Missed and mis-spoken, marked on the words
After a take, Medrak compares what you said against what the script says. A word you skipped gets a dotted amber mark; a word you said differently gets a wavy coral one, with the expected and heard versions side by side. The marks appear on the line in the teleprompter, in the prep read-through, and on the take's waveform, so the fix is always one punch-back away.
It is sensible about the things narrators actually say: reading "31" for "thirty-one" is not an error, and a word you accept once with the Ignore button is never flagged again, anywhere in the project.
Entirely on your machine
Proofing runs a Whisper speech model locally, on your PC. The model downloads once (about 1.5 GB, and it survives app updates), then everything works offline. No audio ever leaves the booth, which is the whole point for NDA and pre-release work. On capable machines the encoder rides the GPU; otherwise it runs happily on CPU.
Line by line or the whole chapter
In auto mode each take is proofed as you finish it, quietly, without slowing the session. Prefer to record hot and check later? Switch to manual and press Proof to sweep every recorded line in the chapter or script at once. In a collaborative book, proofing is per-narrator: you see flags on your own takes only.
Where to get it
AI proofing is part of the free desktop app on the Narrator and Studio plans, and every 14-day trial includes it. It is the one feature that stays desktop-only: running the model beside the recorder is what keeps it private, fast, and free to run.