The transcription pipeline: what happens and how long it takes

Before processing. Only the recording owner can start or regenerate AI processing. Open Profile → AI sharing → Review AI sharing. Select Summaries and extraction with Google Gemini, your needed transcription services, and their separate Allow background processing for this service checkboxes; save with Allow AI sharing. Search indexing has its own optional Google choice. A workspace role or a shared link does not let you approve someone else's processing.

The visible states. The Transcript tab shows Waiting for transcription and Pending while it checks for a saved result. Processing not confirmed means no recent progress has been confirmed. The indicator is not a percentage estimate or proof that a transcription service is working. A saved transcript confirms success; a current failure shows its reason.

AI naming when a transcript is missing. Open your recording and choose the wand beside its title (Rename with AI); on a phone, confirm with Yes. If the saved recording has no transcript, myScribe first transcribes its audio or video, then generates the title. Summary and action-item regeneration use the same sequence. Existing transcripts are reused. Keep the recording open until the requested title, summary or actions finish. Closing it or cancelling stops waiting and prevents the next step; already submitted transcription may continue. If you reopen after it completes, choose the wand again to finish naming.

The same owner and AI sharing background permissions apply. Missing media, refused permission, or a failed transcription stops generation and shows a reason. If transcription is already in progress, the action waits instead of starting another copy. Check the Transcript tab before retrying an unconfirmed request.

What actually runs.

1. myScribe picks a provider from the recording's mode (see Recording audio in the browser). Meeting/video → the diarizing engine, with speaker labels. Personal note → the fast single-track engine.
2. Your teammates' names are passed to the engine as custom vocabulary, which measurably improves name spelling in transcripts. Filling in team profiles improves transcript quality.
3. A provider error may be retried, but each send still needs your current permission. Fallback cannot use an unselected recipient; a missing, changed or unavailable choice stops the request.
4. Long files: transcription runs in a background worker with a 15-minute budget; a long-running engine may be handed to a second worker using the same permission. This is not a guarantee that a long file completes.
5. With the required background choices still valid, later processing can include: speaker identification (Speaker naming: Manage Speakers), entity extraction (spotting the people, companies and topics mentioned), task/summary/title generation, and vector embedding so the recording becomes searchable by the AI chat.
6. When it's all done you get a one-time push notification that your notes are ready (if you've allowed notifications).

Realistic timings. A short note is typically ready in under a minute. Longer meetings scale roughly with length; anything that needs the hand-off worker is several minutes. You can close the app: "transcription keeps running in the background and we'll notify you when it's ready."

Limits and gotchas.

• Meeting-mode transcripts never silently fall back to a non-diarizing engine. If the diarizing engine fails, the recording is marked failed rather than being re-done without speaker labels. That's deliberate, but it means "failed" sometimes means "we refused to give you a worse transcript".

• If progress cannot be confirmed while you are watching, the card shows Processing not confirmed and Check status. This does not prove that the background job failed.
• A retry waits for a new result; the earlier recovery or failure message is not treated as the outcome of the new attempt.
• A recording that has already failed shows "Transcription failed" in the Transcript tab, with the reason underneath and a Regenerate Transcript button. Only the owner sees the button; on a recording shared with you, ask the owner to regenerate it.
• For a stuck bot transcript, choose Check now or Transcribe now in the Transcript tab. Meeting capture with Recall needs its own background choice; any fallback transcription recipient needs a separate choice. The waiting screen reads existing results but does not automatically launch new fallback work. A queued request is not a completion receipt; wait for the saved result or a failure, then review before retrying.