Rissa Cao and Shijia Liao have raised a $50 million seed round for Fish Audio, adding capital to a Palo Alto-based AI voice developer that grew from Liao's one-GPU research project into a creator platform and enterprise API provider. TechCrunch reported the financing on July 28.
Coreline Ventures and Capital Today led the round. The investor group also included 359 Capital, Parable, Play Time, Alphalist Partners, Bayhouse Ventures, Carya Venture Partners and HF0. Fish Audio did not disclose a valuation.
Liao, Fish Audio's co-founder and chief scientist, previously worked as a researcher at Nvidia. He started Fish Speech after growing frustrated with synthetic voices that could sound expressive for a short clip but failed to sustain that quality across longer passages. Liao trained the first model on a single GPU and released the code publicly. The resulting Fish Speech repository has collected more than 31,000 GitHub stars, according to TechCrunch.
Cao, Fish Audio's co-founder and CEO, is leading the commercialization push as Fish Audio tries to turn public model access and creator adoption into enterprise API demand.
A large seed behind self-reported traction
Fish Audio says its hosted and downloadable models have reached more than 8 million users and $21 million in annual recurring revenue since launching in 2025. The user figure combines people using open-weight models with users of Fish Audio's hosted products, so it should not be read as a paid customer count.
Fish Audio has shipped five models over the past year, including four speech-generation systems and one speech-to-text model. Three speech-generation releases have been made available with downloadable weights. Fish Audio sells creator subscriptions, usage-based APIs and enterprise access, and says HeyGen, Sanas and Plaud use its technology.
The financing gives Cao and Liao room to pursue two expensive jobs at once: training larger audio models and building the sales, support and reliability functions required by enterprise customers.
Open weights feed a paid API business
Fish Audio uses downloadable model weights to attract developers and sells usage-based API access. Fish Audio's pricing documentation lists s2-pro/s1 API access at $15 per million UTF-8 bytes.
Fish Audio has also worked to reduce the computing cost behind that API. In an engineering post about its inference stack, Fish Audio says it rebuilt its inference pipeline to reduce the computing cost behind S2 Pro API access.
The underlying S2 release includes model weights, fine-tuning code and an SGLang-based inference engine. Fish Audio says S2 Pro was trained on more than 10 million hours of audio and supports more than 80 languages; its S2 technical report describes pre-training on more than 10 million hours across approximately 80 languages and dialects. The report also says S2 can begin streaming audio in about 100 milliseconds on an H200. Fish Audio designed the model to follow natural-language instructions for emotion, pacing and speaker behavior.
Fish Audio enters the enterprise market against heavily financed competitors. ElevenLabs said in February that it raised $500 million at an $11 billion valuation, with products spanning voice agents, creator tools and developer APIs. Fish Audio's wager is that public model access, lower serving costs and fine-grained voice controls can produce a less capital-intensive route into many of the same accounts.
Voice ownership becomes infrastructure
Fish Audio's creator distribution model carries a rights and consent risk. The Fish Audio website advertises a library of more than 2 million voices, many uploaded by users. In June, the UK performers' union Equity said it had demanded that Fish Audio remove unauthorized cloned voices from the platform.
Those allegations show why reporting, ownership verification and takedowns are core operational functions for Fish Audio as its user-uploaded voice library grows.
The $50 million gives Cao and Liao the resources to push an open-weight research project deeper into commercial voice infrastructure. Their next test is whether developer adoption can keep filling the sales pipeline while enterprise contracts pay for model training, computing capacity and rights controls.