# gpt-realtime-whisper

gpt-realtime-whisper transcribes live audio as people speak. Text arrives while the sentence is still in progress, which is what live captions and call monitoring need.

- **Model ID:** `openai/gpt-realtime-whisper`
- **Type:** transcription
- **Providers:** openai
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** $0.000284/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/gpt-realtime-whisper

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { experimental_streamTranscribe as streamTranscribe } from 'ai';
import { createGateway, gateway } from '@ai-sdk/gateway';
import { readFile } from 'node:fs/promises';

const modelId = 'openai/gpt-realtime-whisper';

// Mint this on your server, then send only the short-lived token to the client.
const { token } = await gateway.experimental_transcription.getToken({
  model: modelId,
});
const clientGateway = createGateway({ apiKey: token });

// Raw 24 kHz, 16-bit signed little-endian mono PCM audio.
const bytes = await readFile('audio.pcm');
const audio = new ReadableStream<Uint8Array>({
  start(controller) {
    controller.enqueue(new Uint8Array(bytes));
    controller.close();
  },
});

const result = streamTranscribe({
  model: clientGateway.transcription(modelId),
  audio,
  inputAudioFormat: { type: 'audio/pcm', rate: 24000 },
});

for await (const part of result.fullStream) {
  if (part.type === 'transcript-delta') {
    process.stdout.write(part.delta);
  }
}

console.log('\nFinal:', await result.text);
```

## About

gpt-realtime-whisper arrived on May 7, 2026 as OpenAI's streaming speech-to-text model for realtime transcription. Whisper-generation models transcribe completed audio, which suits recordings and post-session processing. gpt-realtime-whisper is the streaming counterpart, built for live audio where the transcript has to appear during the conversation.

Latency and accuracy are tunable rather than fixed. A delay setting controls how much audio gpt-realtime-whisper hears before emitting text: lower values surface partial text sooner, and higher values give the model more context and improve transcript quality. You can also pass the languages you expect and vocabulary hints, so product names, acronyms, and identifiers come through correctly.

gpt-realtime-whisper accepts audio and text input and returns text, with a context window of 0 tokens and up to varies of transcript output per turn. Pricing is based on audio duration rather than text tokens, so cost tracks minutes of speech and forecasts cleanly from call volume. Current rates are listed on this page.

Audio support on AI Gateway is in beta through AI SDK 7. Your server mints a short-lived token and the client streams audio over a WebSocket connection, so your AI Gateway API key never reaches the browser. You get the same authentication, observability, and spend controls as your text models, with no markup on provider pricing.

## What to consider

gpt-realtime-whisper runs in a persistent streaming session rather than a file upload, so plan for a connection that stays open for the length of the audio. Your interface also has to handle revision, because early partial text can change as more audio arrives.

A delay setting controls the tradeoff. Lower delay produces earlier text, and higher delay gives the model more audio context before it emits, which improves transcript quality. Tune it against production-like audio, including telephony, accents, background noise, and your own domain vocabulary, rather than clean samples.

## When to use

### Best For

- **Live Captions:** Conferences, webinars, and broadcasts that need text on screen as people speak
- **Call Monitoring:** Support and sales calls transcribed during the call for supervision or analytics
- **Meeting Documentation:** Notes captured while the conversation happens instead of afterward
- **Voice Input Feedback:** Interfaces that show users their words appearing as they speak
- **Duration-Based Budgeting:** Cost that tracks minutes of audio rather than token counts

### Consider Alternatives When

- **Recorded File Transcription:** `whisper-1`, `gpt-4o-transcribe`, and `gpt-4o-mini-transcribe` handle completed audio
- **Spoken Replies Required:** The gpt-realtime family answers in speech rather than returning text only
- **Speech Generation Needed:** `tts-1` and `tts-1-hd` turn written text into spoken audio

## Best for

- **Live Captions:** Conferences, webinars, and broadcasts that need text on screen as people speak
- **Call Monitoring:** Support and sales calls transcribed during the call for supervision or analytics
- **Meeting Documentation:** Notes captured while the conversation happens instead of afterward
- **Voice Input Feedback:** Interfaces that show users their words appearing as they speak
- **Duration-Based Budgeting:** Cost that tracks minutes of audio rather than token counts

## Consider alternatives

- **Recorded File Transcription:** `whisper-1`, `gpt-4o-transcribe`, and `gpt-4o-mini-transcribe` handle completed audio
- **Spoken Replies Required:** The gpt-realtime family answers in speech rather than returning text only
- **Speech Generation Needed:** `tts-1` and `tts-1-hd` turn written text into spoken audio

## Frequently asked questions

### What does gpt-realtime-whisper do?

gpt-realtime-whisper transcribes live audio into text while someone is still speaking. Transcript text arrives incrementally through a streaming session rather than after the recording ends.

### How is gpt-realtime-whisper different from Whisper?

Whisper-generation models transcribe completed audio, which fits recordings and post-session processing. gpt-realtime-whisper is the streaming counterpart, built for live audio where the transcript has to appear during the conversation.

### How do I access gpt-realtime-whisper through AI Gateway?

Audio support is in beta through AI SDK 7. Mint a short-lived token on your server with the AI SDK's gateway provider, then stream audio from the client over a WebSocket connection. Your AI Gateway API key never reaches the browser.

### Can I trade latency for transcript accuracy?

Yes. A delay setting controls how much audio gpt-realtime-whisper hears before emitting text. Lower delay surfaces partial text sooner, and higher delay gives the model more context and improves transcript quality.

### How do I improve accuracy on names and product terms?

Pass vocabulary hints and the languages you expect in the audio. Product names, acronyms, and identifiers come through more reliably when gpt-realtime-whisper knows to expect them.

### What context window does gpt-realtime-whisper support?

0 tokens, with up to varies of transcript output per turn. Streaming transcription commits audio in turns, so a long session produces many turns instead of one large response.

### How is gpt-realtime-whisper priced?

By audio duration rather than text tokens. Rates are listed on this page and update when providers change list prices, and AI Gateway adds no markup.

### Does gpt-realtime-whisper support zero data retention through AI Gateway?

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.

### What are typical latency characteristics?

This page shows live performance metrics measured across real AI Gateway traffic. Your delay setting also shapes how quickly partial transcript text appears.

## Links

- [Model page](https://vercel.com/ai-gateway/models/gpt-realtime-whisper)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://developers.openai.com/api/docs/models/gpt-realtime-whisper)