Fish Audio drops S2.1 Pro, takes aim at ElevenLabs with $52M seed

Training frontier voice models used to require massive compute clusters and millions in runway. Fish Audio just proved you can start with a single graphics card and end up taking aim at the biggest players in the business.

The Palo Alto-based startup just closed a massive $52M seed round co-led by Coreline Ventures and Capital Today, with backing from 359 Capital, HF0, and several others. Alongside the check, the team officially rolled out its flagship S2.1 Pro voice cloning engine.

From a side project to $21M ARR

Fish Audio started when former Nvidia researcher Shijia Liao got fed up with stiff, robotic synthetic voices. He trained an expressive speech model on a single GPU and open-sourced it on GitHub as Fish Speech. That repository quickly took off, racking up over 31,000 stars among indie devs and creators.

Over the past year, the company launched five separate models: four for speech generation and one for transcription. Three of those models remain open source, drawing over 8 million users across hosted and self-managed versions. More impressively, the startup scaled to $21 million in annual recurring revenue in its very first year.

CEO Rissa Cao noted that the project was running efficiently on cash flow from creator plans and didn't initially need venture cash. But enterprise demand took off: companies like HeyGen, LiveKit, Retell, Sanas, and OpenArt are already running Fish Audio in production, prompting the founders to raise outside capital for heavy model training.

Sub-second speed and three-minute takedowns

Unlike its predecessors, S2.1 Pro is proprietary and available exclusively through a paid API. It clones a target voice using just 5 seconds of reference audio, operates at twice the speed of Cartesia, and costs one-sixth as much as ElevenLabs.

It also packs word-level steering for emotion, pacing, and pitch across more than 15,000 natural language controls. The company is so confident in its pricing that it promises any enterprise customer a full year of free service if it cannot slash their existing voice stack bill by at least 50%.

Growth hasn't come without friction. A few months ago, voice actors raised concerns that their voices were uploaded to the platform without permission. While the startup previously relied on standard DMCA requests, it has since deployed an automated portal that pulls flagged voices in under three minutes once a creator uploads proof of ownership.

That said, it remains a reactive fix rather than proactive screening. Looking ahead, Fish Audio plans to ship an audio understanding model later this year, alongside an end-to-end speech-to-speech system.

Why it matters

For voice agent builders and video creators, S2.1 Pro represents an immediate upgrade in unit economics. High per-character API costs have long bottlenecked low-latency call bots and interactive gaming agents. Getting ElevenLabs-grade fidelity with granular word-level control at a fraction of the cost makes this worth testing right away.

Across the AI audio landscape, the battleground is shifting from raw voice naturalness to latency and inference efficiency. By combining developer love from its open-source roots with aggressive enterprise pricing, Fish Audio is forcing established rivals like ElevenLabs and Cartesia into a brutal margin squeeze.

We think the commercial strategy is remarkably sharp, especially the bold 50% cost-cut guarantee. However, voice IP remains a legal powder keg. A three-minute automated takedown is impressive engineering, but reactive policing won't satisfy nervous studio executives or angry artists forever. Building robust, upfront consent verification is the real hurdle between Fish Audio and long-term enterprise dominance.