Course Video, step 1: deck to narrated lesson

Course / training video from slides

Read video-studio first. Proven on a 22-min marketing course built from a PDF deck (the brief was "normal editing, not heavy motion graphics", Indian-English male voice). Rebuild it with the structure below: edit/course.py, render.py, mix.py.

Pipeline

1. Slides → images: pdftoppm -r 200 -png raw/deck.pdf work/slides/p (or export PPTX/Keynote to PDF first). Look at every page. For build-ups, find the boxes of each part (title, bullet 1..n, chart, image) — by eye from the image or with PIL row/column projections of non-background pixels — and store them per slide.
2. Script (edit/course.py): per slide: vo paragraphs + reveals [{box:[x,y,w,h], cue:"spoken word"}] + optional zoom box. Write to teach, not to read the slide: hook the lesson (why it matters), one idea per paragraph, examples, recap at the end of each module, ~140–150 words/min. Hinglish if the audience wants it. Show the script and get approval before any VO (it costs credits).
3. Budget check: eleven.py credits; a 22-min course ≈ 18k characters — Starter has 40k/month, so a full re-voice the same month may not fit. Tell the user the cost. Free scratch VO with macOS say only to test timing.
4. VO: per paragraph eleven.py tts "<text>" work/vo/<slide>_<n>.mp3 --voice 05QRrdNgiglhuYOaTQvw --speed 1.05 --words work/vo/<slide>_<n>.json (Prem, approved). Cached by text — editing one paragraph re-pays only that one. Fix pronunciations by respelling in the text. Concatenate with 0.35 s between paragraphs and 0.8 s between slides; build the timeline JSON (slide start/end, cue word times).
5. Render (edit/render.py): each slide starts clean (title only), and each part fades/slides in on its cue word; slow easing camera push (1.00 → 1.04) and a zoom into the part being discussed for dense slides; push/slide transitions between slides (0.4 s), chapter label top-left, thin progress bar, slide number. Crisp: render slides from the 200-dpi PNGs at 1920x1080, no blur. Low-priority, in chunks per module so a killed render resumes.
6. Mix (edit/mix.py): voice chain (highpass 80, light compression), music bed −20 dB with sidechain ducking (assets/music/*underscore*), soft SFX on reveals (pop/swish −14 dB) and transitions (whoosh −12 dB) — lighter than reels. Master −14 LUFS, --tp -2.0.
7. Preview → approve → full: first module at --scale 0.5, get feedback, then everything. qc.py --platform youtube.
8. Package: YouTube chapters (one per slide group/module), description with lesson outline, per-module files if it's a course platform (exports/vN_module_<n>_<slug>.mp4), thumbnail (thumbnail-designer), optional SRT (captions_ass.py --style clean or from the VO words).

Rules

• Visuals follow speech: nothing appears before it's mentioned; nothing important appears without being mentioned.
• Keep the deck's own design (it's the client's brand); add only the camera, reveals, chapter label and progress bar.
• Numbers and claims only from the deck/user. Spell brand names right on screen even if respelled for the voice.
• If the user goes to sleep and waives approvals, still keep previews and notes so they can review in the morning.