Fish Speech
A multilingual text-to-speech system with voice cloning from a 10 to 30 second clip and inline emotion tags, released with its weights under a research licence that requires a separate agreement for commercial use.
What it does
Fish Speech, whose current model the README calls S2 Pro, is described as a "state-of-the-art multilingual text-to-speech system". It clones a voice from a short reference sample, "typically 10-30 seconds", with no fine-tuning. It accepts inline emotion and delivery tags such as [whisper], [excited] and [angry], and the README says over 15,000 tags are supported. It can generate multi-speaker, multi-turn audio in one pass.
The README says the model was "trained on over 10 million hours of audio data covering more than 80 languages", with Japanese, English and Chinese in the top tier and Korean, Spanish, Portuguese, Arabic, Russian, French and German in the second. Streaming is supported through SGLang, with the project's own figures of about 100 milliseconds to first audio on an NVIDIA H200.
The README states 4 billion parameters; the Hugging Face model card states 5 billion. Both are the project's numbers.
What it replaces
Fish Speech replaces ElevenLabs for research, evaluation and non-commercial use where voice cloning and expressive control matter. For a commercial product it does not, unless you sign a separate agreement with Fish Audio.
That is the whole reason it is labelled source available. The LICENSE file says the agreement "is intended to allow research and non-commercial uses of the Materials free of charge" and that "no commercial rights are granted". Commercial purpose, as the licence defines it, includes a hosted service, an API, and "your business's or organization's internal operations". Distributors must display "Built with Fish Audio". The licence is revocable and forbids using the materials to train other generative models.
How it runs
Installation is documented at speech.fish.audio. The README describes a command line, a web interface, server inference, a Docker image (fishaudio/fish-speech on Docker Hub), and serving recipes for SGLang-Omni and vLLM-Omni. Performance figures are quoted on an H200 GPU; minimum requirements are not stated. Fish Audio runs a hosted playground and sells the commercial licence.
Who it fits
A team evaluating what expressive, cloned speech sounds like today, or a non-commercial project. A team shipping a product should read the licence first and budget for the agreement, or choose Chatterbox or Kokoro, whose weights are MIT and Apache-2.0.
Limits the README states
"We do not hold any responsibility for any illegal usage of the codebase." The latest tagged GitHub release is v1.5.1 from May 2025; the S2 Pro model described in the README has no tag.
Sources
- Fish Speech README: github.com/fishaudio/fish-speech, read 2026-09-16.
- Fish Audio Research License Agreement, last updated 2026-03-07: LICENSE, read 2026-09-16.
- S2 Pro model card: huggingface.co/fishaudio/s2-pro, read 2026-09-16.
More text to speech and voice cloning
A family of text-to-speech models from Resemble AI under MIT, with zero-shot voice cloning from a reference clip and a multilingual model covering 23 languages.
A research text-to-speech model with zero-shot voice cloning whose code is MIT but whose pretrained weights are non-commercial, because of the dataset they were trained on.
A fast, local neural text-to-speech engine maintained by the Open Home Foundation, used by Home Assistant and the NVDA screen reader, under GPL-3.0.
A Docker image that serves the 82-million-parameter Kokoro model through an OpenAI-compatible speech endpoint, with streaming, voice mixing and nine languages, all under Apache-2.0.
Added September 16, 2026. Found an error on this page? Write to reveneau@licheo.com and it gets fixed in the next weekly pass.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.