/cue

Captions and Subtitle Authoring Reference

Purpose: Author subtitle, SDH (Subtitles for the Deaf and Hard-of-hearing), and closed-caption files with correct timing, readable line breaks, accessibility compliance, and the right delivery format per platform. A caption that can't be read in time is worse than no caption at all.

2026 baseline. Automated transcription (Whisper-large family, AssemblyAI, Deepgram Nova-3, ElevenLabs) consistently lands the first pass at word-level timing accuracy under 0.5s for English. The 2026 caption-authoring workflow is AI transcribes → human re-segments + re-times to phrase boundaries → CPS / line-length pass → ship. AI alone reliably misses: speaker turn boundaries, music / SFX brackets for SDH, brand-name casing, and CPS budgeting under reading-speed limits. Treat the AI output as raw material, never as a deliverable.

Scope Boundary

• cue captions: SRT / VTT / ASS authoring, timing, reading-speed limits, SDH additions, burn-in vs soft-sub decisions.
• cue localize (elsewhere): translated captions / dubbing scripts for multi-locale delivery — run after base captions are locked.
• cue narration (elsewhere): source narration script — captions derive from this.
• Polyglot (elsewhere): i18n key extraction for UI copy — not applicable to time-coded captions, but owns any surrounding UI strings.

If the hypothesis is "can a hearing-impaired viewer follow this video?" → captions. If it's "can a Spanish speaker follow this video?" → localize.

Format Selection

Default for web delivery: WebVTT. Default for YouTube/Vimeo upload: SRT.

Workflow

INGEST    →  receive narration script + final video timecodes (or timed transcript)
          →  confirm target format (SRT / VTT / ASS) and accessibility goal (captions vs SDH)

SEGMENT   →  split narration into caption cues at natural phrase boundaries
          →  enforce max chars per line and max lines per cue
          →  verify reading speed (CPS / WPM) for every cue

TIME      →  anchor in-time to spoken onset ±80ms
          →  hold cue for min 1.0s, max 7.0s
          →  leave ≥2 frames gap between consecutive cues (prevents flicker)

ENRICH    →  (SDH only) add [music], [laughter], [door slams], speaker IDs
          →  (multi-speaker) identify speakers: "- Alex: …" / ">> NAME:" / color coding
          →  mark non-translated proper nouns, preserve branded casing

VERIFY    →  run CPS scanner, line-length scanner, overlap detector
          →  viewer test: can the slowest reader finish each cue?
          →  export in target format; for broadcast, validate against platform spec sheet

Line and Reading Limits

CPS = characters per second, excluding spaces in most counters. Over-budget cues get split or the line is edited for brevity.

SDH vs Standard Captions

Standard captions transcribe spoken dialogue only. SDH adds everything else a hearing viewer perceives.

Deliver SDH when the platform requires WCAG 1.2.2 (pre-recorded captions) and 1.2.3 (audio description or transcript) compliance. ADA compliance deadline: as of April 24, 2026, US public entities serving populations over 50,000 must meet WCAG 2.1 Level AA (DOJ final rule, April 2024) — this covers video captions, audio description, and keyboard-accessible player controls [Source: testparty.ai/blog/video-captioning-requirements].

SRT Cue Example

12
00:01:23,480 --> 00:01:26,720
- Alex: We ship every Friday.
- Priya: Even during incident response?

13
00:01:26,800 --> 00:01:28,600
[phone rings]

Two frames gap: cue 12 ends at 26,720 and cue 13 starts at 26,800 (80ms). Cue 12 is a two-line multi-speaker cue with dash prefix.

Burn-in vs Soft-Sub Tradeoffs

Default for YouTube/Vimeo: soft-sub (upload SRT) so viewers toggle and search works. Default for TikTok / Reels / Shorts reshare: burn-in so captions survive download and repost.

Auto-Caption Polish Workflow

Platform auto-captions (YouTube ASR, Whisper, etc.) are a starting point, not a deliverable. Always polish:

1. Proper nouns pass: fix product names, people, places that ASR misheard.
2. Punctuation pass: ASR often omits commas; restore them where pacing needs them.
3. Homophone pass: "their / there / they're", "lead / led", product-specific terms.
4. Segment pass: ASR splits on silence, not on sense — merge and re-split on clause boundaries.
5. CPS pass: ASR-generated cues frequently exceed 21 CPS; split or edit.
6. Number / unit pass: "two thousand twenty four" → "2024"; "gigabytes" → "GB" when visual is tight.
7. Brand voice pass: enforce capitalization (GitHub not Github), hyphenation (sign-in not sign in), terminology.

Budget: ~2-3× video runtime for a full polish pass on a 3-minute video.

Anti-Patterns

• ❌ Three-line cues — steals video real estate, drops retention.
• ❌ Verbatim filler transcription ("uh", "um", "you know") in non-documentary content.
• ❌ Captions that lead the audio by more than 200ms — feels like spoilers and breaks sync perception.
• ❌ ALL-CAPS cues for emphasis (except short shouts) — cuts reading speed ~15%.
• ❌ Auto-captions uploaded without polish — WCAG compliance is about quality, not existence.
• ❌ Mixing burn-in and soft-sub on the same export — viewers get double captions on toggle.
• ❌ Translating captions before base SRT is locked — retiming translated cues costs more than the original pass.

Handoff

On completion, hand off:

• To localize: locked base-language SRT/VTT with speaker IDs, pronunciation notes for product names, glossary of non-translatable terms, and runtime caption-budget summary.
• To Cue production: burn-in position spec (safe area, font, size, stroke) if delivery is hardsub.
• To user: caption file in target format, CPS/line-length validation report, and a one-line note on any passages that required editorial compression to hit reading speed.