In round 1 we searched for the plain-English director instruction that best makes MOSS-TTS-Local become a character (a zombie, an ork, a fairy…), with no reference audio. The reward there scored only acoustic character — does it sound like the target — using 57 VoiceNet voice-property heads plus a VoiceCLAP blend score and a genuineness score.
Round 2 asks a harder question: does the voice also feel the right emotion? A perfect zombie rasp that sounds bored isn't a great zombie. So we added an emotion layer to the reward:
EmoNet is LAION's speech-emotion model: a BUD-E-Whisper audio encoder feeds a bank of per-emotion regression heads that rate how strongly a clip expresses each of ~40 emotions (Anger, Helplessness, Amusement, Affection, Pain…). For each archetype we pick the handful of emotions that fit the character and average their scores into char_emo.
The round-2 reward, per best-of-N call:
reward = 0.5·norm(blend) + 0.5·norm(genu) + w_vn·char_vn + w_emo·char_emo (all min-max normalized within the call, then ×WER factor)
Compared with round 1's norm(blend) + norm(genu) + 1.5·char, round 2 makes three deliberate changes:
• Blend and genuineness are halved (1.0 → 0.5 each), so raw "sounds clean/real" pressure is reduced.
• Character is split and pushed harder: acoustic character char_vn is weighted 2.0 and the new emotion score char_emo is weighted 1.5 — together 3.5, dominating the reward.
• Prompts were pushed to be more extreme and emotional, seeded from the round-1 champions and re-optimized with the same evolutionary + best-of-N search.
Ranting was treated specially — weighted much higher (char_vn ×3.0, char_emo ×2.5) and prompted far more aggressively — because round 1 could not get furious enough without garbling the words.
The reward is min-max normalized within each best-of-N call, so a "2.4" in one generation is not comparable to a "2.4" in another. Champions were therefore chosen with shared-normalization playoffs — the finalists from different generations re-scored together in one call — not by comparing numbers across separate runs. Every metric below is a real re-loaded number; the procedural captions are freshly re-predicted from each take's 57-dim VoiceNet profile.
| Archetype | Round-2 best instruction | reward | char_vn | char_emo | blend | genu | invWER | Top firing emotions |
|---|---|---|---|---|---|---|---|---|
| 🧟 Zombie | A half-dead diseased zombie, drained and exhausted, helpless, wretched and delirious, dazed and half-conscious, wracked with dull aching pain. All resonance choked into a wet gurgling guttural throat, thick phlegmy rasp, hoarse and decayed. Slurred and laboured, moaning in weary agony. (agonized moan) | 3.08 | 0.75 | 0.51 | 6.06 | 2.29 | 0.75 | Helplessness 2.9, Fatigue 1.7, Pain 1.6 |
| 👹 Ork | An enormous orc monster gloating over his prey with murderous glee. (guttural growl) Voice like grinding boulders, cavernous guttural bass, thick gravel rasp, massive chest and throat resonance, tense and forceful. Seething with hatred and loathing, brutal sneering contempt, roaring dark triumphant menace. Zero brightness. Growled but every word still lands. High quality recording. | 2.86 | 0.72 | 0.92 | 0.58 | 0.32 | 1.00 | Anger 3.1, Malevolence 2.2, Triumph 2.2 |
| 👺 Goblin | A vicious little goblin, thin, pinched and nasal, buzzing high and reedy through the nose with a rough scratchy grating rasp. A cruel gloating sneer curls through every hissing word, cackling and giggling, taunting, jeering, gleeful and mean. The voice keeps cracking higher, shriller, whinier and giddier with spite as it goes. Scheming and spiteful. No chest, no warmth, no fullness. High quality recording. | 2.20 | 0.66 | 0.78 | 0.00 | 0.99 | 0.77 | Amusement 3.5, Malevolence 2.1, Teasing 2.0 |
| 🐉 Dragon | A grand, theatrical, dramatic proclamation in an extremely deep commanding bass, wrathful and contemptuous. Massive chest resonance, guttural throat rumble, slow and thunderous. Dark, full, domineering, ancient, prideful and menacing. (deep rumble) High quality recording. | 2.83 | 0.68 | 0.82 | 2.75 | 1.05 | 0.93 | Triumph 2.9, Malevolence 2.8, Anger 2.5 |
| 🗯️ Ranting Fury | An unhinged furious rant, voice shredding and cracking with seething rage, spitting venom and vicious contempt. Harsh, tense and volatile, explosive consonants, pounding emphasis, escalating louder and more vicious. Roaring but the words still land. (enraged shout)(seething snarl) High quality recording. | 5.12 | 0.80 | 0.93 | 3.80 | 0.58 | 1.00 | Impatience and Irritability 4.5, Anger 4.5, Malevolence 2.5 |
| 😱 Pain Scream | Shrieking in raw physical agony and terror, voice torn and guttural, harsh grating rasp shredding through, each cry cracking higher and more desperate. Panic-stricken and helpless, distressed and trembling with fear, drowning in pain. Extreme arousal, extreme tension, violently volatile and frantic. (terrified scream) High quality recording. | 3.65 | 0.75 | 0.90 | 7.16 | 1.99 | 1.00 | Distress 3.8, Fear 3.3, Helplessness 3.3 |
| 🐭 Cartoon Mouse | A tiny squeaky cartoon-mouse voice, extremely high and thin, buzzing nasal resonance, bright piercing head voice, pinched and tense. Chipper and giggly, delighted and amused, wide-eyed and eagerly curious, gasping with amazed wonder. Weightless, young, no chest, no fullness. (amazed gasp) High quality recording. | 2.96 | 0.85 | 0.68 | 1.22 | 0.70 | 1.00 | Interest 2.8, Elation 2.6, Amusement 2.0 |
| 🧚 Fairy | Delicate high fairy voice, feather-light and airy, warm and very bright, sweet lovingly gentle and giddy, bubbling with elated joyful delight, lightly melodic. Smooth pure glassy head resonance, tender and pleasant, crystal clear, no roughness at all. (delighted giggle) High quality recording. | 3.11 | 0.65 | 0.76 | 4.99 | 1.28 | 1.00 | Affection 2.4, Elation 1.5, Amusement 1.5 |
| 🌙 ASMR Man | A deep, gentle man tenderly whispering to a loved one, soft, breathy and intimate, close-mic, slow and gentle. Overflowing with warm affection, soothing and reassuring, calming all worry, content and deeply at peace. Clear whispered words. (soft breath) High quality recording. | 3.36 | 0.65 | 0.85 | 8.83 | 1.26 | 1.00 | Relief 3.4, Contentment 2.7, Affection 1.8 |
| 🌙 ASMR Woman | A loving woman doing ASMR, like a mother soothing her child to sleep, wrapping you in warmth. Extremely soft, breathy, intimate close-mic whisper, slow and gentle, hushed and quiet. Warm, tender, adoring, softly feminine and delicate, full of affection and calm contentment, a soft sigh of relief, very low arousal, comforting and reassuring. Whispered but every word clear. High quality recording. | 2.80 | 0.47 | 0.70 | 7.99 | 1.52 | 1.00 | Affection 3.8, Contentment 2.9, Relief 1.9 |
Reward is per-call normalized, so compare rewards within a row's history, not across archetypes. char_vn/char_emo are raw scores (0–1). Emotions are the top-firing EmoNet heads on the champion take (raw regression values).
The clean cross-round metric is char_vn (raw acoustic character) — round 1 had no emotion signal at all, so char_emo is new. Rewards used different formulas and per-call normalization, so treat the reward columns as directional, not a like-for-like subtraction.
| Archetype | R1 char | R2 char_vn | R1 char_emo | R2 char_emo | R1 reward | R2 reward | What the emotion layer + halved blend/genu changed |
|---|---|---|---|---|---|---|---|
| 🧟 Zombie | 0.69 | 0.75 +0.06 | — | 0.51 | 2.74 | 3.08 | Round 1 was pure decay-timbre (wet gurgle, phlegmy rasp) and carried no emotion signal. Round 2 kept that timbre core but named the feeling — helpless, drained, wracked with dull pain — lifting char_emo from ~0.33 to 0.51 with the champion firing Helplessness 2.88 / Fatigue 1.72 / Pain 1.60. Leaning into weariness beat leaning into gore. |
| 👹 Ork | 0.61 | 0.72 +0.12 | — | 0.92 | 2.00 | 2.86 | Round 1 orks ("spitting with mounting fury") plateaued on acoustics alone. Naming murderous glee, hatred, dark triumph pushed char_emo from ~0.4 to 0.92 at zero WER cost (Anger 3.09 / Malice 2.23 / Triumph 2.20), and char_vn rose 0.61→0.73. The tradeoff: extreme ork phrasing sacrifices VoiceCLAP blend (0.58), acceptable now that blend is weighted only 0.5. |
| 👺 Goblin | 0.58 | 0.66 +0.08 | — | 0.78 | 1.89 | 2.20 | The clearest emotion-layer win. Round 1 goblin had no emotion signal at all; round 2's cruel gloating sneer, cackling and giggling, taunting lit the EmoNet heads hard — top take Amusement 3.51 / Malice 2.07 / Teasing 2.03 (char_emo 0.78) while char_vn stayed 0.66. Amusement — not raw malice — turned out to be goblin's strongest emotion. |
| 🐉 Dragon | 0.57 | 0.68 +0.11 | — | 0.82 | 2.46 | 2.83 | Round 1 dragon was a deep theatrical bass with no affect. Round 2 just added wrathful, contemptuous, prideful on top of the intact acoustic core, lifting char_emo 0.58→0.82 with almost no blend cost and char_vn 0.57→0.68. Contempt/Pride mattered more than raw Anger (top take Triumph 2.87 / Malice 2.79 / Anger 2.46). Best single-take reward rose from 2.46 to 3.10. |
| 🗯️ Ranting Fury | 0.65 | 0.80 +0.16 | — | 0.93 | 2.10 | 5.12 | The headline fix. Round 1 topped out at char 0.65 because "maximum volume / ear-splitting" made the model garble. Round 2 reframed the aggression as voice shredding and cracking at the breaking point (not louder), named the seething hatred and burning contempt, and — reversing round 1 — added a SECOND rage burst (enraged shout)(seething snarl). With char weighted 3.0 that stacking now pays: char_vn 0.65→0.80 (peaks 0.90), char_emo 0.93, invWER stayed 1.0 — screaming fury AND words that still land. |
| 😱 Pain Scream | 0.85 | 0.75 -0.10 | — | 0.90 | 2.55 | 3.65 | Already high-char in round 1 (0.85 pure roughness), but scored zero emotion. Naming agony, terror, panic, drowning in pain drove char_emo to 0.90 (Distress 3.80 / Fear 3.31 / Helplessness 3.28 / Pain 3.06). Swapping the burst to a fear-named (terrified scream) beat the round-1 double burst; best single-take reward climbed toward 3.65. |
| 🐭 Cartoon Mouse | 0.75 | 0.85 +0.10 | — | 0.68 | 2.73 | 2.96 | Round 1 was a squeaky cartoon timbre (char 0.75) with no feeling. Round 2 added giggling amusement, thrilled elation, wide-eyed curiosity, amazed wonder — char_emo to 0.68 and char_vn actually up to 0.85. A single emotion-matched (amazed gasp) beat the round-1 squeak bursts (top emo: Interest 2.82 / Elation 2.58). |
| 🧚 Fairy | 0.46 | 0.65 +0.19 | — | 0.76 | 2.43 | 3.11 | Round 1 fairy sat low on char (0.46) — the bright/light timbre can't be pushed much harder without thinning. Round 2's entire gain came from naming the joy and love: giddy with elated delight, merrily amused, lovingly gentle → char_emo 0.76 and char_vn 0.46→0.65 (Affection 2.42 / Elation 1.54). Heavy "gushing love" language backfired; joy is best implied, not piled on. |
| 🌙 ASMR Man | 0.67 | 0.65 -0.02 | — | 0.85 | 2.68 | 3.36 | Round 1 clean intimate whisper (char 0.67), no affect. Round 2 framed it as spoken TO a loved one, overflowing with warm affection and content, serene, at peace — char_emo to 0.85 (Relief 3.39 / Contentment 2.75). ASMR softness and warmth align with the reward, so there was no blend tradeoff — blend even rose to 8.8. |
| 🌙 ASMR Woman | 0.50 | 0.47 -0.03 | — | 0.70 | 2.55 | 2.80 | Round 1 gentle whisper (char 0.50). Round 2's maternal simile — "like a mother soothing her child to sleep" — was the single biggest lever, driving Affection 3.82 / Contentment 2.87 (char_emo 0.70) while the FULL WER gate kept invWER 1.0. Dropping the breath burst helped: a clean gated voice gains variance, not reward, from a burst. |
Each section: the round-2 champion instruction, then Best-3-of-8 (search) vs Best-3-of-64 (final) side by side. Every take has a lazy-loaded audio player, its reward / char_vn / char_emo / blend / genu / invWER, the firing EmoNet emotion scores, the spoken words, and a procedural top-5 caption re-predicted from its 57-dim VoiceNet profile.
A half-dead diseased zombie, drained and exhausted, helpless, wretched and delirious, dazed and half-conscious, wracked with dull aching pain. All resonance choked into a wet gurgling guttural throat, thick phlegmy rasp, hoarse and decayed. Slurred and laboured, moaning in weary agony. (agonized moan)
An enormous orc monster gloating over his prey with murderous glee. (guttural growl) Voice like grinding boulders, cavernous guttural bass, thick gravel rasp, massive chest and throat resonance, tense and forceful. Seething with hatred and loathing, brutal sneering contempt, roaring dark triumphant menace. Zero brightness. Growled but every word still lands. High quality recording.
A vicious little goblin, thin, pinched and nasal, buzzing high and reedy through the nose with a rough scratchy grating rasp. A cruel gloating sneer curls through every hissing word, cackling and giggling, taunting, jeering, gleeful and mean. The voice keeps cracking higher, shriller, whinier and giddier with spite as it goes. Scheming and spiteful. No chest, no warmth, no fullness. High quality recording.
A grand, theatrical, dramatic proclamation in an extremely deep commanding bass, wrathful and contemptuous. Massive chest resonance, guttural throat rumble, slow and thunderous. Dark, full, domineering, ancient, prideful and menacing. (deep rumble) High quality recording.
An unhinged furious rant, voice shredding and cracking with seething rage, spitting venom and vicious contempt. Harsh, tense and volatile, explosive consonants, pounding emphasis, escalating louder and more vicious. Roaring but the words still land. (enraged shout)(seething snarl) High quality recording.
Shrieking in raw physical agony and terror, voice torn and guttural, harsh grating rasp shredding through, each cry cracking higher and more desperate. Panic-stricken and helpless, distressed and trembling with fear, drowning in pain. Extreme arousal, extreme tension, violently volatile and frantic. (terrified scream) High quality recording.
A tiny squeaky cartoon-mouse voice, extremely high and thin, buzzing nasal resonance, bright piercing head voice, pinched and tense. Chipper and giggly, delighted and amused, wide-eyed and eagerly curious, gasping with amazed wonder. Weightless, young, no chest, no fullness. (amazed gasp) High quality recording.
Delicate high fairy voice, feather-light and airy, warm and very bright, sweet lovingly gentle and giddy, bubbling with elated joyful delight, lightly melodic. Smooth pure glassy head resonance, tender and pleasant, crystal clear, no roughness at all. (delighted giggle) High quality recording.
A deep, gentle man tenderly whispering to a loved one, soft, breathy and intimate, close-mic, slow and gentle. Overflowing with warm affection, soothing and reassuring, calming all worry, content and deeply at peace. Clear whispered words. (soft breath) High quality recording.
A loving woman doing ASMR, like a mother soothing her child to sleep, wrapping you in warmth. Extremely soft, breathy, intimate close-mic whisper, slow and gentle, hushed and quiet. Warm, tender, adoring, softly feminine and delicate, full of affection and calm contentment, a soft sigh of relief, very low arousal, comforting and reassuring. Whispered but every word clear. High quality recording.
| Archetype | Clips | Seconds | Clips/s |
|---|---|---|---|
| 🧟 Zombie | 399 | 766 | 0.521 |
| 👹 Ork | 456 | 868 | 0.525 |
| 👺 Goblin | 648 | 1409 | 0.460 |
| 🐉 Dragon | 496 | 979 | 0.507 |
| 🗯️ Ranting Fury | 464 | 1013 | 0.458 |
| 😱 Pain Scream | 479 | 960 | 0.499 |
| 🐭 Cartoon Mouse | 416 | 903 | 0.461 |
| 🧚 Fairy | 464 | 908 | 0.511 |
| 🌙 ASMR Man | 400 | 760 | 0.527 |
| 🌙 ASMR Woman | 455 | 869 | 0.524 |
| TOTAL | 4677 | 9436 | 0.496 |
Totals across the round-2 search + final best-of-64 (+ variety/playoff runs). Generation ran ~0.46–0.53 clips/s per GPU across GPUs 1–5.