Background Noise Reducer

How the Background Noise Reducer works

Steady background noise — air conditioning, computer fans, tape and preamp hiss, the rumble of traffic through a window — sits underneath everything in a recording. Turning the volume down lowers it along with the voice, and an ordinary noise gate only helps in the pauses. A spectral noise reducer works differently: it splits the sound into about a thousand narrow frequency bands and decides, band by band and moment by moment, which contain wanted sound and which contain only noise.

How the background noise reducer works

The recording is cut into overlapping frames of 2,048 samples, about 46 milliseconds at 44.1 kHz, and each frame is converted to a spectrum with a fast Fourier transform. The noise print is the average spectrum of the section you selected: for every frequency band it records how loud the noise is on its own. Each frame of the full recording is then compared to that print. Bands that rise more than the threshold above the noise are kept untouched; bands that do not are turned down by the reduction amount. The frames are converted back to sound and overlapped so that, with no reduction, the output is identical to the input sample for sample.

Two kinds of smoothing keep the result natural. Each band's decision is averaged with its neighbours, so isolated bands do not switch on and off, and gains that fall are eased down over a few frames while gains that rise open straight away. That keeps the starts of words crisp and avoids most of the chirping, "underwater" sound known as musical noise.

Choosing the noise print

The print is the most important setting. Select a section with only background noise — the second before someone starts speaking, a pause between sentences, or the end of the file after the last word. Half a second is the minimum that works well; two or three seconds is better. Avoid breaths, clicks and room echo of the voice, because anything in the print is treated as noise and will be reduced wherever it appears.

With no selection, the tool picks the quietest 10% of the file's frames and averages those. That works well for interviews and lectures with natural pauses. For music or continuous speech with no gaps, make a selection.

Reduction and threshold

Reduction is how far noise-only bands are turned down. 10 to 15 dB is a clear improvement that still sounds natural; 20 to 25 dB makes hiss nearly vanish; above 30 dB the background can sound unnaturally dead and the voice can pick up a watery edge. Most engineers prefer the lowest setting that makes the noise stop being distracting.

Threshold is how much louder than the noise print a band must be to count as wanted sound. A low threshold, 2 to 4 dB, protects quiet details such as soft consonants and reverb tails but lets more noise through. A high threshold, 10 dB or more, cleans harder but starts to eat into the voice. Start at 6 dB and move it a few steps either way while previewing.

What it cannot fix

Spectral gating works on steady noise. It does very little against noise that changes from moment to moment: other people talking, music in the background, dogs barking, keyboard clicks, wind buffeting a microphone or clipping distortion. Mains hum is steady but is better handled by a notch filter that removes only the hum frequencies; the Hum Remover does that with almost no effect on the voice.

Noise that is louder than the voice in the same frequency band cannot be separated by this method, because there is nothing to tell the two apart in the spectrum.

Tips

• Record a few seconds of room tone. Before you start talking, record silence. It gives a perfect noise print and makes cleaning far easier.
• Clean first, then process. Run noise reduction before compression or normalisation. Those raise the noise floor, so reducing it afterwards is harder.
• Combine with a gate. For speech, follow gentle reduction here with the Noise Gate to silence the pauses completely.
• Check on headphones. Musical noise and dulled consonants are easy to miss on laptop speakers. Listen to the preview closely before exporting.
• Look at the noise. The Spectrogram shows hiss as a haze across the top and hum as horizontal lines at the bottom, which tells you what kind of noise you are dealing with.

Related tools