Fish Audio ships Drama 3 Preview, an open-weights TTS model
- Source
- x.com
- Date
Drama 3 Preview is the #1 open-weights TTS model on @ArtificialAnlys 🎭 Beats Eleven v4 on pronunciation. 5x cheaper. 83 languages. Weights on HF, live on our API today. Best-in-class voice AI, for everyone. And we're just getting started. https://t.co/JFEqj2uhpk
@Mqorbanii @mochacortado @ArtificialAnlys Hey, the 1-month free was set for the first 48 hours after the launch. Please check your dm!
@mochacortado @ArtificialAnlys please check your dm!
Context
Fish Audio's claim that Drama 3 Preview is the top open-weights text-to-speech model comes from independent tests by Artificial Analysis, which the announcement links to. means the model files themselves can be downloaded, here from Hugging Face under the Fish Audio Community License, rather than only reached through the company's paid service. Artificial Analysis describes Drama 3 Preview as a model with 35 billion parameters, of which only 3 billion are active at any one step, supporting 83 languages. It is built for multi-character scenes and long-form content, with voice cloning and control through plain-language directions. Fish had earlier previewed that control as directing tone, emotion and pacing in ordinary words, even changing emotion partway through a line.
The ranking rests on three leaderboards that measure different things. The one Drama 3 Preview leads outright is Pronunciation Robustness, a test Artificial Analysis introduced in September to check whether a voice says tricky text correctly, for example reading St. as Saint in one place and Street in another. In its results post Drama 3 Preview scores 93.2%, ahead of ElevenLabs' Eleven v4 at 91.7% and Eleven v4 Turbo at 90.1%; this is the basis for Fish's line that it beats Eleven v4 on pronunciation. The next open-weights model, Mistral's Voxtral TTS, scores 78.1%.
The other two leaderboards are arenas where listeners compare voices, and scores are Elo ratings, a system borrowed from chess that rises when a model wins head-to-head comparisons. There Drama 3 Preview is the best open-weights model but not the best overall: it ranks 10th in the Provider Voice arena at 1,243, a gain of 101 points over Fish's previous S2.1 Pro, and 14th in the Controlled Voice arena at 1,058. Fish's 5x cheaper claim matches the prices Artificial Analysis lists: $15 per million characters through the Fish Audio API, the same as S2.1 Pro, against $80 for Eleven v4 and $40 for Eleven v4 Turbo. Fish says the model is live on its API now. The license terms, which decide how freely the weights can be used, were not part of the material read here.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
Checking sign-in…
Loading comments…



