Captions

Words and timings

• npx hyperframes transcribe <file> --json and a HeyGen voiceover (/media-use) give every word's start and end. Work from those as [{ "id": "w0", "text": "Hello", "start": 0.0, "end": 0.5 }, …].
• npx hyperframes transcribe defaults to the small.en whisper model, which translates non-English speech into English. Pass the model: npx hyperframes transcribe <file> --model small.en for English, --model small --language <code> otherwise, --model medium.en for speech over music. It writes assets/transcript.json (its JSON names the path) and overwrites it on the next run; --to srt or --to vtt writes a sidecar.
• For Chinese, Japanese and other unspaced scripts, local transcription returns a whole sentence as one "word", sometimes in the other script (Traditional for Simplified speech). Split each segment's text into short phrases and spread its time across them by character count; check the characters against what is said.
• Whisper invents "Thank you." over silence and repeats lines over a silent tail: drop words that start inside a final silence (silencedetect=noise=-35dB:d=0.6; real speech sits around −16 to −26 dB).
• Never put two words in one timed entry, and match transcript words to caption words by position, not text, or a repeated word takes the wrong time.
• Read the transcript before using it. More than 20% ♪ or � tokens, or nonsense words, means it failed: retry with a larger model, or take lyrics from the person. Drop non-word tokens before building groups.

Structure

• One sub-composition host with data-track-kind="captions" holds every caption group, so Studio shows them on one captions track.
• Ready-made styles: npx hyperframes catalog --tag caption-style --json. An installed one is source you can restyle.

Groups

• 2–3 words for high energy, 3–5 conversational, 4–6 calm. Break at a sentence end, a pause of 150 ms or more, or the word limit.
• One group visible at a time. A group opens at or before its first word's start and closes at or after its last word's end. A caption at exactly 0 s looks pre-existing: start 0.1–0.3 s in.
• Every group ends with a hard kill after its exit, or a stale group can stay on screen:

tl.to(group, { opacity: 0, duration: 0.12, ease: "power2.in" }, end - 0.12);
tl.set(group, { opacity: 0, visibility: "hidden" }, end);

tl.set is an instant flip, so setting visibility on the group is fine; never tween it on a .clip.

Fit and place

• Size each group to fit: window.__hyperframes.fitTextFontSize(text, { fontFamily, fontWeight, maxWidth }) returns { fontSize }. Defaults: maxWidth 1600 landscape, 900 portrait; base 78, minimum 42, step 2. When a word scales up for emphasis, divide maxWidth by that scale.
• A full-width absolute container, centred with text alignment. Never left: 50%; transform: translateX(-50%), which clips at the frame edges.
• overflow: visible, not hidden, which cuts off scaled words and glows.
• Landscape: 80–120 px from the bottom. Portrait: about 600–700 px from the bottom. Never over the speaker's face; inside the 80% title-safe box.
• Position each group absolutely. In a flex column a group hidden with visibility: hidden still takes space, so the next one lands lower. In a stack that accumulates, groups in one block share one exit time, or the stack reflows.
• Fade the group or its words, not both: the opacities multiply and read as a pop. A gradient or background-clip: text fill goes on the inner text span, not the transformed container.
• Over footage, choose the blend by the brightness (0–255) behind the text: under 60 screen, 60–180 overlay, over 180 normal with opaque colour. Plain subtitles above brightness 60 get a scrim sized to the text (30–40%), never a full-width bar.
• Footage may already carry burned-in subtitles only mid-clip: scan a 1 fps contact sheet before placing captions.

Text behind the speaker

• Stack the video, then the captions, then a cutout of the person from npx hyperframes remove-background (/media-use), ordered by z-index. Shake or punch in the video and the cutout together in one wrapper.
• check treats a <video> as opaque across its whole box, so a WebM cutout fails every caption behind it with text_occluded: put data-layout-allow-occlusion on the cutout and judge from snapshot.
• A big word behind the head reads best 30–55% covered: under about 15% it floats, over about 65% it is eaten. Re-lay out when a content word is 80% covered, or two adjacent words 50% each. Measure the rendered word boxes against the matte over each caption's own time window (heads move: one clip went from y=224 to y=96 in 8 s); estimates from character widths were wrong both ways.
• The matte model drops mic arms, lets large rigs near the body leak in, and flickers on held objects. Its first run downloads 168 MB, and it runs at about 2 fps at 1080p on CPU.