LiveKit Local Participant

This service is a lightweight direct Python LiveKit RTC participant for job_tracker.

It is intentionally not a LiveKit Agents SDK worker. It does not use worker registration, dispatch, subprocess jobs, VAD, turn detection, or plugin lifecycles. Its only job in this phase is:

join local LiveKit room
    -> subscribe to browser microphone audio
    -> buffer PCM frames only during an active utterance window
    -> wait for utterance_start and utterance_end control packets
    -> fetch current canonical hotwords from the backend
    -> call whisper-service /transcribe
    -> publish final_transcript or transcription_error packets

Required Local Services

• livekit-server --dev
• jobtracker-BE
• whisper-service

Frontend microphone publishing is now handled by the PR 3 browser voice controls.

Environment

Copy .env.example to .env:

LIVEKIT_URL=ws://127.0.0.1:7880
LIVEKIT_API_KEY=devkey
LIVEKIT_API_SECRET=secret
LIVEKIT_ROOM_NAME=job-tracker-local
JOBTRACKER_BACKEND_URL=http://127.0.0.1:8000
WHISPER_SERVICE_URL=http://127.0.0.1:8100
WHISPER_REQUEST_TIMEOUT_SECONDS=120

OS environment variables override values from livekit-agent/.env.

WHISPER_REQUEST_TIMEOUT_SECONDS defaults to 120 seconds, must be a positive number, and is used for the Whisper /transcribe HTTP request only.

Setup

cd /home/aditya/dev-work/job_tracker_assistant/livekit-agent
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run

cd /home/aditya/dev-work/job_tracker_assistant/livekit-agent
source .venv/bin/activate
python agent.py

The participant connects with a fixed local identity:

job-tracker-local-agent

Structured Logs

Logs are written to stdout and include:

• startup and configuration validation
• room connect start / success / failure
• participant identity and room name
• track subscription events
• utterance ids
• recording gate open / close events
• buffer reset and audio duration
• hotword fetch start / success / failure
• Whisper request start / success / failure with configured timeout and request duration
• final transcript publish success
• clean shutdown

Secrets and JWTs are never logged.

Current Limitations

• No VAD or silence detection
• No partial transcripts
• No automatic submission to the typed transcript flow
• No LiveKit Agents SDK worker lifecycle