Fish Speech
fishaudio/fish-speechReviewed by AI
Checked by RepoPicks Radar on
What is Fish Speech?
Fish Speech reads text aloud with the voice you give it: 10 to 30 seconds of sample audio is enough to copy a timbre, inline tags like [whisper] or [angry] steer the emotion, one script can hold several characters, and the newest model can replace a single word in an existing take. You run it yourself through a command line, a web page or an API server, with Python or Docker and an NVIDIA graphics card. The licence is not a normal open-source one: free for personal and research use, limited for business.
How easy is it to install Fish Speech?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- Linux D
Linux is the best-supported path and is still a D: there is no installer, you set up Python or Docker, download the model weights yourself and start a server or web page from the terminal, with an NVIDIA graphics card (AMD only through the separate ROCm compose file). Windows follows the same documented steps with the same requirements, so it is no easier.
Coming soon: install in 1 click with the RepoPicks app.
Fish Speech on GitHub
- Stars
- 33,665on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +84stars in 6 days
Stars over time
Facts
- Runs on
- Windows, Linux
- You need to install
- Python, Docker, NVIDIA graphics card
- To build from source
- Python, Docker, NVIDIA graphics card
- Disk space
- 950 MB–4.3 GB · estimate
- Accounts or keys
- None
- Cost
- Freemium · Free for research, hobby and evaluation use. Business use is only allowed while your yearly revenue is under USD 500,000; above that, and for any service that lets other people generate voices, you must buy a written licence from Fish Audio.
- License
- NOASSERTION (Other)
- Can replace
- ElevenLabs, Descript, Fish Audio
- Last change on GitHub
- Oct 29, 2026
Fish Speech risk signals (5)
- Attentionlicense is not open source
GitHub shows no recognised licence and the LICENSE file is the "FISH AUDIO COMMUNITY LICENSE AGREEMENT": commercial use is allowed "only while Your Annual Revenue is less than USD 500,000", "no license is granted under this Agreement to offer a Hosted Service", and the licence is "revocable". The README adds: "We will take action against any violation of the license."
AI review
- Attentionmodel weights under stricter license
README license notice: "The S2-Pro weights remain under the Fish Audio Research License", a different and narrower licence than the one covering the code, so the older flagship model may not be used for business at all.
AI review
- Attentionvoice cloning can be misused
README: "Rapid voice cloning: 10-30 seconds of reference audio is enough to reproduce a timbre" and "Multilingual voice cloning". Copying somebody's voice without permission can break impersonation, publicity or copyright rules; the README's own warning says "We do not hold any responsibility for any illegal usage of the codebase. Please refer to your local laws about DMCA and other related laws."
AI review
- Infoheavy hardware for the new model
The models table lists Drama 3 at "35B-A3B (3B active per token)" and compose.yml comments "e.g. cuda:0 to keep Drama 3 on one GPU; unset uses up to four", so the flagship model is sized for multiple graphics cards. S2-Pro (4B) is the one described as small enough for a consumer GPU.
AI review
- Infonewest version is a beta
The latest release is v3.0.0-beta (2026-10-13) and the one before it v2.0.0-beta; the last release without the beta mark is v1.5.1 from 2025-05-31.
AI review
Evidence is shown in its original language.
What Radar could not check about Fish Speech
- How much disk and video memory the weights really take: a 35-billion-parameter model is far larger than the checks' 0.95-4.45 GB estimate, which only covers the code and its dependencies.
- Whether it runs at a usable speed on a normal home graphics card, since the published speed figures come from an H200 server card.
- Whether the Windows instructions work as smoothly as the Linux and Docker ones.
- How close the cloned voices and the benchmark scores are to what you get with your own recordings.
- Whether your intended use falls inside the licence's free limits; that is a legal question about your own revenue and service.