Caption Grouping

transcript.json from transcribe.cjs:

Drop type: "spacing" entries; you only need words.

Break boundaries

Cut a new group at ANY of:

1. Pause ≥ 500ms (gap between word.end[i] and word.start[i+1]) — speaker took a breath.
2. Sentence terminator — word ends with ., ?, !, or an em-dash-like pause.
3. Strong comma — , followed by pause ≥ 250ms.
4. Discourse reset — words like “but”, “so”, “and then”, “you know” starting a clause often merit their own or new group.
5. Group reaches 6 words OR 2.5 seconds — whichever first. Long groups feel like subtitles, not embedded typography.

Hard constraints:

• Minimum 2 words per group (1-word exceptions: interjections like “Wait.” or the crown line).
• Minimum 0.5s on screen. If a group is less, merge into neighbor.
• No overlapping groups — at most one visible at a time.

Timing the group

For a group with words w[0]..w[n-1]:

• in = w[0].start - 0.08 (enter slightly before first word)
• out = min(next_group.in - 0.05, w[n-1].end + 0.6) (linger ~0.6s after last word, but don’t collide with next)

The last group extends to the video end if needed.

Style & tone (cross-reference)

See typography-presets.md for how to pick style and tone per group. Work left-to-right through the groups and:

1. First group: default intro + soft.
2. Watch for emphasis signals (ALL CAPS in transcript is rare but possible; more often it’s semantic — superlatives, proper nouns).
3. Escalate tone into present once the monologue shifts from setup to statement.
4. Reserve crown for at most ONE group, typically the final line.

Editorial surgery is allowed

You do NOT have to caption every word. It’s fine to:

• Drop filler like extra “you know”s, “um”s, “I mean”s if they bloat the visual pace.
• Condense a 6-word run into 4 by cutting function words, as long as the meaning and timing remain truthful.
• Skip the whole thing during obvious silence or non-speech (laugh, music interlude).

Editorial rule: you are writing typography to support the speech, not a court transcript. Keep meaning, trim noise.

Example (champion)

Transcript: “You know, for me I’ve had this kind of upbringing, had the great foundation and, you know, I’ve achieved incredible things. I was dreaming of becoming number one in the world and becoming a Wimbledon champion”

Groups after editorial pass:

Plus the crown:

Notice “had” was dropped from cg-2 (“had the great foundation” → “the great foundation”), “I was” was dropped from cg-4, and “a” was dropped from crown — all for visual cadence.

word.start/end inside groups

Pass through the original timestamps from the transcript. Don’t retime individual words — only the group in/out. The word-level karaoke reveal inside each group uses the original w.start.