1. Why this check comes first
An AI reviewer that looks at frames cannot hear. A person watching once can miss a line. But a missing line is a real defect: in one approved-looking episode, the automatic check found that one character’s line had simply not made it into the cut. The owner had watched that episode and liked it.
2. How it works
- Read every scripted line, in order, with its shot number and speaker.
- Transcribe the episode on the Mac with Whisper (the large-v3-turbo model through MLX). A three-minute episode takes about ten seconds and comes back with a time for every word. Give it the character names as hints.
- Line up the script’s words against the heard words, in order. Each line comes out as OK (90% or more heard), CHECK (70–90%, often a name spelt differently) or MISSING.
- Treat contractions, numbers and sung syllables as equal: “I’m” and “I am”, “19” and “nineteen”, “Ti-na” and “Tina”. Keep a small alias file for how the transcriber spells your names.
- List three or more heard words that match no line as extra speech: an ad-lib, a wrong line, or a voice in the background.
- Note lines that carry on after a cut into the next shot.
3. Two things that made it reliable
The second listen. Short lines (“Four.”, “Every morning.”) get lost in a long transcription. Before calling a line missing, cut out the gap where it should be, between its neighbours, and transcribe only that. Give names as hints, never the expected line, so the transcriber cannot just hear what it was told. On one episode this turned three false “missing” lines into confirmed ones.
Dropping invented words. Speech-to-text invents words over music and silence, and the inventions loop: “Look, Look, Look…”, “Thank you. Thank you. Thank you.” Any phrase of one to six words repeated three or more times in a row, and any stretched sound, is removed before scoring.
4. What it cannot judge
It cannot tell whether a voice sounds like the same character, whether the emotion fits, or whether the lips match. Children’s speech and invented names are also harder for transcribers, so a CHECK means “listen once”, not “wrong”. Voice fingerprints, clarity scores and lip-sync timing are the next checks to add; until then, the owner’s one watch with sound covers them.