Fish Audio releases Drama 3, an open-weight TTS model you direct like an actor Other

Fish Audio released Drama 3, an open-weight text-to-speech model for audiobooks, audio drama, games and dubbing that you direct like an actor. It speaks 83 languages and is available now on fish.audio, in the Fish Audio app and through the API, with the weights on Hugging Face.

You write the direction in square brackets inside the line, such as [whispering], [choked up, crying quietly] or [cracking up despite herself], and the voice changes exactly there, even halfway through a sentence. The direction can be in a different language from the line itself. Fish Audio announced it on October 13, 2026.

What it is

• Accurate under emotion. On Artificial Analysis' Pronunciation Robustness benchmark, Drama 3 read 93.2% of hard words correctly (brand names, numbers, codes, words whose reading depends on context), the top score on the leaderboard, against 88.1% for Fish Audio's previous S2.1 Pro.
• Architecture. A 35B-parameter mixture-of-experts with about 3B active per token, built on the language model of Qwen3.5-35B-A3B. It uses hybrid attention (Gated DeltaNet layers plus one full-attention layer in four) and a new codebook-MTP audio decoder.
• Tuning. Instead of rewarding speech that is easy to transcribe, it was tuned with a model of which takes people preferred, so a laugh or a whisper you asked for is not punished.

Price

The license also bans cloning a real person's voice without their consent.

Sources

• Fish Audio: Drama 3, the most emotive text-to-speech model we've built