Paragraph capitalisation

How a transcript paragraph's first letter is drawn, especially after a split. The words are the 6 Oct 2026 screenshot of a real session (IKEA with uxfriends, s1), drawn from the transcription's word list as the page draws them today. Two splits are added in the participant's answer: one after a sentence ends, one in the middle of a sentence. Transcript CSS is the shipped CSS, inlined.

Artefact: the transcript as drawnCommentary: not the product“↵ split” marks where Return was pressed (mockup only)

The options, on the same paragraphs

Commentary

What causes it. The page draws a paragraph from its word timings when it has them, and those words are often lower-case and unpunctuated. The paragraph's stored text is not always better: in this session 17 of 42 paragraphs begin lower-case in the text itself ("um well I think…", "to um yeah…"), because the source transcript breaks turns mid-sentence. So the first letter is inconsistent before anyone splits anything. A split adds two cases: after a sentence ends, the second half's first word was a capital in the text but may be lower-case in the words; in the middle of a sentence, it is a continuation and should arguably stay lower-case.

A · As transcribed. Honest and changes nothing. It reads as raw output: a new speaker's turn starts "what do you think…", and a new sentence after a split starts "normally".
B · Every paragraph. What a tidied transcript looks like. Simple, display only. The cost is that a split in the middle of a sentence reads as two sentences: "So when it's fermenting the air / It's kind of a simple…".
C · Where a sentence begins. A capital when the paragraph starts a speaker's turn, or when the paragraph before it (same speaker) ended with . ? !; lower-case when the same speaker carries on mid-sentence. This is what B is for, without B's false sentence. It needs punctuation to see a sentence end. Where the words have none, the rule can only see the speaker change, so a split after an unpunctuated sentence stays lower-case. The stored text usually has the full stop, so the rule should read the text, not the words.
D · C, with a continuation mark. A muted "…" before a paragraph that carries on the previous one's sentence. It says the paragraph was cut, which is true and useful when reading a split, but it is a mark that is not the speaker's, and it appears after every mid-sentence split.

All four are display only. Nothing rewrites the stored text, which stays what the transcript said; exports and the agent read the text. Changing the stored text at split time (capitalising the second half when the cut follows a full stop) is a fifth option, not drawn: it fixes exports too, but it is the first time the transcript's words would be changed by anything other than the researcher typing.

Languages. Capitalising must use the session's language (toLocaleUpperCase), or Turkish "i" becomes "I" instead of "İ". Japanese, Chinese and Korean have no case, so A–C draw the same there and D's "…" is the only visible difference. Within a paragraph, Whisper's lower-case "i" ("i guess", "i'm") is untouched by every option here: that is a separate fix, and English-only.

Decided 7 Oct 2026: C (owner), built after measuring it on 3,237 real paragraphs (experiments/paragraph-case/), with a guard for words spelled lower-case on purpose (iPhone). Was the recommendation: C, reading sentence ends from the stored text, and from the words only where there is no text. It looks after a turn's start and a split after a sentence, and it leaves a mid-sentence split looking like the continuation it is. D's mark is worth trying only if researchers find mid-sentence splits confusing to read back.