Fish Audio Drama TTS Model

Drama 3 is a 35-billion-parameter mixture-of-experts model with only about 3 billion active parameters per token ^[1][5], supporting 83 languages. Its headline feature, "audio inpainting," lets a user regenerate a single word in a sentence while the model listens to the audio on both sides of the edit and matches the surrounding voice, mood, and rhythm ^[3][5]- rather than forcing a full re-render of the line. It also generates complete multi-speaker scenes in a single pass, keeping each character's voice distinct and responsive to the preceding line of dialogue ^[3]. Fish Audio co-founder and chief scientist Shijia Liao has said the model "adopts a new architecture that allows for better control and instruction following," crediting the redesign for the jump in controllability ^[2].

Fish Audio and Artificial Analysis both describe Drama 3 Preview as the new #1 open-weights model on the Text to Speech leaderboard and the outright leader of the Pronunciation Robustness leaderboard, ahead of ElevenLabs' Eleven v4 and v4 Turbo ^[4]. On that specific metric, Drama 3 scores 93.2% versus 91.7% for Eleven v4 and 90.1% for Eleven v4 Turbo ^[1]. But the claim is narrower than the marketing suggests: on Artificial Analysis' broader Provider Voice Arena, which ranks models by head-to-head Elo across thousands of comparisons, Drama 3 Preview sits 10th overall with an Elo of 1243 - still trailing Eleven v4 Turbo (1330) and Eleven v4 (1328) - even though that marks a 101-point jump over Fish Audio's own prior S2.1 Pro model ^[4]. Fish Audio prices Drama 3's API at $15 per 1 million characters, against $80 for Eleven v4 and $40 for Eleven v4 Turbo ^[4]- a cost argument that may matter to buyers as much as the exact leaderboard position.

Fish Audio published Drama 3's weights on Hugging Face under a license that is free for commercial use for any company and its affiliates with under $500,000 in annual revenue ^[1], explicitly courting self-hosted deployment in VPCs, on-prem environments, sovereign clouds, and air-gapped settings ^[1]- deployment modes a closed, API-only incumbent like ElevenLabs cannot offer. That positioning lines up with how outside analysts frame the competitive map: coverage of the release describes Drama 3 as addressing a gap in professional, long-form voice production - audiobooks, games, multilingual dubbing - where word-level edits via audio inpainting avoid the cost of re-rendering an entire take ^[7]. The flip side of a model this good at preserving voice, mood, and rhythm across single-word edits is that it also raises the fidelity ceiling for synthetic audio manipulation, a risk flagged in broader coverage of Fish Audio's funding round ^[6].

Fish Audio closed a $52 million seed round co-led by Coreline Ventures and Capital Today in late July 2026, when it reported $21 million in annual recurring revenue and more than 8 million users ^[6]. By the Drama 3 launch in October 2026, the company said its user base had grown past 15 million ^[2], with brand search volume up 740% year-over-year ^[2]. Reaction has been positive: Fish Audio's own announcement drew outsized engagement, and an independent reviewer on X said Drama 3 was the first open TTS model that actually sounds right to them. The one hands-on video review found Drama 3 more expressive and emotional than Fish Audio's own earlier S2 and S2.1 Pro models, though it also noted the release was still labeled "preview" at the time of testing. On Reddit, discussion of the newly open weights centered on the practicalities of running the model locally.