Transcript readability: today against the proposed versions

IKEA with uxfriends, s1. Left: what the transcript page draws today. Right: the same stretches, re-transcribed locally with the proposed Whisper settings, then put through the version's paragraph and display rules. Pick a version at the top. The research is in docs/design-transcript-readability.md.

Artefact: the transcript as drawnCommentary: not the product

How this was made

Left column. The project's paragraphs from its database, with word timings joined by time exactly as the importer does (_enrich_words_from_intermediate), so a paragraph shows Whisper's word list when the words pass the 0.9 match and its stored text when they do not. In this project the stored text comes from a 16 Jul Whisper run and the words from a 27 Aug one. Where the stored text is drawn it carries its “(Speaker A)” label; I found nothing that strips it on the page.

Right column. Fresh local transcriptions of the same audio (mlx-whisper 0.4.3, large-v3-turbo, M2 Max, 6 Oct 2026; no paid calls). Speaker badges are borrowed: speaker identification was not re-run (its pass may call an LLM), so each new sentence takes the speaker of today's paragraph it overlaps most. Where today's speakers are wrong, the borrowed ones can be too.

Shown at full strength. The page draws participant paragraphs with no quote in them at 60% opacity. Every paragraph here is drawn as quoted, so the text is comparable. The dimmed filler style is the one new style on the page, and is a proposal.