Evolving extreme genuine vocalizations with emotion LoRAs

An evolutionary search over prompts + inline vocal-burst tags + LoRA strength + sampling, per emotion LoRA, to elicit extreme edge-case expressions — screaming, groaning, mild genuine laughter, crying/whimpering, shivering — on moss-tts-4.55b-voice-acting-v2. Population evolves over generations (elitism + crossover + mutation); fitness = EmoNet target + vocal-burst blend + genuineness + DSP proxies. Each run is compared against the same prompt without the LoRA (λ=0) and a neutral baseline, to isolate the LoRA's contribution. Each target page has a Method & reproducibility box and marks every clip as LoRA ON or LoRA OFF.

Method & reproducibility — how the evolution ran
Base model: laion/moss-tts-local-transformer-4.55b-voice-acting-v2 · LoRA: the matching v3 emotion LoRA (rank 32) applied via PEFT adapter scaling at strength λ.
Algorithm: genetic search — population 24, 6 generations, elitism (top 8). Each genome is generated best-of-4; its fitness is a multi-objective score for the edge case = the EmoNet target + vocal-burst blend + genuineness + DSP proxies (RMS loudness, median pitch, pitch jitter, spectral flatness for screams, 3–9 Hz amplitude tremor for crying/shivering); the best-fitness candidate represents the genome. Offspring = 60% crossover of two elites + mutation, 40% mutation of one elite.
Genome (what evolves): the GENERAL voice description, the SCRIPT cue, the inline vocal-burst tags (e.g. <screams>, <laughs softly>, <sobbing>), the carrier line, the LoRA strength λ, and the sampling params (temperature, top_p, top_k).
Mutation operators: 30% new description, 20% add/remove a vocal-burst tag, 15% swap carrier, 17% jitter λ by ±0.25/0.5, then temperature jitter.
LoRA on vs off: the two LoRA OFF cards are baselines with the adapter disabled (λ=0) — a neutral prompt and the best edge prompt without the LoRA. Every LoRA ON hall-of-fame recipe was generated with the LoRA active at the evolved λ shown on the card.
edge caseLoRAbest fitneutralprompt no-LoRALoRA liftλ*top burst-tags
Fear → ScreamingFear0.7670.2450.387+0.3800.5<screaming> (8), <terrified scream> (5), <screams> (3), <panicked breath> (1)
Pain → ScreamingPain0.6990.2840.506+0.1930.75<cry of pain> (9), [groan] (3), <screams in pain> (2), <agonized scream> (1)
Sadness → Crying / whimperingSadness0.7400.2410.639+0.1010.75<voice breaks> (7), <crying> (5), <sobbing> (4), <whimpers> (3)
Cold → ShiveringFear0.7720.5020.686+0.0860.75[shivers] (8), <shivering> (4), <trembling breath> (3), <teeth chattering> (2)
Pain → GroaningPain0.6360.3330.571+0.0660.75<pained grunt> (5), <heavy breathing> (4), [groan] (4), <low groan> (3)
Amusement → Mild genuine laughterAmusement0.4640.0900.406+0.0581.5<warm laughter> (5), <giggles> (5), <chuckles> (4), <laughs softly> (2)

Conclusions