Speaker Diarization

Lira uses Deepgram Nova-2 for real-time speaker diarization — identifying who is speaking at any moment during the meeting.

How It Works

1. Meeting audio is sent to Deepgram in parallel with the Nova Sonic stream
2. Deepgram returns speaker indices (Speaker 0, Speaker 1, etc.) in real-time
3. Lira correlates these indices with participant names scraped from the meeting DOM
4. The correlation uses a 4-second alignment window to match Deepgram's speaker change events with visible participant highlights in Google Meet

Name Resolution

Deepgram Speaker Index  ←→  DOM Participant Name
        (4s window)

Step	Action
1	Deepgram reports a speaker change (e.g. "Speaker 2 started talking")
2	Playwright scrapes the Google Meet participant list DOM within a 4-second window
3	The active speaker highlight in the DOM is matched to the Deepgram speaker index
4	Once a mapping is established, it persists for the duration of the meeting

Fallback

If Deepgram is unavailable (e.g. API key not configured or service error), Lira falls back to DOM-only polling — periodically scraping the active speaker indicator from the meeting UI. This provides speaker attribution but with lower accuracy than the Deepgram-based approach.

Configuration

Variable	Description
DEEPGRAM_API_KEY	Deepgram API key (optional — falls back to DOM-only if not set)

Transcript Output

With diarization enabled, transcripts include speaker names:

{
  "segments": [
    { "speaker": "John Smith", "text": "Let's discuss the Q3 roadmap.", "timestamp": "00:02:15" },
    { "speaker": "Sarah Chen", "text": "I think we should prioritize the API redesign.", "timestamp": "00:02:22" },
    { "speaker": "Lira AI", "text": "Based on Linear, the API redesign has 3 open issues.", "timestamp": "00:02:30" }
  ]
}