Meetings

Lira joins Google Meet calls as a real participant, listens in real-time, and responds with natural speech when addressed by name.

How It Works

1. Deploy — POST /lira/v1/bot/deploy with a meeting URL
2. Join — Playwright bot navigates to Google Meet, joins the call
3. Listen — Audio captured via WebRTC getUserMedia override, streamed to Nova Sonic
4. Identify — Deepgram provides real-time speaker diarization with name attribution
5. Respond — Nova Sonic generates speech responses, injected into meeting audio
6. Summarize — GPT-4o-mini generates meeting summaries after the call

Key Behaviors

Behavior	Description
Wake word	Only responds when addressed by name (configurable). 4-layer detection.
Speaker ID	Each transcript line tagged with the speaker's real name
Physical mic	Voice commands click the actual mic button in Google Meet
Barge-in	Stops talking immediately when interrupted
Auto-leave	Leaves after 45 seconds if alone in the meeting
Echo gate	Prevents Lira from hearing her own output

Settings

Update AI settings mid-meeting via:

PUT /lira/v1/meetings/:id/settings
Authorization: Bearer <jwt>

{
  "personality": "challenger",
  "wake_word_enabled": false,
  "summary_mode": "detailed"
}

Summaries

Two summary modes:

Mode	Output
short	4–6 sentence overview
detailed	400–700 word breakdown with per-person contributions

GET /lira/v1/meetings/:id/summary?mode=detailed

Task Extraction

Lira automatically extracts action items from the meeting and creates tasks. Tasks can be synced to Linear, GitHub, Slack, or email.

See Tasks API for details.