TL;DR. Best reward is WITH the LoRA at 50% (0.583, beating BASE 0.519). This is the one setting that actually delivers real target strength — emo jumps from 0.029 (prompt-only) to 0.235 — while keeping vocal-burst blend intact at 0.45. The cost is intelligibility and quality: WER roughly doubles (0.269 to 0.558) and quality drops to 0.445. Numbness surfaces as flattened arousal and emphasis (corr neg Arousal -0.79, Emphasis -0.80) plus rising disfluency (corr 0.77); do NOT push to 100/150%, where emo climbs but WER explodes to ~1.0 and reward collapses.
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.52
GENERAL: A voice utterly expressing emotional numbness, a flat, hollow, emotionally numb voice, detached and empty, drained of all feeling, in every breath.
SCRIPT: (flatly, hollow and emotionally numb) "<neutral sentence>" · temp 0.8 top_p 0.9 top_k 40
Best reward is WITH the LoRA at 50% (0.583, beating BASE 0.519). This is the one setting that actually delivers real target strength — emo jumps from 0.029 (prompt-only) to 0.235 — while keeping vocal-burst blend intact at 0.45. The cost is intelligibility and quality: WER roughly doubles (0.269 to 0.558) and quality drops to 0.445. Numbness surfaces as flattened arousal and emphasis (corr neg Arousal -0.79, Emphasis -0.80) plus rising disfluency (corr 0.77); do NOT push to 100/150%, where emo climbs but WER explodes to ~1.0 and reward collapses.