Fish Audio Drama 3 final weights top open TTS arenas

The numbers, per Artificial Analysis. On the Provider Voice arena Drama 3 scores an Elo of 1,278 over 1,512 appearances: #6 overall, #1 among open-weights models, and 38 points above the Drama 3 Preview. On Controlled Voice it's #12 overall and again #1 open-weights at 1,085. It tops their Pronunciation Robustness test at 94.6%, generates about 66.6 characters per second, and the hosted API is $15 per million characters. Every model above it on Provider Voice costs more, from Gemini 3.8 Flash TTS at $16.5 to Eleven v4 at $80 per million.

What's in the model, per the card. It's a 35B mixture-of-experts model with about 3B active per token, FP8 experts, and speaker tags for up to 16 characters in one script. Long text is chunked and streamed automatically. Point tags like [sighs] or [laughs] go where the sound should land. Editing works as continuation: send the earlier takes as context along with the corrected line, and the card's samples show the new read keeping the same voice and delivery, which suits pickups and script fixes. Fish Audio's own H200 tests show first audio in under 100 ms for a single stream.

How to run it. The card shows the weights served through SGLang-Omni or vLLM-Omni, each with a one-line serve command, and the community ComfyUI-Drama3 node runs the earlier release from 8 GB of VRAM by keeping experts in system RAM. We haven't checked whether that node picks up the new weights yet. The licence is the Fish Audio Community License; read it before commercial use. Clone only voices you have permission to use.

Our take: if you voice explainers, ads or short dramas with open models, Drama 3 is now the default to test against ElevenLabs. The edit-by-continuation trick is the practical win for video, because a script fix can match the take you already cut instead of starting from a fresh read.