NCAD AI for Artists Workshop · Week 10

Audio, Voice, Video

and Multimodal Critique

All-tracks week disguised as a video week. Say so early.

60-second recap

The deal

AI for friction, humans for meaning.
Raw output is never finished work.
Verify before trusting. Disclose where it matters.
Skepticism welcome when specific.

Standing recap.

Today

Today's promise

Caption a real clip to broadcast rules — and own a critique workflow that gives you useful, filtered feedback on anything you make, at 2am, forever.

The relay is the most universally applicable workflow in the course.

The week in one line

Machine speed + human verification —
on words, on captions, on critique itself.

Board: MACHINE TRANSCRIBES → HUMAN CORRECTS → TABLE KEPT.

Deliverable

You will leave with

A captioned clip (or critique packet on a still)
A corrections table (heard → corrected)
A prioritized edit list (with minutes)

Show a finished packet briefly.

Today's map

Today's map

  1. Transcription, captions, voice consent — 40 min
  2. The critique relay + sound briefs — 15 min
  3. Demo: dirty audio → captions → relay — 35 min
  4. Lab: your clip, your relay — 70 min
  5. Share: error trophies + rejections — 40 min
Breaks protected.

Part 1

Transcription: the quiet workhorse

~0:10.

What it unlocks

The most reliable AI tool in the building

Captions · interview notes · edit-by-text · searchable footage · meeting minutes

Whisper-class transcription is inside everything now.

Where it fails

Reliably, and exactly where audiences notice

Proper nouns — "NCAD" → "encad"
Irish names and places — mangled
Craft terms — "risograph" → "rise of graph"
Homophones — chosen wrong

"The failures cluster where your credibility lives."

The pattern

Machine transcribes → human corrects → table kept

The corrections table (heard → corrected) is
QA evidence and disclosure in one document.

The table is this week's signature artifact.

Part 1, continued

Captions are craft

~0:20.

Not garnish

Three arguments for captions

Access — a deliverable with legal weight in the EU
Reach — most social video plays silent
Search — captions make work quotable

Different students respond to different ones. Give all three.

The form rules

BBC guideline essentials

BBC Subtitle Guidelines linked on week page — the industry reference.
Try this now

60 seconds

Fix this caption

✗ "and then we printed it at the Black Church
Print Studio which was"

✓ ?

The room fixes it aloud: break at the phrase, keep the name whole. Fast win.

The difference

Sludge vs. professional

Autocaptions = the sludge baseline.

One human pass: names · terms · break points · watch once with sound OFF

"Sludge with your name on it is your sludge."

Part 1, the line

Voice: the brightest line

~0:32. Slow down.
Legitimate

Temp VO

Scratch is craft

Scratch narration for animatics ✓ · accessibility reads of your own texts ✓ · placeholders you'll re-record ✓

Your phone is often better anyway — your breath knows your edit.

Phone-recorded scratch is craft, not compromise.
Never

The line

Cloning a real person's voice
without documented consent.

Not classmates. Not celebrities. Not "just for the rough."

"The brightest line in this course, and it's hardening into law. 'It was temporary' does not work."
The nuance

Your own voice?

Yours: yes

Your voice, your data, your consent.
Label it synthetic where it could mislead.

Never lend your cloned voice to claims you didn't make.

The nuance students actually ask about.

Break

Break — 10 minutes

Stage the relay prompt + your backup WIP still.

Part 2

The multimodal critique relay

Works on everything you make

~1:00. The generalizable half of the week.

The capability

Models can see now

Screenshot anything — frame, layout, model sheet, UI — and run structured critique against your criteria.

Filtered by you: a second pair of eyes. Unfiltered: noise.

The filtering is the skill.

Callback to Week 2's critique loop — this is its visual sibling.

The method

The relay, five steps

  1. Screenshot the WIP
  2. Three named reviewers — 3 observations + 1 priority fix each
  3. You filter — accept/reject with written reasons
  4. Peer challenges one accept + one reject
  5. Triage: blocker / should-fix / polish + minutes
Step 3 carries the week: "At least one critique will be confidently wrong. Rejecting it well is the most professional move you'll make today."

The artifact

The triage list is the deliverable

Blocker / should-fix / polish — with minutes per fix.

It shows you can convert feedback into a plan.
That's a producer's skill, and it's hireable.

"Your next work session, pre-planned."

And sound

Sound briefs + the music wall

Describing sound is a deliverable: "dry clicks, room tone, one swell."

Generated music stays out of deliverables — rights contested, active litigation. Scratch, private, done.

Tools doc has the current state.

Demo

Dirty audio → captions → relay

~1:15. Phone out. The most failure-proof demo of the term.
Demo · move 1

Record, live, badly

45–60 seconds. A stumble. A restart.

"NCAD." A place name. "Risograph."

The room watching you record imperfectly is disarming and perfect. Planted traps included.
Demo · move 2

The error hunt

Find the errors. It's a sport.

Tally by type: name / term / homophone.

Board tally. The errors are guaranteed — that's why this demo can't fail.
Demo · move 3

Cleanup + the table

Caption-ready SRT + the corrections table

Line lengths, sense breaks, rhythm preserved.

Heard → corrected. QA evidence + disclosure. One document.

Name it the teaching object. Read 3 rows aloud.
Demo · move 4

The relay, live

Three reviewers → volunteer filters → room challenges → triage

One reviewer-point will likely misread the image — celebrate the rejection: "It said the type clips the edge. It doesn't. Rejected, with a reason."

Break

Break — 10 minutes

Lab next: your clip, your relay.

Lab slide staged.

Lab · 70 minutes

Screenshot Critique Relay

Record + transcribe (15) → Caption pass (20) → Relay (20) → Triage (15)

Bring any WIP — or use the provided set.

Parts + times on board.
Lab · steps 1–2

35 min

Record, transcribe, caption

45–60s of YOU pitching your project — imperfect on purpose → transcribe → cleanup recipe → corrections table → SRT.

The corrections table is the assessed artifact.

Corridor recording fine. Voice-shy: partner reads the pitch. Circulate: "Show me your heard→corrected pairs."
Lab · step 3

20 min

The relay

Your WIP → three named reviewers → filter every point in writing, with reasons → neighbour challenges one accept + one reject.

Swap personas per discipline: type designer / illustrator / UX researcher / colourist.
Lab · step 4

15 min

Triage + packet

Blocker / should-fix / polish + minutes. Order by impact-per-minute.

Packet = clip + table + filtered critique + edit list.

"Would you start tomorrow's session with this list? Then it worked."
Always available

Alternative

Zero-AI track (fully equal)

Hand-transcribe 30 seconds → format to BBC rules → paper relay with two classmates as named reviewers.

The hand-transcription teaches caption rhythm better.

Say it plainly; it's true.

Share-out · 30 min

Trophies and rejections

Applaud the rejections specifically — that's the skill being trained.
Every week

Integrity check

This week's four

  1. Every name and term verified against the audio?
  2. Every voice: yours, consented, or labelled synthetic? No cloning real people
  3. Relay on a classmate's work: permission first
  4. Music: nothing generated in deliverables
Four beats.

Reflection

Write 4 lines

Two quiet minutes.

Next Wednesday

Coding for Artists

Tiny tools, real builds — the spec-first method. A working app gets built live, from a spec the room writes.

No coding experience expected. Genuinely none. Same deal as Blender week.

End on time; corridor questions after.

Tonight, optional

Go deeper

Final slide stays up while people pack.