Film voiceover
Voice comes AFTER the picture is approved, and is fitted to it. The picture never moves to fit the voice unless the
owner asks. Read references/writing.md (every rejected pattern, with before/after) before writing a word, and use
the copywriting and humanizer skills on every draft.
Gate 1 · Directions and script (no recording yet)
- Understand every shot. For each beat in
storyboard.json, write in one line what the shot is FOR. If you can't, ask the owner. Lines written about what a shot seemed to show (a pet on a wall of tiny items) got rejected. - Directions. When the angle is open, write 2–4 full directions that differ in what they give the viewer, e.g.: Tips the viewer can use · The daily annoyance · How it works · Doubts answered (key safety, cost, platforms). For each: the full transcript, what it's good at, its risk. Put your pick first.
- Put them on the storyboard with REAL frames: render a still per beat (
node scripts/render.mjs --at …), setvoper direction instoryboard.json, buildreview/storyboard.html(film-storyboard's script). The page opens with each direction as one paragraph: the eyes-closed read. - Self-check with the five tests in
references/writing.md(eyes-closed, frame, stranger, slop scan, claims). - Send the storyboard; ask which direction and what to change. Ask the voiceover questions from
product-film/references/question-bank.md(voice, who it talks to, tone, how much voice). Iterate until a clear yes. Write the approved script toVO-SCRIPT.mdwith the time range per line and "where each claim comes from". Commit.
Gate 2 · Pronunciation (before the full record)
- Ask how the product name (and any other unusual word) is said, in the owner's words ("WIZ-get, hard g").
- Test spellings, one short clip each (a short sentence with the name): the plain name, hyphenated, doubled consonant (e.g. "Nimbet", "Nim-bet", "Nimbbet"). Several spellings in one take come out shorter and confused. Name each file by its spelling. Send them; the owner picks.
- Record the winner in
vo.json › voice.note, inVO-SCRIPT.md, and in memory for future films. Never send the correct spelling to the engine if it mispronounces it: send the respelling everywhere.
Gate 3 · Record, place, mix
- Credits first: check the TTS balance. Fish Audio: max ~500 bytes per call, billed per byte. Record per
section (
audio/vo/l1.wav,l2.wav…), not the whole script in one call. - Voice: the one the owner chose from samples (Fish "Ethan" was approved for a calm product film). Fit pace with
per-phrase
tempo≤ 1.1–1.15; never speed up a whole take. - Cut into phrases:
python3 scripts/vo.py --scanlists speech segments per take; writevo.jsonphrases withtake,from,to,text. - Place each phrase's
at(beats = video seconds × 2) so its key word lands with the matching picture or on-screen words. Get video times withnode tools/probe.mjs cutsand"window.UNWARP(<story t>)". Runpython3 scripts/vo.pyuntil it reportsoverlaps: none; leave ≥ 0.25 s between phrases. - Mix and mux:
python3 scripts/sfx_samples.py && python3 scripts/mix.py(-14 LUFS, ≤ -1 dBTP; voice 0, sfx about -10, music per the owner's choice) thennode scripts/render.mjs --mux [--fmt 9x16]. - Send the voiced cut with the transcript and a table of where each line lands. Say you can't hear it; ask about
pronunciation, pace, level and any line that lands on the wrong frame. Iterate. Commit
vo.jsonandVO-SCRIPT.md.
When the picture changes later
Re-probe, re-place phrases (at), re-mix, re-mux. For 9:16, the same audio/mix.wav works if the timing is shared.