Transcripts & notes
How the processing pipeline turns raw audio into a speaker-labeled transcript and structured notes.
When you stop a recording, Ogma runs a pipeline with three stages. Each stage is idempotent and retryable — if anything fails (network, API outage, closing the laptop), you can retry from the meeting's detail view and it picks up exactly where it left off.
1. Transcription (Whisper)
Each 5-minute audio segment is sent to the OpenAI Whisper API and the results are stitched back together with corrected timestamps, producing one continuous, timestamped transcript.
2. Speaker labels & notes (Claude)
The full transcript is sent to the Claude API in a single call that returns two things:
- Speaker attribution — utterances get labeled
Speaker A,Speaker B,Speaker C, … inferred from the conversation text itself (who answers whom, names used, speaking style). - Structured notes — TL;DR, summary, key points, decisions, action items, open questions, and quote highlights that jump-play the corresponding audio in the app.
Speaker labels are best-effort
Ogma does not do acoustic diarization (voice fingerprinting). Labels are
inferred from the text, which works well for structured 2–4 person meetings
and degrades with heavy crosstalk. Utterances the model can't confidently
attribute are labeled Speaker ?.
Renaming speakers
In the meeting detail view you can rename Speaker A → Alice etc. Renames apply across the whole transcript and carry into Notion on the next sync.
3. Notion sync
If a Notion database is connected, the finished meeting is pushed there — see Notion sync.
Retrying and resuming
The pipeline derives its resume point from what's already stored:
- no transcript yet → transcription runs
- transcript but no notes → the Claude step runs
- notes but no Notion page → only the sync runs
So a retry after a failure never redoes (or re-bills) completed work.
Search
Every transcript is indexed with full-text search locally. You can search across all meetings from the app — and AI clients can do the same through the MCP server.