Transcription diagnostics in the Sessions grid

Proposal, drawn on real measurements: five public recordings in five languages, transcribed through the shipped stage 5 with the style prompt and the skipped-window net. Numbers only; no transcript text is on this page. The grid and the slice under a row are the shipped CSS; the diagnostics block inside the slice is new and proposed.

Sessionsartefact

Commentary

What each strip is. All strips share one time scale across sessions, as the session timeline does, so sessions compare by eye. Sound: the loudest tenth of each 10 s (90th-percentile level), with the recording's noise floor dashed. Pace: words a minute. Doubt: share of words Whisper gave under 0.5 confidence, darker is more. Gaps: stretches with sound but almost no words; outlined ones the net re-transcribed, filled ones are still empty.

Placement. Under the session timeline, in the same slice, collapsed by default behind its own disclosure triangle, and present at all only when the Diagnostics menu is on (the user's preference, off by default). The two headline numbers stay beside the heading while the block is closed.

The headline. Beside "Diagnostics", always: the share of decode windows Whisper had to retry, and how long the recording has sound but no words. Of everything measured, these two separated the worst recording (a field tape: 59% retried, 8 minutes with sound but no words) from the rest (1–8% retried, seconds at most). Mean word confidence did not: the field tape scored 0.87, a clean interview 0.86, because it only averages the words Whisper wrote and skipped speech has none. It stays in the Confidence panel as background.

What it is not. An accuracy score. None of these numbers knows what was said; they say where to listen. Overlapping speech, the commonest cause of lost words, is not here yet: the voice pass would have to report it.

Placeholders. Thumbnails, dates and speakers are mockup filler; the session timeline lanes that sit above this block in the shipped slice are left out to keep the page about diagnostics.