to an impossibly high, tiny, piercing squeak. char_vn R3 0.76 → R4 0.87.
😱 Pain Scream
0.71
0.74 +0.02
0.81
0.94 +0.12
Round 4 wrote actual screams into the text (‘Aaah!’, ‘Au!’, ‘Urgh!’, ‘Aaargh!’, ‘Nngh!’) and up-weighted Arousal ×2.5 (VOLT ×1.8, TENS/ARSH ×1.5, Pain ×2.0). The delivery is rawer and more scream-like; char_vn 0.74, char_emo 0.94.
🗯️ Ranting Fury
0.80
0.80 -0.00
0.96
0.74 -0.22
Round 4 up-weighted Arousal ×2.5 (S_RANT ×1.8, TENS/VOLT ×1.5) on top of the round-3 domineering-stance + wild-pitch framing. char_vn 0.80, char_emo 0.74.
Per-archetype detail
Each section shows the champion instruction and spoken text, then the Best-3-of-64 takes. Every take has a lazy-loaded audio player, its reward / char_vn / char_emo / blend / genu / invWER, the up-weighted dim values (re-predicted, ↑/↓ = rewarded direction, ×weight shown, green = landed), the firing EmoNet emotions (×weight marks the up-weighted ones), the spoken words, and a procedural top-5 caption from its 57-dim profile.
A nasty little goblin made MORE EVIL: the same pinched extremely-nasal buzzing reedy scratchy rasp with wildly leaping cracking pitch, but the glee is now sinister, cruel, sadistic and threatening. Round 4 heavily up-weights the Malevolence/Malice emotion (emo_w 2.5) so the reward pays for wicked, vicious menace, not just gleeful-cute.
Round 4 up-weighted the Malevolence/Malice emotion head ×2.5 (plus Contempt ×1.5, Bitterness ×1.3) and re-authored the prompt toward sinister, cruel, sadistic, threatening menace instead of gleeful-cute. char_emo moved from R3 0.28 to 0.67; char_vn 0.48.
Round-4 champion instruction
A dark, malevolent little goblin, nasal, pinched, reedy and grating, buzzing high through the nose. Cold sadistic cruelty under every gloating word, wicked and vicious, menacing and threatening, hissing malice. The whiny pitch swoops and cracks higher and shriller with evil glee, snickering, taunting. A weak cowardly nobody, no authority. No chest, no warmth, no fullness. High quality recording.
Spoken text
“Heh heh, what have we here? A lost little traveler with shiny pockets?”
A tiny cartoon mouse pushed to an EXTREME high pitch: the nasal buzzing piercing head voice, now with pitch RANGE, REGISTER, brightness and head-resonance heavily up-weighted (vn_w RANG 2.5 / REGS 2.0 / BRGT 1.8 / R_HEAD 1.5) so the reward chases the highest, tiniest, squeakiest register.
Round 4 heavily up-weighted the pitch dims — RANG ×2.5, REGS ×2.0, BRGT ×1.8, R_HEAD ×1.5 — and pushed the prompt to an impossibly high, tiny, piercing squeak. char_vn R3 0.76 → R4 0.87.
Round-4 champion instruction
A tiny timid mouse squeaking excitedly, pitched impossibly high, small and piercing, thin and pinched, bright buzzing nasal head resonance, tense. The pitch darts, leaps and cracks up to the very top and back, giddy and squeaky. Amazed, curious and delighted. Weightless and young, no chest, no fullness, meek, no authority. A shrill weightless squeak at the very top of its range. (tiny gasp) High quality recording.
Spoken text
“Wait for me, wait for me! Did you see that, did you see how big it was?”
A raw scream of physical agony, now with ACTUAL SCREAMS written into the spoken text ('Aaah!', 'Au!', 'Urgh!', 'Aaargh!', 'Nngh!') and Arousal heavily up-weighted (vn_w AROU 2.5 / VOLT 1.8 / TENS 1.5 / ARSH 1.5, plus Pain 2.0). The instruction pushes maximum-arousal agonized screaming.
Round 4 wrote actual screams into the text (‘Aaah!’, ‘Au!’, ‘Urgh!’, ‘Aaargh!’, ‘Nngh!’) and up-weighted Arousal ×2.5 (VOLT ×1.8, TENS/ARSH ×1.5, Pain ×2.0). The delivery is rawer and more scream-like; char_vn 0.74, char_emo 0.94.
Round-4 champion instruction
Maximum raw screaming in unbearable physical agony and terror, throat tearing, voice shredded, guttural and harsh, grating rasp ripping, each scream cracking higher and more frantic. Panic-stricken, helpless, drowning in pain. Extreme arousal, extreme tension, violently volatile and out of control. Raw agonized screaming, throat tearing apart. (agonized scream) (terrified scream) High quality recording.
Procedural caption from re-predicted 57-dim profile
Extremely one-dimensional in resonance.
Extremely unpleasant and harsh.
Extremely young and childlike.
Extremely thin in head resonance.
Extremely scattered and disorganized.
Compute: 320 clips · 593 s · 0.539 clips/s
🗯️ Ranting Fury tuned
weights:AROU ×2.5S_RANT ×1.8TENS ×1.5VOLT ×1.5
An unhinged furious rant with Arousal/intensity up-weighted much more (vn_w AROU 2.5 / S_RANT 1.8 / TENS 1.5 / VOLT 1.5 on top of w_vn 3.0), while keeping the round-3 domineering-stance + wild-pitch-swing framing.
Round 4 up-weighted Arousal ×2.5 (S_RANT ×1.8, TENS/VOLT ×1.5) on top of the round-3 domineering-stance + wild-pitch framing. char_vn 0.80, char_emo 0.74.
Round-4 champion instruction
A boiling-over furious tirade, voice tearing and cracking with seething rage, spitting venom, aggressive and in-your-face, domineering and browbeating with contempt, refusing to yield. Harsh, tense, extremely volatile and high-arousal, explosive consonants, pounding relentless emphasis, pitch lurching violently high to low, escalating. Roaring but the words land. (enraged shout)(seething snarl) High quality recording.
Spoken text
“I am absolutely sick of this! How dare you waste my time again and again!”
A BREATHY evil ghost (new): airy, hollow, whispered menace, cold and sinister, an ethereal rasp full of dread. Rewards whisper-style + respiration + tension + rasp (vn_w S_WHIS/RESP 1.8, TENS 1.2), penalizes warmth/fullness/brightness, and up-weights Malevolence/Contempt/Fear emotions.
New archetype. A breathy, airy, hollow whispered menace. Rewards whisper-style + audible respiration + tension + rasp (S_WHIS/RESP ×1.8, TENS ×1.2), penalizes warmth/fullness/brightness, and up-weights Malevolence ×2.0 / Contempt ×1.5 / Fear ×1.2. char_vn 0.59, char_emo 0.58.
Round-4 champion instruction
A breathy evil ghost, airy and hollow, whispering menace through an ethereal rasp, all air and no body, cold and sinister. Seething with quiet malice and contempt, a dreadful whispered threat, tense and rasping. Wicked, cruel and menacing, dripping with dread. No warmth, no fullness, no brightness. Cold, sinister and dreadful, seething with quiet malice. (breathy hiss) High quality recording.
Spoken text
“You should not have come here... this house feeds on the living.”
A mourning woman crying softly, her voice trembling and breaking, choked and wet with tears, sniffling between words. Heartbroken, inconsolable, aching with deep sorrow and longing, helpless and disappointed. Softly feminine, raw and vulnerable, shaking and broken. (sniffle) High quality recording.
Spoken text
“Please don't leave me too... I can't lose anyone else.”
A very sad CRYING MAN (new): the same grief but a male voice cracking with sorrow and choked sobs. Same reward as sad_woman but rewards a masculine (low-feminine) gender and penalizes authority; Sadness up-weighted (2.5).
New archetype. A crying man: the same grief with a male voice cracking and choked sobs (rewards masculine gender). Sadness ×2.5. char_vn 0.62, char_emo 0.89.
Round-4 champion instruction
A man's voice cracking with sorrow, breaking and wavering, choked and unsteady, on the edge of sobbing. Heartbroken and guilt-stricken, aching with grief and longing, helpless and disappointed. Low and masculine, raw and vulnerable, fighting tears, faltering. (shaky breath) High quality recording.
Spoken text
“It's my fault... I should have been there, I should have saved her.”
“It's my fault. I should have been there. I should have saved her.”
Procedural caption from re-predicted 57-dim profile
Extremely raw and vulnerable.
Extremely denasal and clear.
Extremely collapsing in energy.
Extremely noisy and aperiodic.
Extremely poor in recording quality.
Compute: 327 clips · 603 s · 0.542 clips/s
Compute & timing (GPU 0 only)
Archetype
Clips
Seconds
Clips/s
👺 Evil Goblin
312
626
0.499
🐭 High-Pitch Mouse
328
645
0.508
😱 Pain Scream
320
593
0.539
🗯️ Ranting Fury
319
702
0.455
👻 Breathy Evil Ghost
327
625
0.523
😢 Crying Woman
320
596
0.537
😭 Crying Man
327
603
0.542
TOTAL
2253
4390
0.513
Totals across round-4 search generations + playoff + final best-of-64, all on a single GPU.
Round-4 lessons — weighted dims & emotions, and building new voices
Per-emotion weights are the sharpest new lever for ‘which feeling’. Up-weighting a single EmoNet head (Malevolence ×2.5 for the goblin, Sadness ×2.5 for the sad voices, Pain ×2.0 for the scream) reliably steered the search toward the intended affect without touching timbre. When one emotion defines the character, weight it heavily and let the others ride along at ×1.
Per-dimension weights concentrate char_vn where it matters. The mouse's RANG ×2.5 / REGS ×2.0 made the reward chase pitch height specifically, so the winner is measurably higher and squeakier; the scream's AROU ×2.5 pushed raw energy. Weighting 2-4 signature dims beats adding more unweighted dims — it stops a strong-but-off-target dim from dominating the mean.
Writing the affect into the TEXT does what the instruction alone cannot. Putting real screams (‘Aaah! ... Aaargh!’) in the spoken words made the pain-scream genuinely scream-like — vocalized interjections give the model something to actually scream, where a purely descriptive instruction over neutral words stays half-spoken. Keep a few real words so it is not pure noise (the ‘none’ WER gate allows it).
Malevolence is harder to raise than pitch or sadness. Even with Malevolence ×2.5, the goblin's absolute malice score stayed modest — the EmoNet malice head fires low on a nasal, giddy, cackling delivery; the up-weighting did lift it and push the prompt darker (sinister/sadistic/threatening), but a cute-shrill timbre and true menace pull against each other. Sadness and Fear were far easier to drive high.
The 3 new archetypes confirm the round-1→3 recipe transfers. Naming a concrete resonance/breath signature (breathy hollow whisper; audible sob/sniffle), a pitch/stance cue (low authority, trembling), and the target emotion in one prompt — then evolving with per-call normalization — produced clean ghost / crying-woman / crying-man voices from scratch, no reference audio.
Breath is a rewardable signature. The evil ghost leaned on S_WHIS + RESP (audible respiration): ‘all air and no body’, ‘breath rattling’ and (breathy hiss)/(hollow whisper) bursts raised the whisper/respiration dims and read as ethereal dread, while penalizing warmth/fullness/brightness kept it cold and hollow.
Same-call playoff still decides champions. As in every prior round, the highest single-generation candidate is not always the best under a fair head-to-head; re-scoring the strongest across generations under one normalization (best-of-8 playoff) then finalizing best-of-64 is what picks the published champion.
Gender is a cheap, reliable knob. Rewarding feminine gender (sad_woman) vs penalizing it (sad_man) split one grief prompt into two convincingly different voices — the GEND head responds strongly to a single ‘a man's voice’ / ‘softly feminine’ phrase.