Fish Audio, a voice AI company based in Palo Alto, California, has launched Drama 3, a text-to-speech model that users “direct” in plain language rather than simply generating a voice. The company announced the model on 13 October 2026, and its launch blog says it is available on fish.audio, in the Fish Audio iOS and Android app and through its API. Unusually for a commercial voice model, Fish Audio has also released the model’s weights so people can run it themselves.
Text-to-speech software turns written text into spoken audio. “Open weights” means the trained model itself is published for download, though the licence still sets limits on how it can be used.
Update, 29 October 2026: this article has been revised. An earlier version said the release gave no independent benchmarks and listed audio inpainting as an available feature. Fish Audio’s launch blog cites a third-party pronunciation benchmark, and says inpainting is in invite-only testing.
Directions in square brackets, written like notes to an actor
Users write directions in square brackets before the words they apply to, such as “[barely holding it together, forcing a smile]”, and can add sounds like a sigh or a laugh exactly where they should happen. New direction partway through a sentence changes the delivery from that point, so a line can start calm and turn to a terrified whisper. A paired tag limits a delivery to a few chosen words.
The developer documentation says Drama 3 can generate a whole scene with several speakers in one request, switch language for a phrase mid-line (marked as experimental), and fix the pronunciation of a name sound by sound using phoneme tags. The blog says it speaks 83 languages and that directions need not be in the same language as the script.
Fish Audio says it is aimed at audiobooks, audio drama, game and animation dialogue, dubbing and short video. “Voice AI has gotten very good at generating speech. The next step is directing a performance,” said Rissa Cao, co-founder and chief executive.
A 35-billion-parameter model, with some editing tools still in testing
The blog describes Drama 3 as a 35-billion-parameter “mixture-of-experts” model that uses about 3 billion parameters at each step, built from Alibaba’s Qwen3.5-35B-A3B language model. It was tuned with a model of which recordings people preferred, rather than rewards for being easy to transcribe, which Fish Audio says had marked down laughs and whispers.
The press release presented “audio inpainting”, regenerating just one word or segment while keeping the rest of a take, as part of Drama 3. The blog says inpainting and an interactive editing mode are “in small, invite-only testing”, with an editing API “coming soon”.
Price, licence and the company’s figures
The API costs $50 per million bytes of input text, cut to $10 until 12 November 2026, according to the documentation. The press release says businesses switching from another voice provider get their first three months free. The open weights come under a new Fish Audio Community License: free for non-commercial use and for commercial users with under US$500,000 in annual revenue, with larger firms needing a paid licence.
Fish Audio says Drama 3 is first on Artificial Analysis’ Pronunciation Robustness benchmark, reading 93.2% of hard words, numbers and codes correctly against 88.1% for its earlier S2.1 Pro. The release also says the company has grown “from zero to $21 million in annual recurring revenue” with “more than 15 million users”. TSN has not independently checked these figures.
A consent clause matters more than the emotion
Drama 3’s selling point is control: the shift in voice AI is from producing speech to editing a performance, which is what audiobook producers and game studios need when one wrong word should not mean re-recording a chapter. A model that follows direction closely and reads difficult names correctly could cut the cost of narration and dubbing, and open weights let smaller studios and researchers use it without sending scripts to a cloud service.
The same features cut both ways. A model that can make a cloned voice sound convincingly tearful or angry, and that anyone can download, is also more useful for scams and impersonation, the problem behind laws such as Queensland’s proposed offence of creating sexual deepfakes. Fish Audio’s licence bans cloning a real person’s voice without consent, and its launch video used a comedian’s voice “with his permission”, but a licence term is hard to enforce once weights are public. Neither the release nor the launch blog describes watermarking or other ways to detect Drama 3 audio.
Voice cloning done with consent can help people too: ElevenLabs has offered free voice restoration for people with ALS in Spain, and a peer gave the first AI-voiced speech in Parliament. How well Drama 3 performs against rivals such as ElevenLabs beyond one pronunciation test is not yet known.
Sources
- Fish Audio via PR Newswire, “Fish Audio Launches Drama 3: Direct AI Voices in Plain Language, Control Every Word and Shift Emotion Mid-Sentence”, 13 October 2026.
- Fish Audio Blog, “Fish Audio Drama 3: The Most Emotive Text-to-Speech Model We’ve Built”, 13 October 2026.
- Fish Audio documentation, “Drama 3”.
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- LAION Releases Humaneness Voice Converter, an Open Voice-Conversion Model Built on Chatterbox; the Scores Are LAION’s Own
- Amazon Nova 2.5 Sonic: AWS Says Its Voice AI Now Thinks and Answers Faster
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X

