# S2 Pro

Fish Audio's open-source voice generation. S2 Pro covers more than 80 languages and takes prosody direction written in plain language, with published weights and a full serving stack.

- **Model ID:** `fish-audio/s2-pro`
- **Type:** speech
- **Providers:** fish-audio
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** $15/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/s2-pro

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { experimental_generateSpeech as generateSpeech } from 'ai';
import { gateway } from '@ai-sdk/gateway';
import { writeFile } from 'node:fs/promises';

const result = await generateSpeech({
  model: gateway.speechModel('fish-audio/s2-pro'),
  text: 'Hello from the Vercel AI Gateway!',
  // Browse voices at https://fish.audio/app/discovery
  // Open a voice, then use "Copy Model Id" in its "..." menu.
  voice: '933563129e564b19a115bedd57b7406a',
});

await writeFile('speech.mp3', result.audio.uint8Array);
```

## About

S2 Pro is Fish Audio's open-source text-to-speech model, covering more than 80 languages. The release includes the model weights, fine-tuning code, and a complete serving stack, which is what separates it from the hosted-only options in the family.

The architecture is a dual autoregressive design over a residual vector-quantized audio codec. A larger model predicts the primary semantic codebook and a smaller one fills in the remaining residual codebooks at each step, which is how it keeps quality high without running one large model over every codebook.

Prosody control works the same way as in the S2.1 generation. Direction is written inline in square brackets and interpreted as description rather than matched against a fixed vocabulary, so arbitrary phrasing works. Multi-speaker dialogue is supported in a single request.

Integration uses the AI SDK's speech generation function.

## What to consider

The licence is the thing to check first. S2 Pro is published under Fish Audio's research licence: research and non-commercial use is free, and commercial use requires a separate licence agreement. Open weights here do not mean open commercial rights.

S2.1 Pro supersedes this model. It covers three more languages, and listening comparisons favour it, so choose S2 Pro when you specifically want published weights or a self-hosted serving stack, and S2.1 Pro when you want the best available quality through an API.

Self-hosting means operating the serving stack yourself. Through AI Gateway you skip that, but then the open weights are not the reason you are choosing this model, and S2.1 Pro is likely the better pick.

## When to use

### Best For

- **Open-weight voice work** where published weights and fine-tuning code matter
- **Self-hosted serving** using the released inference stack
- **Multilingual coverage** across more than 80 languages
- **Expressive direction** written in plain language rather than a fixed tag list
- **Custom fine-tuning** on top of a published base

### Consider Alternatives When

- **Commercial use** is planned without a separate licence agreement in place
- **Best available quality** is the goal, where S2.1 Pro is preferred in listening comparisons
- **Maximum language coverage** is needed, where S2.1 Pro adds three more
- **You will not self-host**, in which case the open weights bring you no benefit

## Best for

- **Open-Weight Voice Work:** Published weights and fine-tuning code
- **Self-Hosted Serving:** A released inference stack you can operate
- **Multilingual Coverage:** More than 80 languages
- **Expressive Direction:** Plain-language control rather than a fixed tag list
- **Custom Fine-Tuning:** A published base to build on

## Consider alternatives

- **Commercial Use:** Requires a separate licence agreement from Fish Audio
- **Best Available Quality:** S2.1 Pro wins head-to-head listening comparisons
- **Maximum Language Coverage:** S2.1 Pro supports three more languages
- **Hosted-Only Deployments:** Open weights bring no benefit if you never self-host

## Frequently asked questions

### Is S2 Pro open source?

The weights, fine-tuning code, and serving stack are published, but under Fish Audio's research licence. Research and non-commercial use is free; commercial use requires a separate licence.

### How many languages does S2 Pro support?

More than 80. S2.1 Pro extends that to 83.

### How is S2 Pro different from S2.1 Pro?

S2.1 Pro is the improved successor, with better quality, latency, and throughput, and it wins head-to-head listening comparisons. Choose S2 Pro when published weights or self-hosting is the requirement.

### How do I control emotion and delivery?

Write direction inline in square brackets. The model interprets the description rather than matching a fixed tag list, so phrasing outside any documented set still works.

### Can S2 Pro render dialogue?

Yes. Multi-speaker dialogue is supported in a single request, with each turn tagged by speaker index.

### Does S2 Pro support Zero Data Retention?

Zero Data Retention is not currently available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/s2-pro)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.fish.audio/developer-guide/models-pricing/models-overview)
- [Provider pricing](https://docs.fish.audio/developer-guide/models-pricing/models-overview)