Evolving prompt+sampling to push VoiceNet dimension S_TECH in the low direction with the vn_S_TECH__low LoRA. Steering phrase: a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
Best in-direction score 0.922 at 100% vs neutral 0.723 (shift +0.200).
Method & reproducibility — how the evolution ran
Base model:laion/moss-tts-local-transformer-4.55b-voice-acting-v2 · LoRA: the new VoiceNet-dimension LoRAvn_<CODE>__<high|low> (retrained with ≤25% game-character data, rank 32), applied via PEFT adapter scaling at strength λ. Algorithm: genetic search — population 12, 10 generations, elitism (top 5). During evolution every genome is generated best-of-6with the LoRA at λ=100%; its fitness = the mean over those 6 candidates of the target VoiceNet dimension score in the intended direction (for a __high LoRA the raw score; for __low, 1−score), minus 0.2·max(0, WER−0.6). Genome (what evolves): the GENERAL voice description, the SCRIPT cue, the carrier sentence, and the sampling params (temperature, top_p, top_k). λ is fixed at 100% during the search. Mutation operators: 40% regenerate description, 25% swap carrier, 17% jitter temperature ±0.1, 18% change top_k. Offspring = 60% crossover + mutation, 40% mutation. Dose measurement (below): the winning recipe is then re-generated at λ=0 (LoRA OFF), λ=50% and λ=100% (each best-of-6) to show exactly what 50% vs 100% of the LoRA brings over the no-LoRA baseline.
Dose — what 50% vs 100% brings (top recipe)
λ=0 (no LoRA) — dir 0.879 (raw 0.12)
WER 0.17 · genu 0.15
λ=50% — dir 0.933 (raw 0.07)
WER 0.37 · genu 0.19
λ=100% — dir 0.922 (raw 0.08)
WER 0.00 · genu 0.07
GENERAL: A voice that is markedly a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
SCRIPT: (intensely a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style) "I need to stop by the store and pick up a few things on the way home." · temp 1.1 top_p 0.95 top_k 40
Other evolved recipes (100%)
— dir 0.915 (raw 0.09)
WER 0.33 · genu 0.15
A voice that is markedly a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
— dir 0.908 (raw 0.09)
WER 0.33 · genu 0.10
A voice that is clearly a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
— dir 0.900 (raw 0.10)
WER 0.50 · genu 0.15
A voice that is markedly a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
— dir 0.893 (raw 0.11)
WER 0.17 · genu 0.15
A voice that is clearly a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
— dir 0.893 (raw 0.11)
WER 0.00 · genu 0.15
A voice that is strongly a voice with markedly minimal teacher/didactic style, nearly the absence of teacher/didactic style.
