A detailed look at Fish Audio S2 Pro for expressive text-to-speech, instant voice cloning, multilingual generation, direction tags, streaming and developer use.
Voice AI is moving quickly, and one of the more interesting recent releases is Fish Audio S2 Pro. It is a multilingual text-to-speech model that clones voices from very short samples and lets you steer emotion and delivery with simple text tags. It is also available as open model weights, which is unusual in this category.
This guide explains what S2 Pro is, how voice cloning and voice conversion work on Fish Audio, how to get started, and what to check on consent, licensing, and pricing. To compare it with other tools, browse our Voice Cloning & Voice Conversion category.
Fish Audio S2 Pro is the flagship text-to-speech model from Fish Audio. According to the project documentation, it was trained on more than 10 million hours of audio and supports over 80 languages. It uses a Dual-Autoregressive architecture, with a decoder-only transformer paired with an RVQ audio codec, to balance audio quality and speed. Served through vLLM, it outputs 44.1 kHz mono audio.
Fish Audio is also a full platform around the model. Besides text to speech and cloning, it includes a voice changer for recorded audio, a large voice library, audio separation, audio translation, and a Story Studio for longer projects. You can use it through the web app, the API, or by self-hosting the open models.
One naming note: Fish Audio's own voice cloning page mentions a newer "S2.1 Pro" release. Check the official site for the model version currently powering the web app.
Voice cloning creates a reusable digital voice from a short sample of someone speaking. You then type text and the model reads it in that voice.
Voice conversion takes existing recorded speech and re-speaks it in another voice, keeping the original timing and delivery. On Fish Audio, this is the voice changer feature for recorded audio.
Many projects use both: clone a voice once, then either generate new narration from text or convert a live or recorded performance into that voice.
| Capability | What it does |
|---|---|
| Zero-shot voice cloning | Clones a voice from a short reference clip, typically 10–30 seconds |
| Emotion and prosody tags |
| Control delivery at sub-word level with `[tag]` instructions written in natural language |
| Multilingual output | Supports 80+ languages, with cross-lingual cloning |
| Multi-speaker generation | Native support for dialogue with several speakers |
| Voice changer | Converts recorded speech into another voice |
| Voice library | A large library of public voices you can search and use |
| API and integrations | Developer API, plus integrations such as Home Assistant |
| Open-source models | Model weights available for self-hosting under the Fish Audio Research License |
Explore the platform on the official Fish Audio website and the Fish Speech repository on GitHub.
S2 Pro suits creators and developers who want fast cloning from tiny samples, strong multilingual output, and the option to self-host. It is less ideal if you need a heavily polished enterprise suite with every compliance feature built in. Self-hosting also needs serious GPU capacity: the reference setup loads the model at roughly 48 GB of GPU memory.
| Need | Better fit |
|---|---|
| Fast cloning from short samples, plus open models | Fish Audio S2 Pro |
| Established platform with a large ecosystem | ElevenLabs |
| Enterprise voice cloning and custom deployment | Resemble AI |
| Simple built-in voices without cloning | Fish Audio S1 or a basic TTS service |
1. Make sure you have the rights.
Clone only your own voice, or one you have clear written permission to use.
2. Record a clean sample.
Fish Audio says a clip of roughly 10–30 seconds can be enough. Use a quiet room, a steady microphone distance, and natural speech with no music or echo.
3. Create the voice.
In the web app, open the voice cloning section, upload or record your sample, and save the voice. Free accounts include a few public voice slots.
4. Choose the model and test.
Select S2 Pro for text to speech, paste a script, and generate. Try short and long lines, plus names and technical terms.
5. Add emotion and style tags.
Insert `[tag]` instructions inline, such as a calmer or more excited delivery. The tags are natural-language instructions, not a fixed list, so experiment.
6. Use voice conversion if needed.
Upload recorded speech to the voice changer and pick your cloned voice as the target.
7. Connect through the API.
Developers can call the API, or use hosted providers that offer S2 Pro. If you self-host, voice cloning needs both a reference audio file and its transcript.
8. Review before publishing.
Listen for pronunciation, pacing, and artifacts, and label AI-generated audio where appropriate.
The S2 Pro weights can be downloaded and served on your own hardware through tools such as vLLM and SGLang, with an OpenAI-compatible speech endpoint. The model is released under the Fish Audio Research License Agreement, which is not the same as a permissive license like MIT or Apache. Read it carefully before any commercial deployment, and contact Fish Audio if your use case is unclear.
Cloning a voice is powerful and easy to misuse. Fish Audio states that it does not pre-screen individual use cases and may remove content or accounts that violate its terms or the law. Keep these rules in mind:
This guide is not legal advice. Consult a qualified professional for commercial or high-risk projects.
Fish Audio combines subscription plans with usage-based API billing:
Third-party sources disagree on exact monthly prices, which suggests promotions and changes over time. Always confirm current plans, credit rules, and commercial rights on the official Fish Audio site before you buy.
Fish Audio S2 Pro stands out for how little audio it needs to clone a voice, how much control it gives over emotion and delivery, and how wide its language coverage is. The platform adds a voice changer, a large voice library, and an API, and the open model weights give technical teams a self-hosting path that few competitors offer.
The trade-offs are practical. Self-hosting needs strong GPU hardware, the research license limits how freely you can deploy the weights commercially, pricing details shift often, and responsibility for consent and disclosure sits with you. Start with a short sample of your own voice, test a small project, and scale once the results meet your quality bar. Ready to try it? Explore Fish Audio.
It is Fish Audio's flagship text-to-speech model. It supports zero-shot voice cloning, emotion control through text tags, and more than 80 languages. Learn more at fish.audio.
Fish Audio documents typical reference samples of about 10 to 30 seconds. Clean, clear recordings give better results than longer noisy ones.
Voice conversion changes recorded speech into another voice while keeping the original timing and delivery. On Fish Audio this is done with the voice changer feature.
The model weights are available to download and self-host under the Fish Audio Research License Agreement. See the Fish Speech repository and review the license terms before using it commercially.
The model documentation lists more than 80 languages, and it supports cross-lingual cloning, where a voice cloned in one language speaks another.
Yes. S2 Pro accepts inline [tag] instructions written in natural language to adjust emotion and prosody at sub-word level.
Fish Audio has a free tier with a limited monthly credit allowance and a few public voice slots. Paid plans add capacity, private voices, and commercial use. Check the official site for current details.
Only with their explicit permission. Using a cloned voice to impersonate or mislead people is unethical, may be illegal, and can get your content or account removed.
