> ## Documentation Index
> Fetch the complete documentation index at: https://docs.compose.market/llms.txt
> Use this file to discover all available pages before exploring further.

# Realtime

> Realtime audio, transcription, translation, and live sessions through the Responses API.

Compose supports realtime inference through the `/v1/responses` endpoint with realtime-capable models. The system includes 29 realtime models across multiple providers, a WebSocket vendor layer, and realtime-specific settlement.

## How it works

Realtime requests go through the same `/v1/responses` endpoint as text inference. When the resolved model is a realtime type (e.g., `gpt-realtime-2`, `qwen-omni-turbo-realtime`), the engine opens a WebSocket session to the provider, streams audio/text bidirectionally, and meters the result with realtime-specific billing.

```
POST /v1/responses
Authorization: Bearer compose-...
Content-Type: application/json

{
  "model": "gpt-realtime-2",
  "input": [{ "role": "user", "content": "Hello" }],
  "stream": true
}
```

## Realtime models

| Provider   | Models                                                                                                          | Operations                                        |
| ---------- | --------------------------------------------------------------------------------------------------------------- | ------------------------------------------------- |
| OpenAI     | `gpt-realtime-2`, `gpt-realtime-translate`, `gpt-realtime-whisper`                                              | Realtime transcription, translation, conversation |
| Alibaba    | `qwen-omni-turbo-realtime`, `qwen3-asr-flash-realtime`, `qwen3-omni-flash-realtime`, `qwen3-tts-flash-realtime` | Realtime ASR, TTS, omni-modal                     |
| Google     | `gemini-2.5-flash-native-audio-latest`, `gemini-3.1-flash-live-preview`, `gemini-3.5-live-translate-preview`    | Live audio, live translation                      |
| Deepgram   | `deepgram/flux-general-en`, `scribe_v2_realtime`                                                                | Realtime STT                                      |
| ElevenLabs | `scribe_v2_realtime`                                                                                            | Realtime speech-to-text                           |

Use `GET /v1/models?type=realtime` or `GET /v1/models/search?type=realtime` to list all realtime models.

## Appending input to a live session

For ongoing realtime sessions, append audio or text without starting a new response:

```
POST /v1/responses/{responseId}/input_items
Authorization: Bearer compose-...

{
  "type": "message",
  "role": "user",
  "content": [{ "type": "input_audio", "input_audio": { "data": "base64..." } }]
}
```

## Settlement

Realtime sessions are settled like any other inference call — the metering system extracts audio duration, token counts, and session metrics from the provider's usage format, then settles via x402. The receipt includes per-modality token breakdowns (input\_audio\_tokens, output\_audio\_tokens).

## Related

* [Streaming](/inference/streaming) — SSE streaming for text and media
* [Modalities](/inference/modalities/overview) — All supported modality types
* [Speech models](/inference/families/speech-models) — TTS and STT provider details
* [SDK: Inference](/sdk/inference/overview) — TypeScript API for inference