In the app
Three screens, in the order you meet them.
It is mains hum — 50 or 60 Hz from the power around you, not from your microphone. Because it sits at exact frequencies, it can be subtracted precisely instead of smeared away along with your voice. DeHum does that on your iPhone. Nothing is uploaded.
The same ten seconds, before and after, at the Standard depth. Headphones help — on laptop speakers the hum may be inaudible in both.
Speech with a 60 Hz ground loop underneath it.
DeHum’s actual output on that file. Nothing else was changed.
This is a test fixture, not a recording from someone’s phone: clean speech with a 60 Hz ground loop added at a known level. That is exactly what makes the result measurable rather than merely describable — and exactly why it is labelled. It is built from the repo by make website-audio, straight out of the test suite’s own output, so nobody picked it for sounding good.
Before a line of the app was written, the method was put through a gate on 13 fixtures built from real recorded speech with electrical contamination added at a known level. These are those numbers, including the one that fell short.
Correct detections — including all three decoys, among them a sustained musical note at exactly 60.000 Hz, which was left alone.
Damage to the voice on 12 of the 13 hum fixtures, measured against the known-clean original rather than judged by ear.
Hum removed under speech at Standard depth. The engineering target was 20 dB, and the low end of that range did not reach it.
Measured on synthesised fixtures, which is the only way to know the exact answer to compare against; the thirteenth fixture, a control case, showed 1.8 dB. The same subtraction reaches 83 dB on hum with no speech over it, so what limits it is estimating the hum underneath a voice, not the method. None of this is a promise about a particular recording of yours — which is why the app lets you hear the before and after before you export anything.
Ground loops from a cheap interface, dimmer-switch buzz, fluorescent ballast hum, single-coil pickup noise, generator drift. DeHum finds the fundamental, follows it if it wanders, and removes it and its harmonics.
A notch filter removes everything in a band, voice included. Mains hum genuinely is one sinusoid per harmonic with a slowly varying amplitude — so DeHum estimates that sinusoid and subtracts it, which leaves the rest of what lives at that frequency untouched and introduces no ringing.
Never more than the amount that is always there. Hum is steady; a voice is not. That single rule is what stops it eating a low male voice that happens to share a harmonic with the hum — the risk that decided whether this app was worth building.
If DeHum cannot confirm mains hum in a recording, it says so and changes the file in no way at all. It is not a general noise reducer and will not pretend to be one: an app that always finds something to fix is guessing.
“Found 60 Hz hum, 7 harmonics, −26 dB”, with every harmonic it located marked on the spectrum. The detection is stated, never assumed — and when it is uncertain it offers you the 50/60 Hz choice rather than guessing.
No account, no upload, no server — DeHum contains no networking code at all. That is not a policy, it is the architecture, and a check in the build fails the release if a single networking or analytics symbol appears anywhere in it.
Repair the audio inside a video and DeHum puts it back without re-encoding the picture — the visual stream is copied through, and because the repair changes no sample count, sound and picture stay in sync.
Three steps, and you hear the result before you commit to it.
An audio file from Files or iCloud Drive, a video from Photos, or anything handed over through the iOS share sheet. M4A, WAV, MP3, AAC, CAF and AIFF, plus standard MP4 and MOV.
DeHum analyses the first stable seconds, takes the peak nearest 50 or 60 Hz, and confirms it against its harmonic series before believing it. The spectrum comes back with each harmonic marked and the hum’s strength in dB.
Toggle Original against Repaired — both play the whole time, so switching only moves which one you hear and the comparison never jumps. Then export the audio, or send it back into your video.
Three screens, in the order you meet them.
The free tier is capped by length and by nothing else.
What the free tier exports is the same fidelity as the paid one — it is simply shorter. The audio is never watermarked on any tier, because a watermarked repair would not demonstrate a repair. Prices are shown in your own currency in the app before you buy.
A repair tool for the one defect that does not need a guess — on your iPhone, with nothing uploaded.