MOSS-TTS-4.55B voice-acting-v2 — Conditioning Manual

A measured usage guide for laion/moss-tts-local-transformer-4.55b-voice-acting-v2.

• Human / audio version (GitHub Pages): https://laion-ai.github.io/moss-voiceacting-manual/
• Agent-skill (Markdown, no audio): md/index.md

For each target: the best prompting strategy with and without a LoRA, the best merge %, the expected reward = (genuineness + vocal_burst_blend + 1.25·target_emotion) × (1 − min(WER, 1)), and the measured side-effects / correlations on other EmoNet emotions, VoiceNet dimensions, genuineness, vocal-burst blend and speech quality. Data: per target × 5 conditions × 32 samples on neutral sentences.

The reward above was previously written as / (1 + WER). That form is wrong and has been corrected here: about three quarters of candidates have a negative core score, and dividing a negative number by a larger denominator makes it larger — so the division form rewards transcription errors on most of the pool. Confirmed twice, most recently on a 1,000-clip control set where a literal WER × quality filter put 602 of 1,000 candidates at exactly 0. Use × (1 − min(WER, 1)) on sigmoid-squashed (strictly positive) components.

Sections: Emotions (40) — ready · Edge cases · VoiceNet dimensions · Character voices (planned) · Vocal bursts, merge doses & how to evaluate a LoRA — new · Expressive & NSFW recipe book.

tools/md2site.py renders md/**.md into the site's HTML style, so a page's Markdown mirror and its HTML cannot drift apart.