Multilingual speech and voice cloning
fish-speech is a text-to-speech and voice-cloning model that has become one of the better-known open options, now developed under the Fish Audio S2 and S2 Pro branding. It generates natural, expressive speech and can clone a voice from a short sample, across many languages. The README describes the S2 Pro model as trained on more than 10 million hours of audio covering over 80 languages, with a dual-autoregressive architecture. At 30,832 stars as of 2026-06, it is a frequent pick when people want open voice cloning with strong multilingual quality.
There is a history worth knowing: the project carried OpenAudio and S1 branding in earlier phases, so older references and issues mix those names. The current line is Fish Audio S2.
What it does
The core capabilities are multilingual TTS and zero-shot or few-shot voice cloning: give it a reference clip and a transcript, and it produces speech in that voice. It supports expressive control and the major Tier-1 languages (Chinese, English, Japanese) most reliably, with quality varying for less common languages. It ships a web UI, a CLI, and a server mode, and the models download on first use.
Install
pip install fish-speechA Docker image and a Gradio web UI are also available, and the official install and usage docs at speech.fish.audio are the source to follow for the current model and flags, since the project moves quickly. Expect a capable GPU; the larger models are memory-hungry, and several issues report tight VRAM on consumer cards.
The license is the catch
This is the part to read before you build anything on it: fish-speech is not under a standard open-source license. Both the code and the model weights are released under the Fish Audio Research License, which permits research and non-commercial use free of charge but requires a separate license from Fish Audio for any commercial use. That is a deliberate choice by the company behind it, and it is enforced. So fish-speech is fine for experiments, study, and personal projects, but if you plan to ship it in a product or a paid service, you need to contact Fish Audio for terms. Do not assume “on GitHub” means “free to commercialize” here.
Where it fits, and where it doesn’t
Reach for fish-speech when you want high-quality multilingual voice cloning for research or a non-commercial project, when Tier-1 language quality matters, and when you can run a capable GPU. The expressiveness and language coverage are its strong suit.
It is the wrong choice if you need a permissive license for a commercial product, where the research license is a hard blocker unless you pay for commercial terms. It is also heavier on VRAM than some alternatives, and the official training and fine-tuning tooling is not fully open, so custom training is harder than the star count might suggest. For commercial use without a separate agreement, the alternatives below differ sharply on licensing.
fish-speech versus the other open TTS models
| fish-speech | F5-TTS | ChatTTS | VoxCPM | |
|---|---|---|---|---|
| Stars | 30,832 | 14,759 | 39,458 | 28,884 |
| Strength | multilingual cloning, expressive | flow-matching cloning | conversational prosody | voice design and cloning |
| Code license | Fish Audio Research | MIT | AGPLv3 | Apache-2.0 |
| Weights | research, non-commercial | CC-BY-NC | CC-BY-NC | Apache-2.0 |
| Commercial use | separate license required | retrain weights or license | not permitted | permitted |
Counts are from GitHub as of June 2026. F5-TTS has MIT code but CC-BY-NC weights, so commercial use means retraining on your own data. ChatTTS is research-only on both code and weights. VoxCPM under Apache-2.0 is the most commercial-friendly of the group. If licensing is your deciding factor, that column is the comparison that matters.
The star curve
The star curve rose strongly during the open-TTS surge as voice cloning became accessible, and the rebrand from OpenAudio and S1 to S2 marks the shift from a community model to a company-backed product line. The growth reflects genuine capability, tempered by the licensing that limits where it can be deployed.
What the issue tracker warns you about
The recurring issues are practical. VRAM and CUDA compatibility come up often, with reports of tight memory on consumer GPUs and version-specific failures, so check your hardware against the docs. Inference speed is a frequent complaint even on strong cards. The training and fine-tuning workflow is underdocumented and the official tooling is not fully open, which frustrates custom-voice projects. And language quality is tiered: Chinese, English, and Japanese are solid, while less common languages vary, so test your target language before committing.
Related
For MIT-code voice cloning, see F5-TTS. For conversational TTS, see ChatTTS. For an Apache-2.0, commercial-friendly model, see VoxCPM. For what is trending, see LLM tooling, the daily digest, and the weekly report.
FAQ
Can I use fish-speech commercially? Not without a separate license. Both the code and the weights are under the Fish Audio Research License, which is free for research and non-commercial use but requires a commercial agreement from Fish Audio for any product or paid service. Contact them for terms.
Is fish-speech the same as OpenAudio or S1? It is the same project. OpenAudio and S1 were earlier branding; the current line is Fish Audio S2 and S2 Pro. Older references and issues mix the names.
fish-speech vs F5-TTS, which is better for commercial use? F5-TTS has MIT-licensed code, but its pretrained weights are CC-BY-NC, so commercial use means retraining on your own data. fish-speech requires a commercial license for any use. For a permissive default, VoxCPM under Apache-2.0 is the cleaner option.
Which languages does fish-speech handle best? Chinese, English, and Japanese are the most reliable. The model covers many more languages, but quality varies for less common ones, so test your specific target before relying on it.
What GPU do I need for fish-speech? A capable GPU with substantial VRAM. The issue tracker has frequent reports of tight memory on consumer cards and version-specific CUDA failures, so check the official docs against your hardware.