# Transcribe-1

Fish Audio's speech-to-text model. Transcribe-1 transcribes recorded audio across languages, detects the spoken language on its own, and can return word-level timing rather than a single block of text.

- **Model ID:** `fish-audio/transcribe-1`
- **Type:** transcription
- **Providers:** fish-audio
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** $0.0001/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/transcribe-1

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { experimental_transcribe as transcribe } from 'ai';
import { gateway } from '@ai-sdk/gateway';
import { readFile } from 'node:fs/promises';

const result = await transcribe({
  model: gateway.transcriptionModel('fish-audio/transcribe-1'),
  audio: await readFile('audio.mp3'),
});
```

## About

Transcribe-1 is Fish Audio's multilingual speech-to-text model, and the counterpart to the S-series voice models in the same family. Where those generate speech, Transcribe-1 transcribes it.

Language detection is automatic, so you do not have to declare the spoken language ahead of time. That matters for mixed-language archives and for user-generated audio where the language is not known in advance.

When alignment detail is requested, Transcribe-1 returns word-level timestamped segments rather than a single block of text. That is what subtitle generation, transcript search, and clip extraction all depend on, since each needs to map words back to positions in the audio.

Audio is supplied either as a URL or as base64. Speech-to-text through AI Gateway is billed by audio duration rather than by tokens, so cost tracks the length of the recording. See the pricing panel on this page for current rates.

## What to consider

Fish Audio publishes less public documentation for Transcribe-1 than for its voice models, and its own models overview covers the speech generation family rather than transcription. Verify supported languages, file size limits, and duration caps against your own test audio before committing a pipeline to it.

Billing follows audio duration, not tokens. A long recording costs the same whether it is dense speech or mostly silence, so trimming leading and trailing silence before upload is worth doing on large batches.

Transcribe-1 handles recorded audio. A live conversational surface needs a realtime speech model rather than a transcription pass over a finished file.

## When to use

### Best For

- **Recorded audio transcription** of calls, meetings, podcasts, and voice notes
- **Mixed-language archives** where automatic detection avoids declaring a language per file
- **Subtitle generation** using word-level timestamped segments
- **Transcript search** that maps matched words back to positions in the audio
- **Clip extraction** driven by word timing rather than manual scrubbing

### Consider Alternatives When

- **Live conversation** is the surface, where a realtime speech model fits instead
- **Documented limits** must be confirmed in advance, since public documentation is thin
- **Speech generation** is the task, where the S-series models in this family produce audio
- **A specific language guarantee** is required, which is not publicly enumerated for this model

## Best for

- **Recorded Audio Transcription:** Calls, meetings, podcasts, and voice notes
- **Mixed-Language Archives:** Automatic detection without declaring a language
- **Subtitle Generation:** Word-level timestamped segments
- **Transcript Search:** Matched words mapped back to audio positions
- **Clip Extraction:** Driven by word timing rather than manual scrubbing

## Consider alternatives

- **Live Conversation:** A realtime speech model fits a live surface better
- **Documented Limits:** Public documentation for this model is thin
- **Speech Generation:** The S-series models in this family produce audio
- **Guaranteed Language Support:** The supported set is not publicly enumerated

## Frequently asked questions

### What is Transcribe-1 built for?

Transcribing recorded audio into text. It detects the spoken language automatically and can return word-level timestamped segments.

### How do I supply audio to Transcribe-1?

As a URL or as base64-encoded audio.

### Can Transcribe-1 produce subtitles?

Yes. Request alignment detail and it returns word-level timestamped segments, which map each word back to its position in the audio.

### Do I need to specify the language?

No. Language detection is automatic, which suits mixed-language archives and user-generated audio where the language is not known ahead of time.

### How is Transcribe-1 billed?

By audio duration rather than by tokens, so cost tracks recording length regardless of how much speech it contains. See the pricing panel on this page for current rates.

### Can Transcribe-1 transcribe a live conversation?

It is built for recorded audio. A live conversational surface needs a realtime speech model instead.

### Does Transcribe-1 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/transcribe-1)
- [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)