Fish Audio releases Drama 3, a 35B mixture-of-experts TTS for multi-character scenes

Fish Audio published Drama 3, which its card describes as a text-to-speech model "built for multi-character scenes and long-form content". The model is 35B parameters in total with 3B active per token, using a mixture-of-experts text backbone. The metadata lists 83 languages. Text can carry speaker tags from <|speaker:0|> to <|speaker:15|>, so up to 16 characters in one script, plus inline tags such as [sighs] and [laughs]. Voice cloning uses a short reference clip with a transcript, and long text is chunked automatically.

The card reports streaming speed on an NVIDIA H200: about 89 ms to first audio and a real-time factor of 0.131 at concurrency 1. It gives no memory or hardware requirements and no quality scores, and it notes results depend on hardware and sampling settings.

The licence is ambiguous on the card. The metadata says fish-audio-research-license while the licence section links to the "Fish Audio Community License"; check the licence file before use. The card has no release date. The date used is the Hugging Face repo creation date.

It is big, and the card does not say what you need to run it.

Sources

• Fish Audio model card on Hugging Face
• Hugging Face repo metadata (creation date)

Editor's first pass (AI-assisted)

• Real 3.5: Weights are on Hugging Face but a 35B model needs serious hardware and the card gives no requirements.
• Useful 3.5: Multi-character long-form scenes are a real use case.
• Notable 3.5: A large open TTS from a known maker; no quality benchmarks on the card.
• Novel 3.5: Mixture-of-experts at this scale is uncommon for open TTS.