Fish Audio launched Drama 3 on October 13, 2026, a 35-billion-parameter open-weight text-to-speech model that lets creators direct a voice performance in plain language — shifting emotion mid-sentence, generating multi-character scenes, and editing single words without re-rendering. The model leads Artificial Analysis' Pronunciation Robustness benchmark and is the highest-ranked open-weights entry on the Provider Voice Arena.
On October 13, 2026, Fish Audio launched Drama 3, the text-to-speech model it describes as its 'most emotive and controllable' yet. Instead of simply generating a reading, the model takes direction: users describe the desired tone, pacing, and character in plain language, then shape the performance line by line until it is right. The launch, announced from Palo Alto, came with an 80% discount on API usage for the first 30 days and three free months for enterprises switching from another voice provider.
Drama 3 adds four ways to shape a performance: natural-language direction that can change a sentence mid-line — starting calm and turning panicked without a new take; multi-speaker dialogue that generates a whole scene in one pass with each character keeping its own voice; audio inpainting that regenerates only the edited segment while keeping everything around it intact; and long-form consistency that holds one voice across long scripts, with mid-scene language switching and phoneme-level pronunciation control in English, Chinese, Japanese, Korean, and more. Under the hood it is a 35-billion-parameter mixture-of-experts model with roughly 3 billion parameters active per token, initialized from Qwen3.5-35B-A3B, with weights released on Hugging Face under the Fish Audio Community License. 'When you change a single word, Drama 3 listens to the audio on both sides of it and makes the new word fit,' said co-founder and chief scientist Shijia Liao.
Independent benchmarks put the Drama 3 preview at 93.2% on Artificial Analysis' Pronunciation Robustness benchmark — first overall, ahead of ElevenLabs Eleven v4 at 91.7% and v4 Turbo at 90.1%, and up from 88.1% for Fish Audio's previous S2.1 Pro. It also reached a Provider Voice Elo of 1,243, a 101-point gain over S2.1 Pro that made it the highest-ranked open-weights model in a field otherwise led by closed systems. Early pricing coverage listed the API around $10 to $15 per million characters, against $80 per million for Eleven v4. The model supports 83 languages and up to 16 speakers in a single pass.
Drama 3 caps a year of momentum for Fish Audio, which grew out of the Fish Speech open-source project and now serves more than 15 million users and $21 million in annual recurring revenue, ranking No. 1 on Ahrefs' list of the 50 fastest-growing AI companies. Rissa Cao, co-founder and CEO, framed the shift: 'Voice AI has gotten very good at generating speech. The next step is directing a performance.' Caveats temper the launch: the community license requires a separate commercial agreement above $500,000 in annual revenue, and interactive editing and inpainting remain invite-only for now.
