A voice-AI company you may not have heard of just put up numbers most funded startups don't reach in three years: $52M seed, $21M ARR, 8 million users — about a year in. The interesting part for a builder isn't the round. It's how Fish Audio got there: an open-weights model that did the distribution, and a paid API that did the revenue. That's a playbook you can study whether or not you ever ship a single second of audio.

What it is#

Fish Audio makes text-to-speech. The open half of the company is Fish Speech, an open-source, multilingual TTS system with 31.4k GitHub stars. Its flagship open model, Fish Audio S2 Pro (4B parameters, trained on over 10 million hours of audio across 80+ languages), is published on HuggingFace and runs on your own hardware. Three of the company's speech models are open-sourced.

The closed half is S2.1 Pro — the newest and most capable model, available only through the paid Fish Audio API. Through the end of August 2026, S2.1 Pro is free to developers via that API. If you want to hear the best model before you pay, the window is open now.

Who's behind it#

Founder Shijia Liao, a former Nvidia researcher, built the first Fish voice model on a single GPU before open-sourcing it. The July 28 seed round was co-led by Coreline Ventures and Capital Today. The framing in the announcement — "passion project into one of voice AI's fastest-growing companies" — is the tell: the open repo was the growth engine, not a marketing afterthought.

How to start#

Self-host the open weights when you need offline, data residency, or zero per-call cost:

# Open-weights path — you bring the GPU
git clone https://github.com/fishaudio/fish-speech
# follow speech.fish.audio/install for CLI, WebUI, server, or Docker

S2 Pro's weights are on HuggingFace. One caveat: Fish Speech ships under the Fish Audio Research License, not a permissive OSS license — read the LICENSE before you put it in a commercial product.

Call the hosted API when you'd rather not run inference:

# pip install fish-audio-sdk   (Python, TypeScript, and Go SDKs exist)
from fish_audio_sdk import Session, TTSRequest

session = Session("YOUR_API_KEY")            # POSTs to api.fish.audio/v1/tts
with open("out.mp3", "wb") as f:
    for chunk in session.tts(TTSRequest(text="Welcome to the demo.")):
        f.write(chunk)

The hosted path gets you S2.1 Pro with no infrastructure — and it's free for developers through end of August. Confirm the current request schema at docs.fish.audio before you wire it into production. If you're weighing this against the incumbents for a voice agent, we compared the transcription-side options in Deepgram vs AssemblyAI vs Whisper and the build-vs-buy call in OpenAI's Presence vs Realtime API.

Pricing#

The founder takeaway#

Strip away the voice and Fish Audio is a clean open-core machine:

  1. Free, self-hostable weights win distribution. 31k stars and 8M users didn't come from a sales team — they came from a repo people could run.
  2. The newest model is the paid tier. S2 Pro is open; S2.1 Pro is API-only. The community gets a great model; the best model is the upsell.
  3. The managed API monetizes convenience. Most developers don't want to keep a GPU warm. Selling them the endpoint captures the ones the open weights attracted.

If you're sitting on a model — voice, vision, retrieval, anything — that three-step wedge is the repeatable part. Open the thing that drives adoption, gate the newest thing behind an API, and charge for not having to run it yourself. Fish Audio just priced that strategy at $52M.