One Channel Wired Backwards: What the Transcript Hears
How this happens in real life
An XLR cable gets soldered the wrong way round. A TRS adapter swaps tip and sleeve. A mixer channel has a polarity button pushed in from a previous session and nobody notices. A stereo cable is wired one way at one end and the other way at the other.
The result is a file that plays, sounds thin on headphones, and contains almost nothing. Two copies of the same signal, one inverted, sum to zero. Not quiet. Zero.
We built that file on purpose. One 31.6-second mono recording, panned to both channels, then the right channel multiplied by minus one. 5,575,730 bytes of WAV, same as the correct version. A meter would show you that something is badly wrong, but only if you looked.

What came back
The cancelled file uploaded without complaint. HTTP 200. The response took 21.33 seconds, which is the longest of anything we ran today and about three times the normal time for a clip this length.
The text is 2,333 characters and 434 words. Here is what it is made of.
- 53 sentences in total.
- "Thank you very much" appears 53 times, once per sentence.
- One sentence is the phrase "I'm sorry" repeated back to back with no punctuation between the repeats.
- The actual content of the recording - a walkthrough of a dashboard - does not appear at all.
It is the same failure shape we have seen from long silences and from synthetic speech. Given nothing to decode, the model reaches for the phrases that dominate its training data for empty audio. It does not say "I can't hear anything". It invents a polite ending, over and over.

Half a decibel stopped it
Exact cancellation is a knife edge. We made the inverted channel slightly quieter at five settings and uploaded each one.
- Right channel inverted at full level: 434 words, 2,333 characters, 21.33 seconds.
- Right channel inverted and 0.5 dB down: 109 words, 542 characters, 7.10 seconds.
- 1.0 dB down: 109 words, 542 characters, 13.74 seconds.
- 1.5 dB down: 109 words, 542 characters, 6.85 seconds.
- 3.0 dB down: 109 words, 542 characters, 7.10 seconds.
One half of one decibel. That is the difference between a file the recognizer reads correctly and a file it fills with a stock phrase 53 times. And 0.5 dB is well inside the tolerance of any real-world wiring error, which means the failure is not reliably reproducible from the outside.
That is the part worth remembering. You cannot look at a bad transcript and conclude the polarity is flipped, because a file that is 0.5 dB off from perfect cancellation transcribes perfectly.
The four channel faults that were harmless
To be fair to the format, most channel problems do nothing at all. We tested five layouts on the same recording.
- Identical audio in both channels: 109 words, 543 characters.
- Speech in the left channel, digital silence in the right: 109 words, 543 characters.
- Speech in the right channel, silence in the left: 109 words, 543 characters.
- Speech left, pink noise at -60 dBFS right: 109 words, 543 characters.
- Right channel 12 dB below the left: 109 words, 543 characters.
Every one of them matched the reference. The endpoint downmixes or picks a channel and gets on with it. A missing channel, a dead channel, a quiet channel - none of it costs a word.
So the four things people worry about are free, and the one thing nobody checks is the one that destroys the file. That asymmetry is the entire reason this test was worth running.

How to catch it before you upload
You cannot hear exact cancellation on a phone speaker, and you cannot see it in a file size. Two checks work.
Sum the two channels to mono and look at the level. A file that is correctly wired gets louder or stays the same. A file with one channel inverted gets much quieter, because the sum is the difference between the two. In our case the sum was roughly 40 dB down.
Or load it in any editor that draws both channels separately. If the two waveforms are exact mirror images - every peak on one side matched by a trough on the other - the polarity is wrong. Mirror images are not a stylistic choice; real stereo never looks like that.
What the fix costs
Invert the channel back. That is arithmetic on the samples, not a re-encode, so there is no quality question. Do it before the recording gets long, because the repair is per-file and every file you already sent through has to be done again.
The alternative is to record in mono until you have checked the chain. Mono cannot cancel against itself, which is why the safe habit for interviews and lectures is one channel and one microphone.
Where your file goes
All six uploads went to /api/transcribe over TLS, were processed once in memory, and were discarded when the response came back. The cancelled file was processed like any other: the endpoint had no way to know it was empty, and that is exactly why it returned 2,333 characters of nothing.
Got a Reel you want as text? Get an Instagram transcript as TXT, SRT, or VTT →
A cable wired backwards is a two-minute fix, but only before you record the next hour. Wondershare UniConverter opens the file on your own machine and lets you invert or swap a channel without re-encoding the speech, so the repair is a copy of the waveform rather than a new generation of it. It does not transcribe - the words come back from this page.
FAQ
What happens if one channel of my recording is wired backwards?
The two channels cancel and the file contains near-silence. We uploaded exactly that and got HTTP 200 with 434 words and 2,333 characters in 21.3 seconds. The content was 53 sentences of "Thank you very much" and a run of "I'm sorry", repeated. None of the actual speech survived.
How much level difference does it take to break the cancellation?
Half a decibel. With the right channel inverted at full level we got 434 words of nonsense. At 0.5 dB below the left channel, the same recording returned 109 correct words and 542 characters. That is inside the tolerance of most real wiring errors.
Does a missing or silent channel hurt the transcript?
No. Speech in the left channel with digital silence in the right returned 109 words, identical to a properly wired stereo file. So did speech in the right with silence in the left, a -60 dBFS noise floor on the empty side, and a 12 dB level difference between channels.
How do I check a file for inverted polarity?
Sum both channels to mono and watch the level. A correct file gets louder or holds; an inverted one drops sharply, about 40 dB in our test. Or look at the waveform: exact mirror images between the two channels means the polarity is wrong.