Fish Audio, a Palo Alto voice AI company trading legally as Hanabi AI, has closed a $52 million seed round, a size Tech Times notes is more typical of a Series B. The round, confirmed on 28 July, was co-led by Coreline Ventures and Capital Today with seven other backers. The company says it reached $21 million in annual recurring revenue and more than eight million users in its first year without outside money.
The funding lands while the company is still dealing with complaints from performers whose voices turned up on its platform without consent. Earlier this year a Bristol voice-over artist, Faye Dicker, found a clone of her voice that had been downloaded more than 900 times before she knew about it, and the performers' union Equity said it had lodged claims for more than 20 members in similar situations. A French voice artist wrote publicly that a removal she obtained in May did not stop her voice being generated, because taking a voice off the site is not the same as erasing it.
Chief executive and co-founder Rissa Cao has said takedowns are now automated: a performer can prove ownership with a short sample or a contract and have a clone pulled from fish.audio in under three minutes. Tech Times argues that this fixes only half the problem. Three of the company's four speech models, released as Fish Speech, are published openly on GitHub, and anyone who downloads the weights can clone voices on their own machine without touching Fish Audio's servers. A takedown on the hosted site has no effect on those copies, and the company has no way to see or revoke that use.
Osuke Honda of Coreline, one of the lead investors, said creator trust would decide whether the company's community approach lasts, and called for consent, transparency and attribution to be designed in from the start, along with verified voice ownership, clear licences and revenue sharing across the industry. According to the piece, he did not speak to the open-weight issue.
The technology explains both the low cost and the consent gap. The flagship S2.1 Pro model splits speech generation into two stages, one planning what is said and in what emotional register and one rendering the fine acoustic detail that makes a voice recognisable, with the first audio arriving in about 90 milliseconds. Cloning is zero-shot: a clip of five to ten seconds is enough, because the model was trained once on more than ten million hours of audio in 83 languages and applies what it learned to any new voice. More than 15,000 inline tags such as whisper or laugh let users direct delivery word by word, which the company pitches to game studios and other clients. Its own blind tests say listeners preferred S2.1 Pro 67% of the time, a figure that has not been independently audited.
Co-founder and chief scientist Shijia Liao, formerly a video researcher at NVIDIA, built the first Fish Speech model on a single gaming graphics card at home and released it as open source, where it gathered more than 31,000 GitHub stars. The commercial service launched in 2025 with a shared library of user-submitted voices whose owners are paid when their voices are licensed; that library now holds more than two million voices. HeyGen, Retell, LiveKit, Sanas and OpenArt build on its API, and customers in regulated fields can run it on their own premises with no data retained and HIPAA-compliant settings.
Cao has said the company did not need outside capital at first and raised to build more advanced models and to serve a rush of enterprise customers. Rico Mallozzi of 359 Capital said the small team had matched far better-funded labs in making synthetic voices sound human, which he gave as the reason for investors' interest. The money is earmarked for voice-native language models, speech-to-speech translation, an audio understanding model and a larger sales team. A free S2.1 Pro API period was set to end in August, and turning that developer base into paying accounts is the near-term test. The rival it chases is far larger: ElevenLabs raised $500 million at an $11 billion valuation in February and had passed $500 million in annual revenue by May, while Cartesia has raised $191 million in total and Speechify and WellSaid compete for the same budgets. Fish Audio's pitch rests on two bets: that the open-weight community it built before hiring any sales staff works as a distribution channel, and that its architecture lets it serve customers more cheaply per generation than closed rivals.
The legal ground is moving. The NO FAKES Act, which would create a federal right over a person's voice and likeness with liability and a notice-and-takedown duty for platforms, cleared the Senate Judiciary Committee unanimously in June but is not yet law. EU AI Act rules on disclosing synthetic voices apply from 2 August, Tennessee's ELVIS Act already bars unauthorised voice cloning and can reach platform providers, and an Illinois biometric privacy class action filed in May targets ElevenLabs and eight other voice companies, with statutory damages of $1,000 to $5,000 per violation for each person that could add up quickly across a class.
None of those measures separates a hosted service that can delete a voice from a published model that cannot be recalled. Tech Times concludes that this gap in the law matches the gap in Fish Audio's own consent system, and that the voice industry has yet to answer it.
Sources · 1