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": ""
  },
  "diarize": true,
  "timestampGranularities": ["segment"],
  "language": ""
}

For uploaded files: "type": "file", "fileId": "" 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": "",
  "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.

Option B — mistral_chat with json_schema (core profile, no extra flags needed):

{
  "model": "mistral-small-latest",
  "temperature": 0,
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "speaker_classification",
      "schema": {
        "type": "object",
        "properties": {
          "action_items": { "type": "array", "items": { "type": "string" } },
          "decisions": { "type": "array", "items": { "type": "string" } },
          "open_questions": { "type": "array", "items": { "type": "string" } }
        },
        "required": ["action_items", "decisions", "open_questions"]
      }
    }
  },
  "messages": [
    {
      "role": "user",
      "content": "\n\nExtract: (1) action items — tasks explicitly committed to, with any deadlines. (2) decisions — choices confirmed or made. (3) open questions — unresolved questions raised. Base your extraction only on the text above."
    }
  ]
}

Step 4 — Build per-speaker dispatch

Aggregate results into a per-speaker structure:

• Action items: tasks explicitly committed to, with any mentioned deadlines
• Decisions: choices confirmed or made
• Open questions: unresolved questions raised

Step 5 — Output

Present a summary table followed by a flat TODO list:

MEETING DISPATCH
────────────────────────────────────────
Duration: [Xmin]  Speakers: [N]  Segments: [N]

| Speaker | Action Items | Decisions | Open Questions |
|---------|-------------|-----------|----------------|
| spk_0   | 2           | 1         | 1              |
| spk_1   | 1           | 0         | 2              |

ACTIONS BY SPEAKER
──────────────────
[spk_0]
  □ [action item text] [deadline if mentioned]

DECISIONS
─────────
  • [decision text] (spk_X)

OPEN QUESTIONS
──────────────
  ? [question text] (spk_X)

Offer to pass the action items to french_meeting_minutes for a formatted compte-rendu, or to mistral_chat to draft follow-up emails per speaker.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Swih
• Source: Swih/mistral-mcp
• License: MIT

Install and usage instructions live in the source repository linked above.