Transcription

After recording, open the Meeting Page panel and click Transcribe.

How it works

Audio is converted to a structured transcript with speaker labels and timestamps, saved as <recording>.transcript.json alongside the audio in your vault.

The route depends on your plan and your Transcription model choice in Settings → AI:

• Pro All-in-One / Trial — audio routes through Pensum’s worker, which streams it directly to Deepgram and returns the transcript. Nothing is persisted on Pensum’s side beyond a billing metadata row (account, duration, timestamp).
• Pro BYO — your plugin sends audio directly to the provider that owns the selected model. Pensum’s servers are not involved.
  • Deepgram Nova-3 (default, recommended): speaker diarization native, ~$0.26/hour
  • OpenAI Whisper / gpt-4o-transcribe: ~$0.36/hour

Privacy and data handling

Path	Where audio goes	Retention
Managed (All-in-One) under 20 min	Synchronous through Pensum’s worker → Deepgram	Held in worker memory for the request only. No disk, no R2 bucket, no D1 row referencing content.
Managed (All-in-One) 20+ min	Async via R2 staging → Deepgram	Audio + transcript deleted as soon as the plugin confirms receipt; 1-hour R2 TTL backstop covers disconnects.
BYO (Deepgram / OpenAI)	Direct from your device to the provider’s API	Subject to that provider’s retention policy; Pensum has no role.
Local storage	The recording (.webm) and transcript (.json) sit in your vault	You control the Audio retention setting (indefinite, 90 days, or 30 days) plus your own backups.

Public claim: we do not retain your audio, we do not retain your transcript, we log only billing metadata. The full data-flow diagrams are on the architecture page.

Local transcription (e.g. whisper.cpp) is a future direction, not in v1.0 — there’s no “Local transcription URL” setting yet.

Speaker identification

After transcription, click Identify Speakers. Pensum sends the first 5 minutes of transcript plus your attendees list to the AI, which maps speaker labels to names based on context clues (introductions, who addresses whom by name).

For this to work, add attendees to the meeting frontmatter:

attendees:
  - Jamie
  - Sam
  - Taylor

High and medium confidence mappings are applied automatically. Low confidence mappings are skipped.

Trial cap

The 14-day trial includes 3 hours of transcription total. Text AI features stay unbounded. The cap exists because transcription is the only AI feature with meaningful per-use cost; 3 hours covers serious evaluation (3–6 meetings) without making trial-rotation worth anyone’s time.

Requirements

For BYO mode, a key for your chosen transcription provider (Deepgram or OpenAI) configured in Settings → AI. For managed mode, an active Pro All-in-One subscription or trial.