audio-dispatch

Audio dispatch

Transcribes a multi-speaker audio recording with Voxtral diarization, classifies each turn, and produces structured action items and decisions per speaker.

Profile requirements:

• voxtral_transcribe — available in core profile (default)
• mistral_chat (for classification fallback) — available in core profile
• files_upload (to upload a local audio file) — requires MISTRAL_MCP_PROFILE=full
• mistral_classify (optimized classification) — requires MISTRAL_MCP_PROFILE=full
• batch_create (bulk processing of many speakers) — requires MISTRAL_MCP_PROFILE=full

Core-compatible path: if running on the default profile, classification uses mistral_chat with json_schema instead of mistral_classify — same quality, slightly different call shape (see Step 3).

Steps

Step 1 — Get the audio file

Ask the user for one of:

• A public URL to an audio file (MP3, WAV, M4A, FLAC, OGG — up to ~2h) — works with core profile
• A local file path → upload with files_upload (requires MISTRAL_MCP_PROFILE=full), note the file_id
• A language hint (ISO 639-1 code, e.g. "fr", "en") — optional, improves accuracy

Step 2 — Transcribe with speaker diarization

Call voxtral_transcribe with diarization enabled:

{
  "audio": {
    "type": "file_url",
    "fileUrl": "<URL>"
  },
  "diarize": true,
  "timestampGranularities": ["segment"],
  "language": "<ISO 639-1 code, or omit for auto-detection>"
}

For uploaded files: "type": "file", "fileId": "<file_id>" instead.

structuredContent.segments contains turns with speakerId, text, start, end, score.

Step 3 — Classify speaker turns

Group segments by speakerId. For each speaker, concatenate their text.

Option A — mistral_classify (requires MISTRAL_MCP_PROFILE=full):

{
  "model": "ministral-3b-latest",
  "inputs": "<speaker text>",
  "labels": ["action_item", "decision", "open_question", "context", "social"]
}

For >5 speakers or very long transcripts, use batch_create (requires full) to process all speakers concurrently.