AI Transcription¶

AI transcription is available in beta. When enabled, a recording session is automatically transcribed and the result is delivered to a LaSuite Docs instance, where the room owner can read and edit the transcript.

Prerequisite: Transcription requires the Recording feature to be fully set up and working first. Transcription uses the same LiveKit Egress, MinIO, and webhook infrastructure.

Known limitations:

• Participant identification is not implemented. Speakers are labeled generically (e.g., SPEAKER_00, SPEAKER_01)
• Transcription relies on WhisperX, which does not provide an OpenAI-compatible API

Hard dependency: LaSuite Docs Transcription output is sent to a LaSuite Docs instance (POST /create-for-owner). Without a running LaSuite Docs service, transcription cannot deliver its results. This is a hard dependency of the current implementation.

How it works¶

sequenceDiagram
  participant Backend as Backend API
  participant Summary as Summary Service
  participant Celery as Celery Worker (transcribe-queue)
  participant MinIO as MinIO (Object Storage)
  participant STT as WhisperX API
  participant Docs as LaSuite Docs

  Backend->>Summary: POST /api/v1/tasks/ (bearer token)
  Note right of Backend: Payload: owner_id, filename, email, sub, room, recording_date, recording_time

  Summary->>Celery: Register task (transcribe-queue)
  Celery->>MinIO: Fetch audio file
  Celery->>STT: Transcribe audio (WhisperX)
  STT-->>Celery: Segmented transcript
  Celery->>Celery: Format transcript
  Celery->>Docs: POST /create-for-owner (title, content, email, sub, api token)
  Docs-->>Celery: Acknowledgement

Docker Compose setup¶

Step 1: Configure the summary service¶

The summary service is the FastAPI application that handles transcription tasks. It is published on Docker Hub as lasuite/meet-summary.

Create its environment file env.d/summary:

# Redis (dedicated instance for the summary service)
REDIS_URL=redis://redis-summary:6379/0

# MinIO/S3 (to download audio recordings)
AWS_S3_ENDPOINT_URL=http://minio:9000
AWS_S3_ACCESS_KEY_ID=minioadmin
AWS_S3_SECRET_ACCESS_KEY=minioadmin123
AWS_STORAGE_BUCKET_NAME=meet-media-storage
AWS_S3_SECURE_ACCESS=False

# Meet backend (to report transcription results)
WEBHOOK_URL=https://meet.example.com/api/v1/tasks/callback/
WEBHOOK_API_TOKEN=internal-api-token

# WhisperX STT API
WHISPERX_BASE_URL=https://your-whisperx-instance.example.com
WHISPERX_ASR_MODEL=large-v3
WHISPERX_API_KEY=your-api-key

You need a running WhisperX instance. An Open Source implementation combining WhisperX and FastAPI is available at github.com/suitenumerique/meet-whisperx.

Step 2: Add to compose.yml¶

redis-summary:
  image: redis:7
  restart: unless-stopped
  networks:
    - internal

app-summary:
  image: lasuite/meet-summary:latest
  restart: unless-stopped
  env_file:
    - env.d/summary
  depends_on:
    - redis-summary
  networks:
    - internal

celery-summary-transcribe:
  image: lasuite/meet-summary:latest
  restart: unless-stopped
  command: celery -A summary.core.celery_worker worker --pool=solo -Q transcribe-queue
  env_file:
    - env.d/summary
  depends_on:
    - redis-summary
    - minio
  networks:
    - internal

Step 3: Connect the Meet backend to the summary service¶

Add to your .env:

SUMMARY_SERVICE_ENDPOINT=http://app-summary:8000/api/v1/tasks/
SUMMARY_SERVICE_API_TOKEN=internal-api-token

Restart the backend:

docker compose up -d --force-recreate backend

Step 4: Configure the WhisperX STT backend¶

The summary service connects to a WhisperX API for speech-to-text. Point it to your WhisperX instance in env.d/summary:

WHISPERX_BASE_URL=https://your-whisperx-instance.example.com
WHISPERX_ASR_MODEL=large-v3
WHISPERX_API_KEY=your-api-key

For production, a GPU-enabled WhisperX instance is strongly recommended. Transcription on CPU is very slow for anything beyond short recordings.

Kubernetes setup¶

Transcription in Kubernetes uses the summary service and Celery workers included in the Meet Helm chart. The recording infrastructure (MinIO, Egress, ingressMedia) must be set up first. See Recording: Kubernetes setup.

Step 1: Enable the summary service¶

Add to your values.yaml:

summary:
  replicas: 1
  envVars:
    REDIS_URL: "redis://redis-master:6379/1"
    AWS_S3_ENDPOINT_URL: "http://minio:9000"
    AWS_S3_ACCESS_KEY_ID: "minioadmin"
    AWS_S3_SECRET_ACCESS_KEY: "minioadmin123"
    AWS_STORAGE_BUCKET_NAME: "meet-media-storage"
    AWS_S3_SECURE_ACCESS: "False"
    WEBHOOK_URL: "https://meet.example.com/api/v1/tasks/callback/"
    WEBHOOK_API_TOKEN: "internal-api-token"
    WHISPERX_BASE_URL: "https://your-whisperx-instance.example.com"
    WHISPERX_ASR_MODEL: "large-v3"
    WHISPERX_API_KEY: "your-api-key"

You need a running WhisperX instance reachable from within the cluster. For production, a GPU-enabled instance is strongly recommended.

Step 2: Enable the Celery transcribe worker¶

celeryTranscribe:
  instances:
    - name: transcribe-1
      replicas: 1

celeryBackend:
  replicas: 1

Step 3: Connect the Meet backend to the summary service¶

Add to backend.envVars in your values.yaml:

backend:
  envVars:
    SUMMARY_SERVICE_ENDPOINT: "http://meet-summary:8000/api/v1/tasks/"
    SUMMARY_SERVICE_API_TOKEN: "internal-api-token"

Apply the updated chart:

helm upgrade meet meet/meet --namespace meet --values values.yaml

Full configuration reference¶

Supported audio formats¶

The summary service accepts: .mp4, .webm, .wav, .mp3, .ogg.

LLM summarization (optional, separate feature)¶

Summarization (generating a written summary from the transcript using an LLM) is a separate optional feature, not part of basic transcription. It requires an additional Celery worker and LLM API access. Refer to the upstream summarization documentation for setup details.