Fish Audio ships Drama 3, an open-weight text-to-speech model
A 35B mixture-of-experts voice model follows plain-language direction mid-sentence and ships free for commercial use under $500,000 in annual revenue.
Fish Audio released Drama 3 on 13 October, an open-weight text-to-speech model built for directed performance rather than flat narration. It’s a 35B-parameter mixture-of-experts model with about 3B active parameters per token, initialised from the Qwen3.5-35B-A3B language model, and it runs at 271 frames per second on a single H200, against 177 for Fish’s previous model, S2.
You direct it with square-bracket stage notes in plain language, and it can shift mid-sentence between tears, suppressed anger or laughter, across 83 languages. Audio inpainting lets you regenerate a single word in an existing take without re-recording the rest. Weights are on Hugging Face and the code is on GitHub, under the new Fish Audio Community License v1.0: free for research, and free for commercial use if your organisation and its affiliates bring in under $500,000 a year. Above that, or if you’re offering it as a hosted TTS or voice-cloning service, you need a commercial licence. The licence also bars voice cloning without consent.
Why it matters: this is a self-hostable voice model with real directorial control, not just a sentence-to-speech pipeline, and the licence puts it within reach of small teams building audiobooks, games or video dialogue without a per-call API bill.
The caveat: the only benchmark cited, Artificial Analysis’s pronunciation robustness score, is a single third-party number (93.2% versus 88.1% for the previous model), not a broad independent evaluation.