Caption Animation

Caption Animation

Turn a transcript or voiceover into word-timed, scroll-stopping captions: each word pops in on the syllable, the active word highlights, and the text stays glued to the narration. Build the kind of captions that drive watch-time on TikTok, Reels, and Shorts — readable, on-beat, and inside the safe area.

When to use

• Word-by-word or karaoke captions on short-form vertical video (9:16).
• Burning open captions onto a voiceover/narration track from a transcript.
• Turning a Whisper/SRT transcript into animated text in Remotion or on the web.
• Restyling auto-generated subtitles into a branded, "Hormozi-style" reveal.

The pipeline

Two non-negotiable rules

1. Word-level timing, not line-level. Karaoke reads as magic only when each word lands on the syllable. Always transcribe to word timestamps; never fake them by splitting a line evenly over its duration — drift is instantly visible.
2. Readability beats style. Captions are read on a phone, in sunlight, muted. Bold sans-serif, heavy stroke or shadow, high contrast first; decoration second. A caption nobody can read is decoration, not a caption.

Word timing from a transcript

Whisper emits word-level timestamps directly. Normalize every source into one flat token shape so the renderer never cares where the words came from:

// Caption token — the one shape everything downstream consumes
type Token = { text: string; startMs: number; endMs: number };

// Parse an SRT cue block "00:00:01,200 --> 00:00:01,640" into ms
const toMs = (t: string) => {
  const [h, m, rest] = t.split(":");
  const [s, ms] = rest.split(",");
  return ((+h * 60 + +m) * 60 + +s) * 1000 + +ms;
};

If only sentence-level cues exist, re-transcribe for word timing — do not interpolate. Whisper's medium.en is ~2x faster than large and accurate enough for clean voice audio.

Per-word pop-in (Remotion)

Drive each word's scale and opacity off a spring anchored to its own startMs. The micro-overshoot is what makes it feel "alive".

import { useCurrentFrame, useVideoConfig, spring, interpolate } from "remotion";

const Word: React.FC<{ token: Token; active: boolean }> = ({ token, active }) => {
  const frame = useCurrentFrame();
  const { fps } = useVideoConfig();
  const enter = (token.startMs / 1000) * fps;            // word's own entrance frame
  const p = spring({ frame: frame - enter, fps, config: { damping: 12, mass: 0.6 } });
  const scale = interpolate(p, [0, 1], [0.6, 1]);        // pop from 60% → 100% (overshoots)
  return (
    <span style={{
      display: "inline-block",
      transform: `scale(${scale})`,
      opacity: interpolate(p, [0, 1], [0, 1]),
      color: active ? "#FFE45E" : "#FFFFFF",             // active-word highlight
      transition: "color 80ms linear",
    }}>{token.text}&nbsp;</span>
  );
};

Compute active by testing frame against each token's [startMs, endMs]. Keep the highlight a single accent color or a filled "box" behind the live word — never animate every word's color at once.

Pages, not walls of text

Show 1–4 words at a time. Use Remotion's createTikTokStyleCaptions() (or group manually) and tune combineTokensWithinMilliseconds: ~200–500ms for true word-by-word, ~1000–1500ms for short readable phrases.

import { createTikTokStyleCaptions } from "@remotion/captions";
const { pages } = createTikTokStyleCaptions({
  captions, combineTokensWithinMilliseconds: 1200,       // ~2–3 words per page
});

Web/CSS pop (no Remotion)

For a DOM/live player, give each word its own animation-delay equal to startMs:

.word { display:inline-block; opacity:0; animation: pop .26s cubic-bezier(.34,1.56,.64,1) forwards; }
@keyframes pop { from { opacity:0; transform:translateY(.18em) scale(.7) } to { opacity:1; transform:none } }
.word.active { color:#FFE45E; }

words.forEach(w => { const el = mk(w.text); el.style.animationDelay = `${w.startMs}ms`; track.append(el); });

Readable type and safe-area placement (9:16, 1080×1920)

Place captions in the center band — not pinned to the very bottom, where the platform UI (username, audio, buttons) lives. Pad text to a max of ~2 lines.

Sync to a voiceover / narration

The transcript already carries the timing of the exact audio. Lay that same audio under the composition and keep timestamps in the audio's timebase — captions and voice stay locked with zero manual nudging. If the VO is re-recorded, re-transcribe; never hand-shift offsets.

Burn-in vs sidecar SRT/VTT

Emit both: burn the animated captions for social, and write a plain SRT/VTT from the same tokens for accessible/SEO playback.

Output checklist

• Word-level timestamps (real, from transcription) — no even-split fakery.
• Per-word pop with micro-overshoot; one accent color for the active word.
• 1–4 words per page; never a wall of text.
• Bold sans, 56–80px, 2–6px stroke + shadow, high contrast.
• Center band, clear of bottom ~280px and side rails.
• Audio laid under composition; timestamps in the audio timebase.
• Ship burn-in for social and a sidecar SRT/VTT for accessibility.

Deliver & verify (rendered stills → MP4)

Packaged helper (scripts/): tile your stills with scripts/contact-sheet.sh sheet.png f-hook.png f-mid.png f-end.png, then assert the encode with scripts/probe-mp4.sh out.mp4 [WxH] [fps]. See scripts/README.md.

Captions ship as a Remotion composition (<Composition> + zod schema + defaultProps) — all word motion frame-driven off useCurrentFrame(), never Date.now() / Math.random() / timers. Deliverable = out/*.mp4 (burned-in) + the project + the sidecar SRT/VTT. 9:16 vertical (1080×1920) is the default.

Verify loop — render stills → inspect → encode. Word timing is the thing that breaks; check it at exact frames before you spend an encode.

# Stills at start / a sampled active-word frame / end — WITH SHIPPED PROPS (real tokens + audio)
npx remotion still Captions out/f-start.png --frame=0   --props='{"captionsSrc":"vo.json"}'
npx remotion still Captions out/f-mid.png   --frame=90  --props='{"captionsSrc":"vo.json"}'
npx remotion still Captions out/f-end.png   --frame=N   --props='{"captionsSrc":"vo.json"}'  # N = durationInFrames-1

# Inspect each PNG:
#  - the word highlighted at frame 90 is the word whose [startMs,endMs] contains 90/fps (no drift)