Fish Audio's Drama 3 brings MoE to open TTS

Fish Audio has released Drama 3, an open-source text-to-speech model built on a mixture-of-experts architecture. The model carries 35 billion parameters in total but activates roughly 3 billion per inference pass, a design that aims to deliver the quality of a large model while keeping the compute cost of generation closer to that of a much smaller one.

Beyond raw synthesis, Drama 3 is pitched around two capabilities that matter for expressive speech: voice cloning and directed performance. The latter suggests finer control over how a line is delivered — tone, emotion, and pacing — rather than just converting text to flat narration. That makes it a candidate for use cases like audiobooks, dubbing, and interactive characters, where delivery is as important as intelligibility.

Why it matters

Mixture-of-experts has reshaped open large language models over the past two years, but it remains relatively uncommon in speech generation. Applying the approach to TTS is a notable bet that the same efficiency gains can carry over to audio.

• Multilingual: the model lists Chinese, English, and Japanese among supported languages, with more indicated.
• Efficient by design: only about 3B of 35B parameters are active per pass.
• Open weights: the model is published on Hugging Face, though under a custom "other" license worth reviewing before commercial use.

As with any open release, the practical question is how Drama 3 holds up against both proprietary speech APIs and other open TTS efforts. For teams already building on open audio stacks, an efficient MoE model with explicit performance controls is a meaningful new option to evaluate.