# Muse Voice Transcribe 1.0 (Realtime) - Core.Today AI API

> Meta's Muse Voice Transcribe over a WebSocket - stream 16-bit PCM (24 kHz or 16 kHz mono) and receive partial and final transcripts while you speak, with endpointing and diarization. Billed per second of audio Meta processed (0.11625 credits/s, 35 credits per 5-minute block), sessions up to 20 minutes.

- **Provider**: Meta
- **Model ID**: meta/muse-voice-transcribe-1.0/realtime
- **Category**: Audio & TTS
- **Credits**: 35 per 5-minute block (0.11625 credits/s)
- **Speed**: Fast
- **Quality**: High

## Features
- Streaming partial and final transcripts while the caller is still speaking
- 25 languages with code-switching, Korean included
- Speaker diarization and endpointing in DIARIZATION/ENDPOINTING modes
- Keyword and language biasing, same as the file transcription endpoint
- Sessions up to 20 minutes; settlement follows Meta's measured audio, not wall-clock time

## Use Cases
- Live captioning or subtitles for a call or presentation
- Voice-driven interfaces that need transcripts while the user is talking
- Real-time meeting transcription with speaker turns as they happen

## API Endpoint

Base URL: https://api.core.today/v1
Create Prediction: POST /predictions
Get Status: GET /predictions/{job_id}

## Authentication
Header: X-API-Key: YOUR_API_KEY

## Input Parameters

### Optional
- **audioEncoding**: string (default: PCM_24KHZ) - Raw PCM 16-bit mono sample rate of the binary frames you send.
  Options: PCM_24KHZ, PCM_16KHZ
- **mode**: string (default: PUSH_TO_TALK) - PUSH_TO_TALK returns one running transcript; ENDPOINTING splits into turns at speech boundaries; DIARIZATION additionally labels speakers.
  Options: PUSH_TO_TALK, ENDPOINTING, DIARIZATION
- **partialMode**: string (default: CUMULATIVE) - CUMULATIVE partials replace the previous one; DELTA sends only new text.
  Options: CUMULATIVE, DELTA
- **emitAudioProgress**: boolean (default: false) - Forward Meta's ~80ms audioProgress events (audioProcessedMs) to your client.
- **keywords**: array - Names, acronyms or product terms to bias recognition toward.
- **language_bias**: array - Language hint (steers recognition, does not force it). Omit for automatic detection across all 25 languages.
  Options: Arabic, Bengali, Dutch, English, French, German, Hebrew, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malay, Mandarin Chinese, Marathi, Polish, Portuguese, Spanish, Tagalog, Tamil, Telugu, Thai, Turkish, Vietnamese

## Examples

### Handshake frame
The first text frame you send after the WebSocket connects, in ENDPOINTING mode with a Korean language hint. Never put a real API key in client-side/browser code - use a server-side connection instead.
```json
{
  "model": "meta/muse-voice-transcribe-1.0/realtime",
  "input": {
    "model": "meta/muse-voice-transcribe-1.0/realtime",
    "audioEncoding": "PCM_24KHZ",
    "mode": "ENDPOINTING",
    "partialMode": "CUMULATIVE",
    "emitAudioProgress": false,
    "keywords": [
      "Core.Today"
    ],
    "languageBias": [
      "Korean"
    ],
    "customerId": "member-123"
  }
}
```

## Response Format
```json
{
  "job_id": "abc123",
  "status": "pending | processing | completed | failed",
  "result": "URL or data (when completed)"
}
```

## Usage Flow
1. POST /predictions with model and input -> receive job_id
2. GET /predictions/{job_id} -> poll until status is completed or failed
3. Result contains output URL(s)

## Tips
- Send audio at real-time pace and pad silence when there's nothing to send - Meta closes the session after 10 seconds without audio ingress
- Keep sessions under the 20-minute cap and reconnect for longer calls rather than expecting one session to run indefinitely
- Use ENDPOINTING (or DIARIZATION for multi-speaker) for live captioning; PUSH_TO_TALK is meant for a single running transcript
- Never embed your API key in browser/client-side code - connect from your own server and relay audio from there

## Documentation
https://dev.meta.ai/docs/speech-to-text