Transcription and its reliable failures; captions as craft; the voice consent line.
Motion · animation · UX · all tracks — 4 hours, drop-in friendly
Audio, Voice, Video, and Multimodal Critique
The post-production assist kit — transcription, captions done properly, temp VO with consent rules — and the multimodal critique relay: useful, filtered feedback on any work-in-progress, any discipline, at 2am.
01This Week at a Glance
You will make: a captioned clip (or critique packet on a still) + a corrections table + a prioritized edit list.
The multimodal critique relay; sound briefs; the music rights wall.
Dirty audio recorded live → transcript error hunt → professional captions → the relay on a real WIP.
Your clip, your relay; error trophies and defended rejections.
Why this matters in real work
Captions aren't garnish: they're an accessibility deliverable with legal weight in the EU — and most social video plays silent, so captions are reach. And the critique relay generalizes past video: models can see now. Filtered by you, that's a second pair of eyes that never sleeps. The filtering is the skill.
02Key Concepts
1 · Transcription: the quiet workhorse
Captions, interview notes, edit-by-text, searchable footage. Its reliable failures cluster exactly where audiences notice: proper nouns ("NCAD" → "encad"), Irish names and places, craft terms ("risograph"), homophones.
The professional pattern
Machine transcribes → human corrects → the corrections table is kept (heard → corrected). That table is QA evidence and disclosure note in one document.
2 · Captions are craft (not autocaption sludge)
- ≤ 42 characters per line, 2 lines max.
- Break at sense boundaries — never mid-name.
- Keep the speaker's rhythm — captions preserve voice, they don't formalize it.
- Sound information when it matters: [door slams], [music swells].
- Watch once with the sound OFF — the silent-autoplay test.
3 · Temp VO and the voice consent line
Legitimate: scratch narration for animatics, accessibility reads of your own texts, placeholders you'll re-record — and phone-recorded human scratch is often better for timing anyway.
The brightest line in the course
Never clone a real person's voice without documented consent. Not classmates, not celebrities, not "just for the rough." This is hardening into law — and "it was temporary" does not work.
4 · The multimodal critique relay (works on everything)
- Screenshot the WIP — frame, layout, model sheet, UI.
- Three named reviewers — 3 observations + 1 priority fix each, no softening.
- You filter — accept/reject each point with written reasons. At least one critique will be confidently wrong; rejecting it well is the skill.
- Peer challenges one accept and one reject.
- Triage: blocker / should-fix / polish, with minutes per fix.
In real work
The triage list with time estimates is a producer's artifact — it shows you can convert feedback into a plan. That's hireable.
5 · Sound briefs and the music rights wall
Describing sound is a deliverable skill ("dry clicks, room tone, one swell at the title"). AI drafts sound briefs from your treatment. Generated music stays out of deliverables — rights contested, active litigation. Scratch-only, private.
03This Week's Words
TranscriptionSpeech-to-text. The quiet workhorse; verify names and terms.
SRTThe standard caption file: numbered cues, timecodes, text.
Captions vs. subtitlesCaptions = same-language + sound info; subtitles = translation.
Caption QA passThe human round: names, terms, breaks, timing.
Corrections table"Heard → corrected." QA evidence + disclosure in one.
Temp VO / scratchPlaceholder narration for timing. Phone-recorded is craft.
Voice cloningSynthesizing a real person's voice. Documented consent or never.
Multimodal critiqueModel-as-reviewer on screenshots, steered by your criteria.
Critique filteringAccepting/rejecting feedback with reasons. The taste skill.
Edit triageBlocker / should-fix / polish, with minutes. The producer's artifact.
04Demo: Dirty Audio to Professional Captions (plus the Relay)
I made this, so can you
The instructor records 45–60 seconds live — imperfect, with a stumble — and takes it all the way to broadcast-rule captions, then runs the relay on a real WIP.
Prepared demo artifact
Instructor backup / finished example: captions-pipeline.html. Use it if the live tool run fails, or to reveal the finished structure after critique.
- The recording: live, with planted traps — "NCAD," an Irish place name, "risograph," a restart.
- The error hunt: raw transcript on screen; the room finds the failures as sport, tallied by type.
- The cleanup: caption-ready SRT (line lengths, sense breaks, rhythm preserved) + the corrections table — the teaching object.
- The relay: a volunteered WIP through three named reviewers; the volunteer filters aloud with reasons; the room challenges; the triage list assembles itself.
What to notice
The machine did volume; the human did names, rhythm, and judgment. One critique point was confidently wrong and got rejected with a reason — the most professional moment of the demo.
05Step-by-Step Workflows
Caption pipeline (any clip, forever)
- Transcribe with whatever's nearest (editor built-in, Descript-class, phone).
- Correct against the audio: names, terms, homophones. Log every correction.
- Format: ≤42 chars/line, 2 lines max, sense-boundary breaks, rhythm preserved.
- Export SRT → import to your editor → watch with sound OFF.
- Keep the corrections table with the project.
The critique relay (any WIP, any discipline)
- Screenshot the work (yours, or with permission).
- Three named reviewers fitted to the work — pick personas who'd disagree.
- Demand 3 observations + 1 priority fix each, no softening.
- Filter with written reasons — accept/reject every point.
- Peer challenge: one accept + one reject defended.
- Triage: blocker / should-fix / polish + minutes. Your next session, pre-planned.
06Prompt Recipes
The caption cleanupClean this raw transcript into captions. Fix obvious mis-hearings — it's "NCAD", "risograph", [your names/terms]. Break into caption-length lines: max 42 characters per line, max 2 lines per cue, break at sense boundaries, never mid-name. Keep my actual speech rhythm — don't formalize my voice. Output as SRT with placeholder timings. Then give me a corrections table: heard → corrected.
The three-reviewer relayCritique this [frame/layout/sheet/UI] as three different people: [reviewer 1], [reviewer 2], and an accessibility reviewer. Each gives exactly 3 observations and 1 priority fix. Reference what you can actually see. Don't soften, don't compliment.
The edit triageI've accepted these critique points: [paste your accepted list]. Merge them into one prioritized edit checklist: BLOCKER (breaks the piece) / SHOULD-FIX (hurts it) / POLISH (nice). Estimate minutes per fix. Order by impact-per-minute.
The sound briefFrom this treatment/edit description: [paste], draft a sound brief for a sound designer (or future me): overall feel in 2 sentences, moment-by-moment cues tied to shots, what silence is doing, references as QUALITIES (e.g., "close, dry, single-mic intimacy"), and what to avoid. No music generation — this is a brief for humans.
The footage loggerHere is the transcript of my interview footage with rough timestamps: [paste]. Build me a log: topic per segment, the 5 strongest quotable moments (with timestamp + why), repeated themes, and a paper-edit order suggestion for a [60-second] cut. I make the final selects.
07Lab: Screenshot Critique Relay (70 min)
Cold start friendly. Bring any WIP — or use the provided set.
- Record + transcribe (15 min): 45–60 seconds of you pitching your current project. Phone voice memo, imperfect on purpose → transcribe.
- Caption pass (20 min): cleanup recipe → corrections table (the assessed artifact) → SRT.
- The relay (20 min): your WIP → three reviewers → filter every point in writing with reasons → neighbour challenges one accept and one reject.
- Triage (15 min): blocker/should-fix/polish + minutes. Assemble the packet: clip + table + filtered critique + edit list.
Zero-AI track (fully equal)
Hand-transcribe 30 seconds (real subtitlers still do), format to BBC rules, then run the relay on paper with two classmates as named reviewers. The hand-transcription teaches caption rhythm better.
08Integrity Check
This week's four
- Captions: did a human (you) verify every name and term against the audio? Autocaption sludge with your name on it is still your sludge.
- Voice: every voice in your work is yours, consented, or clearly synthetic-and-labelled. No cloning real people. Ever.
- Relay on a classmate's work: permission first, always.
- Music: nothing generated in deliverables — rights contested. Scratch-only, private.
09Troubleshooting
| Problem | Fix |
|---|---|
| Transcription mangles Irish names/places | It will. The corrections pass exists for this — and feed correct spellings in the prompt next time |
| Captions read like a legal document | You let it formalize. "Keep my actual speech rhythm" — restore the "like"s that are voice, not noise |
| SRT timing drifts | Placeholder timings need an editor pass — nudge cues to the actual speech. Five minutes |
| All three reviewers agree | Personas too close. Pick reviewers who'd fight: a minimalist and a maximalist |
| The critique misreads the image | Multimodal models do that. Reject-with-reason — "it claims the type clips the edge; it doesn't" — that's why you filter |
| The edit list is huge and demoralizing | Triage IS the cure: blockers only for the next session; polish lives in someday |
10Resources
- StandardLinkedBBC Subtitle GuidelinesThe professional captioning reference — line lengths, breaks, timing, sound labels. Today's caption rules come from here.
- StandardLinkedMaking Audio and Video Media Accessible — W3CCaptions, transcripts, audio description — what each is for and when it's required, from the standards body.
- ToolLinkedWhisper — OpenAI (open source)The model behind most transcription tools — free and local, for bulk transcription without uploading audio anywhere.
- ToolVerify before teachingDescriptThe edit-by-text workflow leader; features and pricing shift — check the current onboarding demo.
- PolicyVerify before teachingElevenLabs voice documentation + consent policyIf you demo voice synthesis, read the current consent/verification requirements first — they change, and they matter.
- VideoVerify before teachingA current captions-advocacy video (deaf creator perspective)First-person "why captions matter" content lands better than any guideline — pick a current one before class.
Optional extension
- Caption a minute of someone else's published video (privately) to BBC rules; compare against their official captions — who made better break decisions?
- Run the relay weekly on your capstone WIP from here to Week 15 — the trajectory of your accepts/rejects is taste, documented.
- Hard mode: record a 60-second audio description of one of your own silent pieces. Describing your work precisely will change what you make.
Next week: Coding for Artists → — tiny tools, real builds, and the spec-first method.