MOSS Voice-Acting — Manual & Studies

💥 Newest (Aug 2026) — bursts, merge doses & evaluation: for a vocal burst inside a sentence, λ = 0.75 is the knee — 91 % of full-merge burst presence at a much better blend, and it is where genuineness peaks; burst@1.0 + emotion@0.5 is the best cell of 2,304 generations, while a half-merged burst adapter paired with a half-merged emotion adapter is the worst (they compete). Cross-lingual reference audio (Japanese anime → DE/EN) does not transfer the voice — similarity 0.188 over 1,280 generations against a 0.105 unrelated-speaker floor — but pitch still tracks the reference (r = 0.591, compressed toward the model's own register), which makes it a cheap source of new character voices rather than a cloning method. And always evaluate against a real-audio ceiling control: a 13-cell sports-commentator sweep looked like a win until real human commentary scored the same as the un-adapted base. Read the page →

🧪 What's new (Aug 2026): an autonomous search agent now discovers recipes on its own. Key validated learnings, folded into the manuals: the reward is multiplied by (1−WER) to stop over-acted, unintelligible takes; at high energy, add the stabilizer adapters vn_ARSH_high@0.5 + vn_BRGT_high@0.5 (and vn_S_NARR_high for storytelling) to keep speech clear; and a cheap API brain (gpt-5.6-luna) matches the local model at cents per task. Example — a favourite audiobook-narrator recipe: vn_S_STRY_high@0.65 + vn_ARSH_high@0.20 + vn_BRGT_low@0.20 → storytelling 0.998, aesthetics 0.59, WER 0.000, ~11 s. Brain used in the comparison: google/gemma-4-31B-it-qat-w4a16-ct (see MODEL_COMPARISON.md).