Remove noise from a recording: wind, hum and hiss without wrecking the take

Three noises, not one

When a take is noisy, the reflex is to turn up a noise reducer until the noise goes away. Sometimes that works and sometimes it leaves the recording worse, because the noise in a take is hardly ever one thing. In vocals, guitar or outdoor recordings made on a handheld recorder or a phone, three show up most: wind, electrical hum and hiss.

All three read as "noise", but they behave in opposite ways. Wind is low and arrives in bursts. Hum is fixed and sits at exact frequencies. Hiss is constant and covers the whole spectrum. Each one is removed with a different technique, and the one that works for one damages the music if you use it on another.

Wind: turbulence at the capsule

The wind that ruins a recording isn't the sound of wind in the trees, which is a sound like any other. It's air hitting the microphone capsule and forming eddies right on top of it. That turbulence moves the diaphragm directly, without travelling through the air as a sound wave, which is why it arrives with so much energy: in a real outdoor take we measured gusts lifting the low end 20 to 30 dB above the music.

Almost all that energy is below 150 Hz, with a tail reaching 500 or 800 Hz, and it comes in bursts. It sounds like rumbling or blowing into the mic, and it also makes any compressor downstream pump.

How to avoid it while recording, from most to least effective:

• A furry windshield ("dead cat") over the recorder. It slows the air before it reaches the capsule and is the only thing that works in real wind.
• The foam cover that comes with the recorder. Enough for a breeze or for walking; in strong wind it falls short.
• Your body as a shield. Stand with your back to the wind and the recorder in front of you, or find a wall, a car or a tree that breaks the airflow.
• The recorder's high-pass filter, at 80 or 100 Hz. It doesn't remove the wind, but it stops the lowest gusts from overloading the input.
• Recording in 32-bit float if your recorder supports it. In the outdoor take we used for calibration, some peaks went above 0 dBFS and were recovered intact because the file was float. It protects the file, not the capsule: if the mic itself or the preamp distorts, that can't be recovered.

How wind is removed: what two mics know

Turbulence is local: the eddies forming over one capsule aren't the same as those forming over the one next to it. Music, on the other hand, is a wave that reaches both microphones at almost the same moment. In an XY pair, like the one on handheld recorders, that difference can be measured. In the real take, the correlation between the two channels was 0.8 in the mids, where the music dominated, and practically 0 in the low end during gusts.

That's where the method comes from. In the low zone the two channels are split into their sum and their difference — the same idea as mid/side processing. The music is in the sum; the wind is spread across both. The difference is discarded, which leaves the low end in mono, and whatever in the sum resembles what was in the difference is subtracted. Above about 800 Hz nothing is touched. Leaving the low end in mono isn't a loss: it's what you'd do in a mix anyway so the bass doesn't depend on phase between channels.

It has a cost, and you should know it. At the recommended strength, the method removed 16 dB of wind below 150 Hz on average, but the instrument also lost 3 to 4 dB of low end. That's why strength matters and why you have to compare: with the cleaned take and the original matched in loudness, switching between them at the same point. If the instrument ends up thin, those decibels can be given back afterwards with a gentle low shelf, now that the wind isn't there to come up with them.

In a mono take, like almost any phone's, there's no second microphone to compare against: all that's left is a high-pass acting during gusts, and it works much worse.

Hum: 50 or 60 Hz and all its harmonics

Hum comes from the mains supply, which oscillates at 50 Hz in Europe and much of South America and at 60 Hz in the United States, Mexico and other parts of Latin America. It gets in through ground loops, poorly shielded cables, power supplies or nearby transformers, and it never arrives as a pure tone: it arrives with its harmonics, the exact multiples of the fundamental. In the real hums we analysed, between 45 and 125 harmonics showed up, all the way to 8 kHz.

That explains why cutting 50 Hz barely helps: the low end goes and a hum keeps ringing in the mids, which is exactly where the ear is most sensitive. It's removed with a very narrow notch on each harmonic, about 3 Hz wide. It's the same filter as an EQ band with an extremely high Q, repeated dozens of times. Widening them seems safer and it's the opposite: with a hundred wide notches the filter eats a noticeable chunk of the music.

The tricky part is precision. The mains doesn't run at exactly 50.00 Hz, and the error multiplies at every harmonic: 0.03 Hz of error on the fundamental becomes 0.6 Hz at the 20th harmonic, enough for a 3 Hz notch to land beside the peak instead of on it. That's why the fundamental has to be measured in the take itself, to the hundredth of a hertz, rather than assumed.

Even so, something remains: in one real hum, about 11% of the energy was noise clinging to the harmonics, which no notch reaches. That needs a second stage of spectral reduction, the same one used for hiss.

A separate case is the buzz of a guitar with single-coil pickups. It also comes from the mains, but almost all its energy is between 2 and 8 kHz, spread over so many closely spaced harmonics that in practice it behaves like noise rather than a clean comb. Treating it like hiss works better.

How to avoid it while recording:

• Plug all the gear into the same power strip. Two devices plugged in at different outlets and joined by an audio cable form a ground loop, the most common cause.
• Use short, balanced cables, and check any that crackle when you move them: broken shielding hums.
• Keep the recording away from transformers, external power supplies, power strips and light dimmers.
• With a single-coil guitar, turn until you find the orientation where it hums least, and move away from the computer and the screen.
• Record ten seconds of "silence" with everything switched on and listen on headphones before the real take.

Hiss: the noise you learn from silence

Hiss is the background breath of a preamp, an interface with high gain, or tape. It's broadband and nearly constant, and that's precisely why it can be learned.

Spectral reduction listens to a stretch with nothing but noise and notes how much there is in each band: the hiss fingerprint. Then it runs through the take and, at every moment and in every band, compares what's there with that fingerprint. Where the noise dominates the music, it lowers it; where the music masks it, it leaves it. That second part is the key: removing hiss while the music plays would take the instrument's brightness with it.

With two seconds of silence to learn from, our tests removed 12 to 18 dB of hiss in the silent parts, about 5 dB in note tails and only 3 or 4 dB while the music was playing. That asymmetry is the right one: hiss is heard in the gaps, not under a strum. With no silence to learn from at all, the reduction managed less than 1 dB.

Overdoing it has a recognisable sound, musical noise: background bubbling or chirps. It's worse than the original hiss, because the ear ignores a constant noise but not one that changes.

To avoid recording it, the remedy is proper gain staging: enough preamp gain that the voice or instrument sits well above the noise, rather than recording too low and raising it later in the mix, which raises the hiss with it. And if the noise comes from the room — a computer, a fridge, the air conditioning — switch it off for the take.

All three at a glance

In what order, and how much

If the take has more than one noise, order matters: hum first, then wind, and hiss last. Hum goes first because it's a fixed, exact signal; if another process touches it first, its peaks get smeared and the notches can't find them. Wind goes before hiss because it comes in bursts: if it's still there when spectral reduction learns its fingerprint from the silences, it learns it as constant noise and applies it to the whole take.

The amount is decided by ear, comparing original and clean at the same loudness and the same point. Listen to the gaps to judge the noise and to the musical passages to judge what you've lost. If the music has changed character, lower the strength: a little residual noise hides in the mix, a thin instrument doesn't.

A cleaned vocal take is then processed like any other: the vocal mixing guide picks up from here. Bear in mind that compression and saturation raise the quietest parts of the signal, and with them any remaining noise: what was inaudible in the raw take can surface after 6 dB of compression.

What can't be fixed without AI

All three techniques work because the noise differs from the music in something measurable: incoherence between mics, exact frequencies or constancy over time. When it doesn't, no filter can separate it.

• Room reverb: it's the music itself, repeated and attenuated. Prevent it by recording closer to the source or with blankets and cushions around.
• Background voices, traffic or birds: they have the same structure as what you want to keep and they change constantly.
• Heavy clipping: the part of the waveform that was cut off isn't in the file. It can be disguised a little, not recovered.

That takes trained models that separate sources, with variable results. If the take can be redone in better conditions, that almost always beats any cleanup.

Try it on your take

Everything in this guide is what Clean Your Take does: it detects wind, 50 or 60 Hz hum and hiss, cleans them in that order and lets you compare the original and the clean version matched in loudness, at three strengths. It runs in your browser: the take never leaves your device. If your recording has no silences, it tells you, and you'll know what to do differently next time.