Fish Audio Releases Drama 3, a 35B Open-Weight MoE Text-to-Speech Model

Fish Audio released Drama 3, a 35-billion-parameter mixture-of-experts text-to-speech model with about 3 billion active per token. Writers add plain-language acting cues in brackets, and up to 16 speakers can render in a single request across 83 languages. Weights are on Hugging Face under a community license that's free for commercial use below $500,000 in annual revenue and prohibits cloning a real voice without consent.

Fish Audio reports it ranks first on Artificial Analysis' pronunciation robustness test at 93.2%, though it trails ElevenLabs' Eleven v4 Turbo on the quality leaderboard. Self-hosting needs at least a 48GB NVIDIA GPU.

For small studios, podcasters and accessibility projects, this is a meaningful release. Directing a scene in plain language and rendering a whole cast in one pass, on hardware you control, used to be the preserve of expensive closed APIs.

The open weights cut both ways. The license bans non-consensual voice cloning, but a license is a promise, not a lock. Once weights are public, enforcement depends on goodwill.

There's also a labor story. Voice actors in audiobooks, dubbing and games have been warning about exactly this: expressive, directable synthetic performance at a price that undercuts sessions. Better tools for creators can coexist with fair treatment of performers, but only if studios choose it.

We'd call it a strong technical step with a familiar set of social costs, and I'm watching how voice-actor unions and platforms respond.