Fish Audio open-sources Drama 3, tops open-weight TTS rankings

Fish Audio has released Drama 3, an open-source text-to-speech model with weights available under the Fish Audio Community License, now on Hugging Face, with API pricing at $15 per million characters. Artificial Analysis published several evaluations the same day, with Drama 3 emerging as the leading open-source speech model.

Confirmed

• Provider Voice Arena: Elo of 1,278 (1,512 appearances), 6th overall and first among open-weight models; strong performance in scenarios like customer service.
• Pronunciation robustness benchmark: Overall score of 94.6%, ranking first among all tested models (including closed-source ones). Human reviewers judge whether models correctly read four categories of difficult text; scores on subcategories such as abbreviation expansion exceed 96%.
• Controlled Voice Arena (all models use the same cloned voice): Overall score of 1,085, 12th overall and first among open-weight models.

Why it matters

• Open-weight models have, for the first time, matched or even surpassed closed-source TTS in overall voice arena rankings and pronunciation robustness, narrowing the open–closed gap in speech synthesis.
• Releasing weights under a Community License alongside a low-priced API offers developers and use cases requiring local deployment a new option.

2026-10-29 ~ 2026-10-29 · 5 related posts

• [source] Fish Audio's open-weights Drama 3 tops TTS arenas with best-in-class pronunciation robustness — ArtificialAnlys · 2026-10-29
• Fish Audio Drama 3 ranks #6 in Speech Arena, #1 among open weights — ArtificialAnlys · 2026-10-29
• Drama 3 Tops Open-Weights Voice Arena at 1,085, Ranks #12 Overall — ArtificialAnlys · 2026-10-29
• [source] Drama 3 tops pronunciation robustness at 94.6%, beating all proprietary TTS — ArtificialAnlys · 2026-10-29
• [source] Fish Audio open-sources Drama 3 TTS at $15 per 1M characters, near top speed — ArtificialAnlys · 2026-10-29