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

# Audio API

> Use OpenAI-compatible file transcription and progressively streamed text-to-speech.

The LLM Router includes OpenAI-compatible endpoints for bounded speech-to-text
and streamed text-to-speech. Use the same each::labs API key and base URL as
chat completions:

```text theme={"dark"}
https://api.eachlabs.ai/v1
```

<Warning>
  The audio API is in staged preview and is enabled per environment. A `404`
  response means that the capability is not enabled for your deployment yet.
  [Contact us](https://www.eachlabs.ai/contact) before depending on it in
  production.
</Warning>

## Current compatibility

| Capability         | Endpoint                        | Current preview models                                     | Behavior                                         |
| ------------------ | ------------------------------- | ---------------------------------------------------------- | ------------------------------------------------ |
| File transcription | `POST /v1/audio/transcriptions` | `openai/whisper-large-v3`, `openai/whisper-large-v3-turbo` | Synchronous JSON response; no prediction polling |
| Text-to-speech     | `POST /v1/audio/speech`         | `deepgram/aura-2`                                          | Progressive raw MP3 or PCM bytes                 |

Realtime WebSocket transcription is not part of this preview. File
transcription waits for the complete upload and returns one response. Speech
generation streams raw audio bytes, not Server-Sent Events.

## Transcribe an audio file

The transcription endpoint accepts one audio file as `multipart/form-data`.
Files may be up to 25 MB and must be FLAC, MP3/MPEG/MPGA, MP4/M4A, OGG, WAV,
or WebM.

<CodeGroup>
  ```python Python theme={"dark"}
  from openai import OpenAI

  client = OpenAI(
      api_key="YOUR_EACHLABS_API_KEY",
      base_url="https://api.eachlabs.ai/v1",
  )

  with open("speech.wav", "rb") as audio:
      transcript = client.audio.transcriptions.create(
          model="openai/whisper-large-v3",
          file=audio,
      )

  print(transcript.text)
  ```

  ```bash cURL theme={"dark"}
  curl https://api.eachlabs.ai/v1/audio/transcriptions \
    -H "Authorization: Bearer $EACHLABS_API_KEY" \
    -F "model=openai/whisper-large-v3" \
    -F "file=@speech.wav"
  ```
</CodeGroup>

The default response format is `json`. Use `response_format=verbose_json` to
request `language`, `duration`, and segment metadata. Word or segment
timestamps are available only with `verbose_json`:

```bash theme={"dark"}
curl https://api.eachlabs.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $EACHLABS_API_KEY" \
  -F "model=openai/whisper-large-v3" \
  -F "file=@speech.wav" \
  -F "response_format=verbose_json" \
  -F "timestamp_granularities[]=word" \
  -F "timestamp_granularities[]=segment"
```

## Stream text-to-speech

The speech endpoint begins returning raw audio while synthesis is still in
progress. Read the response body as a byte stream instead of waiting for a
prediction or parsing JSON.

The current preview profile supports:

| Model             | Voice              | Formats      | Maximum input    |
| ----------------- | ------------------ | ------------ | ---------------- |
| `deepgram/aura-2` | `aura-2-thalia-en` | `mp3`, `pcm` | 2,000 characters |

<CodeGroup>
  ```python Python theme={"dark"}
  from pathlib import Path
  from openai import OpenAI

  client = OpenAI(
      api_key="YOUR_EACHLABS_API_KEY",
      base_url="https://api.eachlabs.ai/v1",
  )

  with client.audio.speech.with_streaming_response.create(
      model="deepgram/aura-2",
      voice="aura-2-thalia-en",
      input="Hello from each labs.",
      response_format="mp3",
  ) as response:
      response.stream_to_file(Path("speech.mp3"))
  ```

  ```bash cURL theme={"dark"}
  curl --no-buffer https://api.eachlabs.ai/v1/audio/speech \
    -H "Authorization: Bearer $EACHLABS_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "deepgram/aura-2",
      "voice": "aura-2-thalia-en",
      "input": "Hello from each labs.",
      "response_format": "mp3"
    }' \
    --output speech.mp3
  ```
</CodeGroup>

If `response_format` is omitted, it defaults to `mp3`. The response
`Content-Type` identifies the returned audio:

* MP3: `audio/mpeg`
* PCM: `audio/pcm` with its sample rate and channel parameters

The response also includes `X-Execution-ID` for support and billing
correlation. When the provider supplies one, `X-Generation-Id` is included as
well.

## Validation and errors

Both endpoints use the OpenAI-compatible error envelope. Unsupported models,
voices, formats, and fields return `400` before a provider request is sent.
Common statuses include:

| Status | Meaning                                                     |
| ------ | ----------------------------------------------------------- |
| `400`  | Invalid or unsupported request                              |
| `401`  | Missing or invalid Bearer token                             |
| `402`  | Insufficient balance                                        |
| `404`  | Capability is not enabled for this deployment               |
| `413`  | File or request is too large                                |
| `429`  | Provider rate limit                                         |
| `503`  | Authentication or audio provider is temporarily unavailable |
| `504`  | Audio operation timed out                                   |

<Note>
  Supply API keys only as `Authorization: Bearer ...`. Query-string API keys
  are rejected so credentials do not leak into URLs or access logs.
</Note>