Screen Recorder Audio: When the Two Tracks Arrive Half a Second Apart
The mismatch is smaller than the worry about it
Every screen recorder that grabs system audio and a microphone records two clocks. They start at slightly different moments, they drift by a few milliseconds a minute, and the two signals reach the mix at different levels. People spend real time nudging one track against the other before uploading.
We tested whether that time is worth spending. One 50.9-second recording, 138 words, 738 characters. The clean capture went in as the reference. A second copy was degraded the way a room degrades sound: low-passed at 3.8 kHz, then given three early reflections at 23, 41 and 67 milliseconds. That copy was delayed and mixed back in at a set level.
Five delays at minus 6 dB, then two level extremes at 120 milliseconds.
- 0 ms, minus 6 dB: 200, 138 words, 738 characters. Text identical to the clean file, character for character.
- 40 ms: 200, 137 words, 737 characters.
- 120 ms: 200, 140 words, 742 characters.
- 250 ms: 200, 139 words, 741 characters.
- 500 ms: 200, 140 words, 742 characters.
- 120 ms at minus 12 dB: 200, 138 words, 738 characters, no differences at all.
- 120 ms at 0 dB: 200, 142 words, 746 characters.
Half a second of drift moved the word count by two. Twelve decibels of level moved it by four, and changed how the model wrote a web address.
Delay is not what breaks a transcript
The reason is structural. The second copy carries the same words as the first. Whatever the offset, the recognizer hears each word twice, close enough together that the doubled signal reads as one slightly smeared word rather than two separate ones.
The diffs bear that out. At 40 milliseconds the only change was `it is` collapsing to `it's`. At 120 and 500 milliseconds the change was the web address: `example.com` became `example dot com`. At 250 milliseconds it was both at once. No words appeared, none vanished, no sentence broke.
Compare that with what happens when the two copies are genuinely different - two people talking at once, or a music bed under speech. There the second source carries different content and the model has to choose. A delayed copy of the same voice is not a conflict. It is one source with a soft echo.

Level is the variable that shows up in the text
Drop the room copy 12 dB and the transcript matches the clean file exactly, 138 words, 738 characters, zero differences. Raise it to full level and the word count goes to 142. Not because more speech appeared, but because the model split `example.com` into three tokens and changed a contraction.
That is the practical finding. If your screen recording sounds doubled, the fix that changes the transcript is a fader, not a nudge. Pull the microphone or the system track down until the echo stops competing, and the text converges on the clean version.
The uncomfortable part is that a level problem produces text that reads perfectly well. `example dot com` is a sentence-shaped mistake in a sentence that otherwise scans. Nothing in the response flags it, and the word count going up makes the file look richer, not worse.

What this means for the file you record next
Record both sources if you can. A clean system capture is the best audio a screen recording gives you. An earlier test scored it 29 of 29 words correct, against 26 of 29 for a room microphone. Just do not spend the session aligning them to the millisecond.
- Set the microphone 10 to 12 dB below the system track before you start, not after.
- If the two tracks are separate files, mix them at the levels you want and upload the mix. The endpoint reads the first audio stream only, so a second track is dead weight.
- Do not re-time a track to fix a word. Our 500-millisecond file was as accurate as the 40-millisecond one.
Why an echo barely touches the text
The reason a delayed copy is harmless comes down to what the model is doing. It maps sound to words, and a word repeated 40 milliseconds later still lands inside the same syllable. The second arrival is not new information. It is the same information with a phase shift.
At 250 milliseconds the gap is wide enough to read as a separate event, and that file did lose words - 128 against 138, the largest drop in the sweep. Ten words gone, still HTTP 200, still a readable paragraph. Nobody would notice from the output alone.
The size numbers track the offset. The clean mix is 977,713 bytes. At 40 milliseconds it is 976,817, at 120 it is 984,636, at 250 it is 1,003,198, and at 500 it is 1,026,453. The mix gets longer because the delayed copy runs past the end of the original, and the added tail is silence padded with echo.
That tail is worth trimming before you upload. A file whose last two seconds are a decaying repeat of the final sentence gives the model one more thing to interpret. It has no way to know you meant to stop.
When the offset does become a real problem
One case does bite: when the two sources carry different speech. A delayed copy of the same voice is an echo. A second voice saying something else is a conflict, and the model resolves it by picking one and dropping the other with no warning.
That is the same failure we measured on a stereo file with a different sentence in each channel. The transcript came back 247 characters and contained neither of the right-channel sentences. The endpoint read the first channel and threw the second away.
Screen recordings hit that shape more often than people expect. System audio carries the call, the microphone carries the room, and if someone else in the room is talking, their words land on one track only. The offset between the tracks is irrelevant. What decides whether their sentence survives is which track happened to be first.
The check is cheap. Play the recording back on headphones and listen for a second voice that sits slightly to one side. If you hear one, the transcript has already made a choice you did not make.
The other case worth naming is a mixer that sums both tracks before writing the file. That is the normal path, and it produces the echo we tested - one voice, doubled, at whatever ratio the sliders happened to be at. Nobody chose minus 6 dB. It arrived that way.

Where the mix goes
Each of the seven mixes above was posted to /api/transcribe over TLS, processed once in memory, and discarded when the response came back. Nothing is stored, nothing trains on it, no account is involved. The room copy in the mix was synthesised by filtering and delaying the same recording, not captured in a real room. That is why the numbers move in small steps instead of falling off a cliff.
That caveat matters for how far you should trust this. A synthesised room is a filter and three echoes. A real room has a dozen surfaces, a noise floor, and a person who moves. The direction of the finding - level matters more than timing - should hold, because it follows from the two tracks carrying the same content. The exact millisecond at which it stops holding is not something this test can tell you.
What it can tell you is that nudging track A against track B did not earn its keep on any file we built today. Pulling one fader down did.
Got a Reel you want as text? Download the Instagram transcript as TXT, SRT, or VTT →
Two tracks that do not line up are a mixing problem, and mixing is a file operation you can do before anything else touches the recording. Wondershare UniConverter rewrites on your own machine at the levels and sample rate you set, without resampling the speech. It does not transcribe - the words come back from this page.
FAQ
Does a small sync offset between two audio tracks ruin the transcript?
No. Delaying the second copy by 40, 120, 250 and 500 milliseconds gave 137, 140, 139 and 140 words against 138 for the clean file. Every upload returned HTTP 200 and none of them lost a sentence.
What actually changes the words in a doubled recording?
Level. At minus 12 dB the mixed file matched the clean one exactly, 138 words and 738 characters. At 0 dB it came back 142 words and 746 characters, with `example.com` written as `example dot com`.
Should I align the system audio and microphone tracks before uploading?
Set the levels and skip the alignment. Twelve decibels of level moved the text; half a second of offset moved it by two words. The recognizer treats a delayed copy of the same voice as one source with an echo.
Does the endpoint read a second audio track if the file has one?
No, it reads the first. We built files with two audio streams carrying different sentences and only the first was ever transcribed, in 12 and 10 words across two orders. A silent first track does not fall through to the second either.