Transcription

Overview#

Minutes uses Whisper Large V3 via Groq to transcribe meeting audio into structured, speaker-attributed text. Each transcript is divided into time-stamped segments with speaker labels, enabling precise attribution and searchability.

Upload Workflow#

Transcription follows a two-step process:

1. Create the meeting with metadata (title, type, participants, language)
2. Upload the audio file to trigger transcription

Step 1: Create the Meeting#

curl -X POST http://localhost:8013/api/v1/meetings \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{
    "title": "Ministry AI Strategy Discussion",
    "meeting_type": "committee",
    "language": "ar",
    "participants": [
      "Dr. Ahmed Al Mansoori",
      "Eng. Fatima Al Hashimi",
      "Mohammed Al Zaabi"
    ],
    "scheduled_at": "2024-01-15T14:00:00Z"
  }'

Meeting Types:

Step 2: Upload Audio#

curl -X POST http://localhost:8013/api/v1/meetings/{meeting_id}/upload \
  -H "Authorization: Bearer $TOKEN" \
  -F "file=@committee_meeting.mp3"

The upload triggers automatic transcription. The meeting status transitions from created to transcribing to transcribed.

Transcript Structure#

Retrieve the transcript after processing:

curl http://localhost:8013/api/v1/meetings/{meeting_id}/transcript \
  -H "Authorization: Bearer $TOKEN"

Response:

{
  "meeting_id": "mtg-a1b2c3",
  "segments": [
    {
      "speaker": "Speaker 1",
      "start": 0.0,
      "end": 12.5,
      "text": "بسم الله الرحمن الرحيم، نبدأ اجتماع اليوم بمناقشة استراتيجية الذكاء الاصطناعي"
    },
    {
      "speaker": "Speaker 2",
      "start": 13.0,
      "end": 28.3,
      "text": "شكراً، أود أن أبدأ بعرض التقدم المحرز في المرحلة الأولى"
    }
  ],
  "full_text": "...",
  "language": "ar",
  "duration": 3540.0
}

Segment Fields#

Speaker Diarization#

Minutes identifies distinct speakers in the audio and labels segments accordingly. When participant names are provided in the meeting creation, the system maps speaker labels to names based on speaking patterns.

Diarization works best with:

• Clear audio with minimal background noise
• Distinct speakers with minimal crosstalk
• Meetings with 2-10 participants

Language Support#

Minutes supports three language modes:

Audio Requirements#

Higher quality audio produces more accurate transcriptions. For best results, use a dedicated meeting recording device or conferencing system with good microphone placement.