Fish Audio

Fish Audio grew out of a personal project by former NVIDIANVIDIANVIDIA is an American semiconductor and computing technology company that provides graphics processing units, artificial intelligence infrastructure, networking interconnects, and computing software, and is listed on Nasdaq under the ticker NVDA.Open the full entry researcher Shijia Liao. While working as a research engineer on video at NVIDIA, he watched VTuber streams in his spare time and trained voice models on a single 4090 GPU in his bedroom, then open-sourced the work as Fish Speech.⁠[1][2]

Chief executive officer Rissa Cao wrote: “Fish Speech went from a repo to a company because of game developers, anime fans, and people dubbing characters for fun.”⁠[1]

On November 2, 2024, Shijia Liao and six co-authors submitted the Fish-Speech technical report to arXiv, describing a multilingual text-to-speech framework that uses a serial fast-slow Dual Autoregressive (Dual-AR) architecture to stabilize codebook-based sequence generation and replaces traditional grapheme-to-phoneme conversion with large language models.⁠[3]

On March 9, 2026, Fish Audio open-sourced S2, a text-to-speech model that supports fine-grained inline control of prosody and emotion. S2 uses a Dual-AR design with about 4 billion parameters along the time axis and about 400 million along the depth axis, and was trained on more than 10 million hours of audio; its weights and fine-tuning code were released on GitHub and Hugging Face together with an SGLang-based inference engine. The company published Seed-TTS Eval word error rates of 0.54% for Chinese and 0.99% for English, and a time to first audio of about 100 milliseconds on a single H200 GPU.⁠[4]

On July 27, 2026, marking its first anniversary, Fish Audio announced a $52 million seed round led by Coreline Ventures and Capital Today, with participation from 359 Capital, Play Time, HF0, 645 Ventures, Parable, Carya Venture Partners and angel investors.⁠[1][2]

Co-founder and chief executive officer Rissa Cao previously spent years in voice AI at Amazon Alexa and Meta, and Shijia Liao is chief scientist. The company said that in its first year it shipped five audio models (four speech-generation models and one speech-to-text model), grew from 3 to 22 people, reached $21 million in annual recurring revenue, and passed 8 million users and 2 million voice models in its community library.⁠[1]

Fish Audio is based in Palo Alto; three of its four speech-generation models are open source, while the latest, S2.1 Pro, is available only through a paid API. It sells monthly plans for creators and teams as well as an enterprise API and platform, with HeyGen and Sanas among its customers. Rissa Cao said the open-source project with creator plans had run efficiently without outside money and that the company raised capital to build more advanced models and serve enterprises; it planned to release an audio understanding model in 2026 and was building a speech-to-speech model.⁠[2]

On October 13, 2026, the Fish Audio team submitted the Drama 3 technical report to arXiv. Drama 3 is an open-source speech generation model built on a 35-billion-parameter mixture-of-experts backbone with 3 billion parameters active per token; it turns text into directed performances and supports voice cloning, multi-character scenes, speech editing and streaming audio. The fish-speech repository lists it as the newest flagship, with S2-Pro as the previous-generation flagship.⁠[5][6]

Fish Audio pays users whose voices are used for training. Some creators alleged that their voices had been uploaded to the platform without consent and that DMCA takedowns were slow; the company has since automated takedowns, which it says take less than three minutes after a creator submits a voice sample or contract. Uploads are not blocked automatically, and a voice remains on the platform until the rights holder files for removal.⁠[2]

Fish Audio offers text-to-speech, voice cloning and API services, and the platform is operated by Hanabi AI Inc. As of November 4, 2026, the fishaudio/fish-speech repository on GitHub showed about 33,500 stars; its code and model weights are released under the Fish Audio Community License, while the S2-Pro weights remain under the Fish Audio Research License.⁠[7][6]

Fish Audio: Official website: https://fish.audio/ (opens in a new window).⁠[7]

X: Public account: https://x.com/FishAudio (opens in a new window).⁠[7]

YouTubeYouTubeYouTube is Google’s video-sharing and creator platform, founded in 2005 by Steve Chen, Chad Hurley, and Jawed Karim and acquired by Google in 2006. Built around video uploading, viewing, and creator earnings, the platform has developed Shorts, YouTube Music, Premium, YouTube TV, and products for children, and is headquartered in San Bruno, California, United States.Open the full entry: Public channel: https://www.youtube.com/@fishaudio (opens in a new window).⁠[7]