What does proof-listening find?
Places where the recording and the manuscript differ: a word read differently, a word or line not heard, and a word added. Each comes with its time in the recording and a button to hear it.
Proofing a recorded chapter
Proof-listen a recorded chapter against the manuscript
Give this page a recorded chapter and the text it was read from. A speech recogniser transcribes the recording, the transcript is lined up with the manuscript word by word, and you get every place where they differ — a skipped line, an added word, a misread — with its time in the recording and a button to hear it. Sign in to start; the model then runs in your browser and your audio and text never leave this tab.
- Nothing is uploaded
- Every difference with its timecode
- A place to listen, not a verdict
Compare a recording with its text
Or paste the chapter text instead
Sign in with Google to use this tool.
The account only starts the comparison: we count how many you run, never what was recorded or written. The speech recogniser is downloaded from this site and runs on your device.
Ready.
Read this before trusting the list
A recogniser’s mistake looks exactly like a misread
Every difference is a place to listen
The transcript comes from Whisper, a speech-recognition model, and it mishears: an unusual name, a quiet word, a sentence over music. When it does, the difference it produces looks the same as a narrator’s slip. So nothing here says the narrator was wrong. Each line is a timecode and a Hear it button; your ears decide.
Which ones are probably the recogniser
When the written word and the heard word sound alike — their and there, a name and its nearest dictionary word — the line says probably the recogniser. That is a sound-alike test, not a confidence score: the model this page runs does not report one per word.
What is folded away before comparing
Capitals, punctuation and apostrophes are ignored, and numbers are compared as words, so a manuscript that says seven and a transcript that writes 7 agree. Timecodes are placed inside the phrase the recogniser returned, so treat them as within a second or two.
What your browser downloads, once
Whisper tiny.en is about 44 MB and its engine about 14 MB, both from this site, the first time you press Compare. English only. A chapter of twenty minutes takes a few minutes on a recent laptop, and the page shows how far it has got.
The rest of the check
Words are one half; the audio is the other
Measure the chapter against ACX
The chapter checker measures level, peaks, noise and format and fixes what it can.
Before the next session
A pronunciation guide prevents the misreads this page finds.
Straight answers
Questions about proof-listening
Can I trust every line of the list?
No, and the page says so. The transcript comes from a speech-recognition model, and a word it mishears looks exactly like a narrator's misread. Every line is a place to listen, not a verdict. Differences where the two words sound alike are marked as probably the recogniser.
Is my recording uploaded?
No. The recording and the manuscript are read in this browser tab, and the speech recogniser is downloaded from this site and run on your own device.
Why do I have to sign in?
Because the speech model is served from this site, and serving it costs us bandwidth. Signing in with Google starts the tool: we count how many sessions you run (10 a day and 60 a month on the Starter plan), never what was recorded or written. The model then runs on your own device.
Which recordings and languages does it take?
WAV, AIFF, MP3 and M4A up to an hour, in English: the recogniser, Whisper tiny.en, reads English only.
Use the book workspace to review decisions alongside your chapters, save the project and prepare the delivery package.