Cut-point checker
Drop your audio and paste the caption file you already have. The cut-point checker measures, for every cue, where the transcript claims the word starts and where the audio says it starts. Whisper word starts run 50 to 100 milliseconds late, and a cut placed on one lands inside the word.
Given your audio and the caption file you already have, produce the gap between each word's transcript timestamp and its real acoustic onset, in under ten seconds.
Your audio is decoded and measured inside this browser tab and released when you leave. Nothing uploads, nothing is stored, and there is nothing to delete.
The measurement, and why it is not obvious
Two things make a cut audible, and both are invisible until they are not. The cut-point checker measures the first: a boundary placed on a transcript's word timing lands inside the word, because the timing is late against the acoustic attack. The consonant gets clipped, the ear hears a click rather than a cut, and nothing in the transcript suggests anything is wrong.
The second is the frame grid. An un-snapped fractional cut makes a concatenating encoder pad the joint with digital silence, which produces an audible room-tone dropout and a progressive timeline stretch that desyncs every caption after it. Both survive in shipped products for the same reason: they are inaudible on the first listen and obvious on the tenth.
How to find a cut point by hand
- Zoom into the waveform, not the transcript. At a two-second view every word looks like a block. The attack is only visible at 50 milliseconds per division or tighter.
- Find the quiet before the word. Look for the sustained floor between the previous word's decay and this word's attack. That floor is what an onset is measured against.
- Place the cut in the gap, not on the word. About 40 milliseconds before the attack. Any later and the first consonant goes; any earlier and the previous word's tail comes with it.
- Check it against the frame grid. A cut at 50.113 seconds on a 23.976 fps timeline is not on a frame. Whatever rounds it will round it somewhere, and it may not round it the way you would.
Where the manual way breaks
Finding a cut point by hand works, and takes about thirty seconds per boundary once you know what to look for. That is the problem: a twelve-minute recording with the filler removed has a few hundred boundaries, so a technique that is correct and cheap per cut is an afternoon per recording.
It also does not scale down. The reason to place one boundary carefully is that a single clipped consonant is audible; the reason nobody places three hundred carefully is that nobody has the afternoon. What usually happens instead is that the boundaries are placed on the timestamps, the result sounds slightly wrong everywhere, and the cause is attributed to the material.
Never writes one JSON document — the Frame-Locked EDL — and a deterministic kernel renders it, snapping every cut to a frame boundary and moving each boundary outside the measured onset instead of trusting the transcript. The measurement on this page is that pass's first half, given away; why it changes how an edit sounds is the rest of it.
Read the transcript
Cut-point checker Drop your audio and paste the caption file you already have. The cut-point checker measures, for every cue, where the transcript claims the word starts and where the audio says it starts. Whisper word starts run 50 to 100 milliseconds late, and a cut placed on one lands inside the word.
Frequently asked questions
Why are transcript word timings late?
Because speech recognition aligns to where a word becomes recognisable, not to where its sound begins. A plosive or a fricative starts tens of milliseconds before the vowel that identifies the word, so the timestamp lands after the attack — consistently, in one direction, by 50 to 100 milliseconds.
What does the cut-point checker do with a cue it cannot measure?
It says so rather than guessing. A measurement needs a quiet gap in front of the word to measure the onset against; inside continuous voicing there is no gap, and the honest answer is that a timestamp-placed cut has nowhere good to land there at all. Those cues are marked, not estimated.
Does my audio get uploaded?
No. The cut-point checker decodes your file with the Web Audio API and measures it inside the browser tab. Nothing is uploaded, nothing is stored, and the analysis is released when you leave the page.
Is this the same measurement Never uses?
The same envelope and the same constants: a 5 millisecond RMS envelope at 16 kHz, with quiet defined as 18 dB below the median level of the speech that follows and six consecutive frames required before a gap counts. What this page does not do is move the cut — that part is the product.