Fish Audio open-weights Drama 3: 35B-A3B directable TTS, #1 on Artificial Analysis Pronunciation Robustness at 93.2%, 83 languages

Why it matters

【Key points】

Fish Audio released Drama 3 on 2026-10-13 as open weights: a directable TTS model for audiobooks, audio drama, game/animation dialogue and short video. Architecture: 35B MoE (~3B active per token; 8 of 256 routed experts plus one shared), backbone from Qwen3.5-35B-A3B; hybrid Gated DeltaNet + full attention and a new codebook-MTP audio decoder. Supports plain-language bracket directions, paired tags, multi-speaker scenes and 83 languages. Weights on Hugging Face (fishaudio/drama-3), code on GitHub (fishaudio/fish-speech). License: Fish Audio Community License v1.0 — free for research/non-commercial and commercial use under US$500k affiliate annual revenue; higher revenue or offering TTS/cloning/inference as a service needs a separate license; no cloning a real person's voice without consent.

Independent Artificial Analysis Pronunciation Robustness: 93.2% overall (vs 88.1% for S2.1 Pro), claimed #1. Live on fish.audio web, iOS/Android app and API (model header drama-3); promo ~$10 per million characters through Nov 12. Vendor-reported decode ~271 frames/s on one H200. Interactive editing and audio inpainting remain invite-only.

【Verifiable】

Blog: https://fish.audio/blog/fish-audio-drama-3/

Product: https://fish.audio/drama-3/

HF: https://huggingface.co/fishaudio/drama-3

GitHub: https://github.com/fishaudio/fish-speech

Tech report: https://github.com/fishaudio/fish-speech/blob/main/docs/assets/Fish_Audio_Drama_3_Technical_Report.pdf

【Our read】

1. Open voice AI is moving from “can speak” to “can be directed”: Drama 3 trains with a human-preference reward so laughs and whispers are not punished.

2. The 93.2% pronunciation score is third-party (Artificial Analysis); emotion still lacks a shared public eval.

3. Open weights ≠ open commerce: the $500k revenue cap and ban on offering TTS as a service will block many startups.

4. Local inference is heavy (≥48GB NVIDIA GPU for FP8 per docs); most creators will use the API first.