In the app
Three of the screens, in roughly the order you meet them.
Two microphones, one conversation — and every word is on both tracks. Mic Bleed Remover subtracts each voice from the other’s recording, using the opposite microphone as the reference rather than guessing from one signal. Each track ends up gateable, levellable and editable on its own. On your iPhone, with nothing uploaded.
Removing the other person is the easy half. The half that decides whether an app like this is worth using is what happens to the voice you are keeping — so that is the clip published here.
Six seconds of one speaker on their own microphone, untouched.
The same six seconds, through the app’s processing. It should sound the same.
Not a recording from anyone’s phone: this is a synthetic test fixture, clean speech with a measured amount of the other voice mixed into it, which is what makes an exact answer available to compare against. It is built from the repository by make website-audio, out of the blind listening test’s own render — specifically the worst single-talk window in the session, the one the harness ranks as most likely to reveal damage. Both clips carry the same gain. Nobody picked this clip for sounding good.
Before a line of the app was written, the method went through a gate on six synthetic two-microphone sessions built so the right answer is known exactly. These are those numbers — including the half of the bar that is still unanswered.
How far the other voice came down, across all six fixtures and both directions, measured on the windows where only that voice was present.
How close the hardest fixture — a room with a long tail — came to its own theoretical ceiling. What limits it there is the filter length, not the method.
The gate’s second half asks for no audible hollowing of the voice you keep. That has not been measured yet, so the gate returns “could not run” rather than a pass.
Every fixture here is synthetic, and that is a real limit as well as the source of the exactness: a wireless transmitter runs automatic gain and a lossy radio codec, which makes the bleed path change with level, and no fixed test corpus can show that. None of these numbers is a promise about a recording of yours — which is exactly why the app measures your own session and puts that number on the screen before you spend anything.
Two people at one table on a dual-channel wireless kit, two phones on the desk, or a phone and a lav. Each mic hears both voices; this takes each voice out of the track it does not belong in.
Every other cleanup tool estimates what to remove from the one signal it has. This one already holds the interfering signal — the other microphone recorded it — and subtracts what that path predicts. Using the real reference instead of a guess is the only way to take a voice out without hollowing the voice you keep.
They never start together and never keep the same time. The offset comes back in plain words — “Ben’s mic starts 12.4 s after Ana’s” — and the drift between two independent clocks is tracked and corrected across the whole session, not just fixed once at the top.
A dual-channel wireless kit in split mode puts one speaker on each channel of a single recording. The app recognises that and never asks you for a second file. Two separate files work too, and they can arrive minutes apart — the second is paired when it lands.
It finds the worst moment of bleed in your session by itself, plays it solo’d, and tells you how far down the other voice came — measured where only that person was talking. Press and hold to flip to the original at the same instant. All of that is free, at full quality.
No account, no upload, no per-minute credits; it works in airplane mode. There is no model and no server in the story at all — it is deterministic signal processing on the phone, and a check in the build fails if a networking or analytics symbol appears anywhere in it.
Two cleaned tracks, named after your speakers, at their original length and start offset, so both drop straight back into place in your editor. Audio outside the overlap passes through bit-identical and is labelled as untouched. M4A, or WAV 24-bit on Premium.
Three steps, and the result is yours to judge before anything is asked of you.
Two files from Files, the share sheet, Photos, AirDrop or an SD card reader — or one dual-channel file, which is recognised as two microphones on its own. Audio or video; two phones running Voice Memos works too.
The offset between the two recordings is found and stated in plain words, clock drift is corrected continuously across the session, and you get a map of where the two actually overlap.
The app finds the worst bleed in the session and plays it to you with the measured figure, press-and-hold A/B, and any other moment you care to scrub to. Then export both cleaned tracks.
Three of the screens, in roughly the order you meet them.
The whole result is free. What Premium buys is the file.
There is no length cap on the free tier, because a length cap refuses the actual job: you would hear thirty seconds of a two-hour interview and learn nothing about whether this works on it. You hear the whole thing, at full quality, with the measured number from your own session — and you pay only when you want the files. Prices are shown in your own currency in the app before you buy.
The one repair that does not have to be a guess — on your iPhone, with nothing uploaded.