Fish Audio Drama-3: 36B TTS model trending on Hugging Face
Fish Audio's open-weight 35B-A3B MoE TTS for multi-character scenes and long-form content: plain-language voice direction, mid-sentence voice shifts, and single-word editing.
Why it matters
Direct a voice performance in plain language instead of generating it: shift emotion mid-sentence, generate 16-character scenes in one output, and fix a single word without re-rendering - now as open weights.
What you could build with it
An indie game studio could prototype full character dialogue with Drama-3's expressive multi-speaker synthesis before hiring voice actors. The result: playtested narrative pacing with real-sounding voices at zero recording cost.
Does it hold up?
Usable today for dialogue and long-form work: tryable in the Fish Audio playground and API ($50/1M chars, 80% off until Nov 12), day-one SGLang serving, and a ComfyUI Partner Node; independent technical analysis confirms the architecture claims, though independent quality-of-controllability benchmarks are still missing.
Built with Fish Audio Drama-3
- Fish Audio Drama 3: a 35B MoE TTS whose codebook decoder attends over all past frameserogol.com · Independent 9-minute technical teardown (Oct 14) of the arXiv report: hybrid Gated DeltaNet + MoE backbone, 13-layer codebook-MTP decoder, RL with one learned preference reward, decode at 271 frames/s on H200.
- Fish Audio Drama 3: Most Controllable TTS? How to Test Itexplainx.ai · Independent hands-on evaluation guide: how to test the preview (mid-sentence shifts, multi-character scenes, single-word fixes), what is confirmed vs not, updated Oct 13 for the open-weights release.
- Fish Audio Launches Drama 3 to Direct AI Voices and Edit a Single Wordunite.ai · Third-party analysis placing Drama 3 in Fish Audio's research lineage (S2 natural-language control), with practical caveats for production teams: results vary between takes and the model keeps changing.
- fishaudio/fish-speechgithub · Official open Fish Speech codebase (Fish Audio Community License) with S2-Pro and Drama 3 install, inference and API-server recipes.
- sgl-project/sglang-omnigithub · SGLang-Omni serving stack with launch-day Drama 3 support and cookbook: `sgl-omni serve --model-path fishaudio/drama-3` exposes an OpenAI-style /v1/audio/speech endpoint with voice cloning and PCM streaming.
Learn more
First spotted on hf: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72Mistral Large 4 (Le Chonk): 1.05T-param open-weight MoEMistral's new flagship: a 1.05T-parameter mixture-of-experts multimodal model with 49B active params and 1M-token…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71JetBrains Mellum 2.1: 12B MoE coding model trained for agentsJetBrains released Mellum 2.1, a 12B MoE model (2.5B active parameters, 128K context) under Apache 2.0, trained mainly…models · JEV 0.69DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.