Fish Audio launches Drama 3 voice AI model

Fish Audio markets Drama 3 as open-weight, and the model card backs that framing with downloadable weights on Hugging Face ^[1]under a new Fish Audio Community License v1.0. But the license is not a blanket giveaway: it is free for research and non-commercial use, and free for commercial use only below $500,000 in annual revenue - anyone above that threshold, including larger studios or platforms planning to resell inference, must negotiate a separate commercial agreement. Independent benchmarking group Artificial Analysis, even while crowning Drama 3 the top open-weights text-to-speech model, described the arrangement as open weights without commercial-use rights rather than unrestricted open source ^[2]. The license also addresses a more immediate risk of the technology itself: Fish Audio's official terms explicitly prohibit cloning a voice without the speaker's consent, an acknowledgment that as voice models become more controllable and realistic, the risk of unauthorized voice replication rises in step ^[3]. That distinction matters for who actually benefits: solo developers, indie game studios, and small audio-drama producers get a free, state-of-the-art voice engine, while the open-weight headline does little to open Fish Audio's technology to the well-funded platforms whose margins a genuinely unrestricted release would pressure.

Drama 3's real departure from prior TTS releases is not the pronunciation score, it is what Fish Audio calls audio inpainting: the ability to re-synthesize a single word or phrase inside an already-generated audio file while the model listens to the audio on both sides of the edit and matches the new word's voice, mood, and rhythm to its neighbors ^[3]. Co-founder and Chief Scientist Shijia Liao, a former Nvidia researcher, frames the underlying architecture change as what makes that control possible: "Drama 3 adopts a new architecture that allows for better control and instruction following. When you change a single word, Drama 3 listens to the audio on both sides of it." ^[4]Technically, the feature sits on top of a 35-billion-parameter mixture-of-experts model with roughly 3 billion active parameters per token, a Qwen3.5-35B-A3B backbone, and a new Codebook-MTP audio decoder that causally decodes 10 codebooks ^[1]. For production workflows in audiobooks, dubbing, and multi-character game dialogue, where a single mispronounced word previously meant regenerating a whole passage and hoping the new take matched the old one, that is a change to the workflow itself, not just an incremental quality bump.

On Artificial Analysis's Pronunciation Robustness leaderboard, Drama 3 Preview scored 93.2 percent, placing it ahead of every model the firm tracks, open or closed, including ElevenLabs' v4 (91.7 percent) and v4 Turbo (90.1 percent), and well above Fish Audio's own prior flagship S2.1 Pro (88.1 percent) ^[3]. That is a meaningful, third-party-verified result. Online developer reaction, though, has been less unanimous: some commentary raised a now-familiar critique, that a leaderboard win does not automatically settle how a voice actually sounds in practice, with doubts also raised about quality in languages beyond the headline set ^[2]. Hands-on creators told a more favorable story: one AI filmmaker, who said voice had long been the weak link in his generated video work, described Drama 3 as the first model where plain-language direction over delivery, such as a mid-line emotion shift, actually felt real. The split is a familiar pattern in AI model launches: an automated leaderboard metric, however rigorously measured, does not by itself settle a debate about how a voice sounds to a human ear across 83 languages and a wide range of real-world use cases.

Fish Audio is roughly one year old, founded by former Nvidia researcher Shijia Liao, and closed a $52 million seed round on July 28, 2026, co-led by Coreline Ventures and Capital Today ^[6], at which point it reported $21 million in annual recurring revenue and more than 8 million users ^[5]. By the time Drama 3 shipped roughly two and a half months later, that user count had grown past 15 million ^[4]. Fish Audio's $21M ARR is a small fraction, roughly 4 percent, of ElevenLabs' reported ~$500 million annual run rate ^[5], yet Drama 3 now tops an independent leaderboard that ElevenLabs' flagship models also compete on. Shipping an open-weight, benchmark-leading model functions as a distribution strategy as much as a research result: it converts developer goodwill and GitHub/Hugging Face visibility into paid API and subscription customers faster than a traditional enterprise sales motion, letting a much smaller company contest a much larger incumbent on cost-efficiency rather than marketing spend alone.