Tutorials Audio

How Noise Reduction Software Tells Noise from Signal

By Infinite Creation · Intermediate · ~14 min

A 1960s mid-century comic book style editorial illustration showing two audio engineers at a mixing desk, with a mixing desk, a studio microphone on a boom arm, an audio waveform display visible in the frame.

Overview

Noise reduction software doesn't magically know what's 'noise' versus what's your voice. It typically works from a sample of the noise alone, then subtracts that specific noise pattern from the rest of the recording, and understanding this explains both why noise reduction works well in some cases and why it introduces watery, robotic artifacts in others.

Steps

The simple version: it learns what noise sounds like, then removes that pattern

At the simplest level, most noise reduction tools need a sample of the noise by itself (a moment with no wanted signal, just the background hum or hiss) to learn what that specific noise's frequency pattern looks like, then apply that learned pattern to reduce the same noise throughout the rest of the recording.

What's really happening: spectral subtraction across frequency bands

Technically, most noise reduction tools work by analyzing audio in many narrow frequency bands simultaneously (not as a single overall signal) and subtracting the learned noise pattern's energy from each band where it's present, leaving bands that contain mostly wanted signal relatively untouched, a technique broadly called spectral subtraction or spectral gating.

Why aggressive settings cause watery or robotic artifacts

When noise reduction is pushed aggressively (removing more than the software can cleanly distinguish from the actual wanted signal), it starts removing parts of frequency bands that contain a mix of both noise and wanted signal, producing the characteristic watery, robotic, or underwater-sounding artifacts noise reduction is known for: this happens specifically because noise and voice often overlap in frequency, and no algorithm can perfectly separate them.

Why a clean noise sample matters more than aggressive settings

Because the process depends on an accurate sample of the noise alone, a longer, cleaner noise-only sample (with no wanted signal bleeding in) generally produces better, more artifact-free results than pushing the reduction amount higher on a poor or brief noise sample, getting a good sample matters more than maximizing the reduction slider.

Tested: what a denoiser takes from the noise and from the voice

We tested steps 2 to 4 on 2026-09-28 with FFmpeg 8.1.2 and its afftdn filter, which reduces noise by frequency band from a noise sample, as step 2 describes. We made a 15-second recording with a computer: 3 seconds of steady hiss alone, then a speech-like signal with pauses, 20 dB above the hiss. The speech has voiced sounds built from harmonics, and short 's' sounds in the same high frequencies as the hiss. Because the speech and the hiss were made separately, we could measure what the filter took from each. It learned the noise from the first 2 seconds, then ran at five settings of its reduction amount. The two speech columns compare the result with the clean speech, so before any reduction the hiss itself adds 0.7 dB to the 's' sounds. The last column counts isolated spikes in the pauses, the short tones heard as watery or twinkling "musical noise".

Reduction amount (nr)	Noise left in the pauses	's' sounds (4 to 9 kHz)	Voiced speech (100 Hz to 4 kHz)	Isolated spikes in the pauses
No noise reduction	0.0 dB	+0.7 dB	+0.0 dB	0.1%
6	-6.0 dB	-5.0 dB	+0.0 dB	0.2%
12 (the default)	-11.8 dB	-9.8 dB	+0.0 dB	0.7%
24	-22.0 dB	-15.3 dB	+0.0 dB	2.8%
48	-26.7 dB	-16.6 dB	+0.0 dB	7.9%
97	-26.7 dB	-16.6 dB	+0.0 dB	28.0%

Step 2 holds. The voiced speech came through every setting unchanged, to within 0.1 dB, because in its bands the speech was far louder than the hiss. The 's' sounds are where step 3 shows. They sit in the same frequencies as the hiss, and the filter took them down with it: 9.8 dB at the default setting and 15.3 dB at 24.

Pro Tip 2 holds, and the last column is the reason. Above 48 the filter removed no more hiss (26.7 dB at 48, 26.7 dB at 97), but the isolated spikes rose from 7.9% to 28.0% of the pauses. In the untreated hiss it was 0.1%. So past a point, turning the amount up buys nothing but artifacts, and turning it down (Pro Tip 3) is the fix.

Step 4 held for a clean sample, but length mattered less than it says. With samples of 0.05, 0.5 and 2 seconds the hiss fell by 22.6, 22.0 and 22.0 dB at nr 24. Our hiss never changes, so a twentieth of a second described it. A longer sample earns its keep when the noise varies, which we did not test. A contaminated sample was a different story. Taken over two seconds of speech, it left the hiss almost untouched (-0.3 dB), most likely because the filter learned the voice's shape, loud in the low bands and quiet where the hiss is, as the shape of the noise.

The sample is not the whole story in FFmpeg. Its source code uses the sample only for the noise's shape across frequency, and takes the noise's level from a separate noise floor setting (nf), -50 by default. At that default, even the maximum amount of 97 removed only 10.9 dB of our hiss, while still raising the spikes to 20.1%. The table uses -40, the value in the filter's own documented example of sampling noise. If afftdn removes less than you expect, try the noise floor before the amount. The learned profile also only takes effect from the moment sampling stops, so take the sample from the start of the file.

The documented pattern, learning the noise from the first 2 seconds of a file: ffmpeg -i input.wav -af "asendcmd=c='0 afftdn sn start',asendcmd=c='2 afftdn sn stop',afftdn=nr=12:nf=-40" output.wav

Validated. Measured 2026-09-28 with FFmpeg 8.1.2 (afftdn, asendcmd) and numpy. The speech and hiss are computer-generated, so the figures describe this recording, not every voice or room. The hiss is steady, and we did not test noise that changes, such as traffic or a fan speeding up. Every figure is measured from 6 seconds in, after each sample had ended, with the filter's delay removed. The spike count is a measurement of musical noise, not a listening test. Other denoisers work differently.

Pro Tips

Capture a longer, cleaner noise-only sample (room tone with absolutely no talking) before recording, rather than relying on the software to guess from a brief or contaminated sample.
Use moderate noise reduction settings rather than maximum, pushing too aggressively removes overlapping frequency content that contains your actual voice, causing watery or robotic artifacts.
If noise reduction artifacts are audible, try reducing the effect amount rather than assuming you need a different tool. The artifact usually comes from over-aggressive settings, not the software being fundamentally inadequate.

Knowledge Base

What You'll Learn

Noise reduction works by learning a noise sample's specific frequency pattern and subtracting it from the rest of the recording across many frequency bands, and understanding this explains why a clean noise sample matters more than aggressive settings, and why pushing too hard causes characteristic artifacts.

The Plain-English Version

Noise reduction software needs to hear what the noise sounds like by itself first, the same way you'd need to hear a specific hum in isolation before you could describe it precisely: it then looks for and reduces that same pattern throughout the rest of the recording, which works well when the sample is clean and fails when it isn't.

The Misunderstanding This Clears Up

People sometimes assume noise reduction is a blunt volume-based tool, and are confused when pushing the reduction amount higher makes voices sound worse rather than cleaner. The artifacts come from frequency overlap between noise and voice, not from the tool being poorly made, and a better noise sample usually solves it more reliably than a higher reduction setting.

Where This Fits

Noise reduction works best on a clean recording, and Audio Recording Fundamentals explains the mic distance and room choices that keep the noise floor low.

FAQ

Q: Why does aggressive noise reduction make my voice sound robotic or underwater?
A: This happens because noise and voice often share overlapping frequency content, and pushing noise reduction aggressively removes parts of that overlapping content along with the noise: the artifact is a direct result of the algorithm's frequency-band subtraction approach, not a sign of a broken or low-quality tool.

Q: What can I do to get better noise reduction results without artifacts?
A: Capture a longer, cleaner sample of the noise by itself (with absolutely no wanted signal in it) before or after your recording, and use moderate rather than maximum reduction settings. A clean noise sample generally improves results more reliably than pushing the reduction amount higher.