Audio Noise Reduction

Every recording made outside a treated room carries something you did not want: the fan in the ceiling, a fridge two rooms away, traffic through a window, the faint hiss of the preamp itself. None of it is loud enough to notice while you are recording and all of it is obvious the moment someone else listens on headphones. Removing it is the single largest improvement available to most recordings, and it costs nothing but a little care with the strength control.

• Podcasts and interviews: clear the room tone under dialogue so an episode sounds like it was recorded in a studio rather than a spare bedroom.
• Video voiceovers: match narration recorded on different days, where the background changed and the voice did not.
• Lectures and meetings: rescue a recording made on a laptop microphone across a table, where the fan is closer than the speaker.
• Digitised records and tapes: take the surface hiss and crackle off a transfer without dulling the music underneath.
• Field recordings: reduce wind rumble and handling noise that were unavoidable at the time.

Noise reduction is not one technique. The right choice depends entirely on what is in the recording, and picking the wrong one is the usual reason people conclude that noise reduction 'ruins' audio.

• AI Voice — a neural network trained to recognise human speech, which keeps the voice and discards everything else. Because it understands what a voice is rather than how loud the background is, it removes noise that changes over time — a door, a keyboard, passing traffic — that no threshold-based filter can touch. It is the strongest option available for speech by a wide margin. The catch is the same thing that makes it strong: anything that is not speech is noise to it, so on music it will remove most of the recording.
• Spectral — splits the audio into frequency bands and attenuates each one that sits below a noise floor you set. It has no opinion about content, so it is safe on music, ambience, sound effects and anything else. It is the right default when you are not sure, and the only mode that can be tuned without listening for a silent passage first.
• Noise profile — the same spectral engine, but the noise floor is measured from your recording instead of guessed. Select a stretch that is nothing but noise and the exact shape of that noise is subtracted across the whole file. Where a plain threshold removes everything quiet, a profile removes the specific thing that was in the way, which means less of the signal goes with it.

The practical rule: speech gets AI Voice, music gets Spectral, and anything with a second of usable silence in it gets the profile. If a mode removes too much, the tool will tell you — it re-measures the result and refuses one where the recording itself has collapsed along with the noise.

This is the technique studio engineers have used since long before software could do it automatically, and it remains the most accurate one available. Rather than deciding that everything below a certain level is noise, the tool looks at a passage you know is only noise and measures its exact spectrum — how much energy sits at 60 Hz, at 400 Hz, at 8 kHz and everywhere between. That shape is then removed from the entire recording.

• Switch the mode to Noise profile. The waveform above becomes selectable.
• Find a stretch with no voice or music in it — the moment before someone starts talking is usually perfect — and drag across it. A third of a second is enough; two seconds is better.
• The selection is shown in seconds, so you can check it lines up with the gap rather than clipping the first word.
• Run it. Because the profile describes your actual room, you will usually need less strength than a plain threshold needs to achieve the same result — which is exactly why less of the voice goes with it.

Almost every noise reduction tool hands back a file and a reassuring label. This one measures the finished audio with the same meter it used on the original, and shows you both numbers, so you can tell whether the settings you chose actually did anything.

• Floor before and after: the level of the quietest sustained content in the file, read from the bytes you are about to download rather than predicted from the settings.
• How much came off: the difference between the two, in dB. Ten dB means the background is about a third as loud as it was.
• What it cost: how much of the recording's own energy changed. A good result removes a lot of noise and almost none of the signal; a bad one removes both, and this is the number that shows it.
• A refusal when it went wrong: if the settings removed the recording along with the noise, the tool says so and names the control to change instead of handing back an empty file that looks correct in every other respect.

Some of what spoils a recording is not really noise at all, and treating it as noise removes the wrong thing. These run before the denoiser, because each of them changes what the denoiser would otherwise measure.

• Mains hum: the steady buzz from a ground loop or a nearby transformer, at 50 Hz in most of the world and 60 Hz across North America. It is a harmonic series, not a single tone, so the tool notches the fundamental and as many harmonics as you ask for — and it listens to your recording first, so it can tell you which one is actually there.
• Rumble: sub-bass energy from handling, footsteps, wind and air conditioning. It is usually inaudible on laptop speakers and enormous on headphones, and cutting it before the denoiser stops it dominating the noise estimate.
• Clicks and crackle: the impulses a record or a damaged tape adds. A click is broadband, so leaving it in poisons every measurement downstream of it — which is why de-click runs first in the chain.
• Clipped peaks: the flat tops left when a recording was tracked too hot. The shape is reconstructed from the waveform either side, and it runs before everything else because a clipped sample carries energy that was never in the room.

The failure mode of noise reduction is not leaving noise behind — it is a voice that sounds like it is underwater. Every tip here is about the same trade.

• Start low and stop early. Raise the strength until the noise disappears and then stop. Going further does not remove more noise, because there is none left; it removes the voice instead.
• Use smoothing before lowering strength. If the result sounds metallic or watery, raise the smoothing control first. It spreads the reduction across neighbouring frequency bands and is the specific remedy for that artifact.
• A/B against the original, on headphones. The switch keeps your position, so you can flip back and forth on the same word. Artifacts that are invisible on speakers are obvious on headphones, and vice versa for low-frequency rumble.
• Fix hum with the notch, not the strength. Turning the reduction up until a buzz goes away costs far more signal than a notch that removes only the buzzing frequencies.
• Normalise afterwards, not before. Removing noise lowers the overall level slightly, so set the final loudness once the noise is gone.

Powered by optimised audio processing algorithms.

• Audio normalizer
• Vocal remover
• Audio equalizer
• Silence remover
• Audio compressor
• Audio converter