You recorded something important with your laptop's microphone, without realizing the computer's fan, the AC, or background traffic got recorded right along with your voice. That constant hum or hiss in the background is exactly the kind of problem a spectral noise reducer is designed to fix — but only if you understand how it works and where its limits are.
Why this isn't the same as muting the audio
A noise gate simply cuts the audio below a certain volume — useful for removing background noise during pauses where nobody's talking, but useless while someone is actually speaking, because the background noise is still there, mixed with the voice, and can't be "muted" without muting the voice too.
Spectral noise reduction (afftdn in FFmpeg) does something different: it analyzes the audio's frequency spectrum, estimates what the "profile" of the constant background noise looks like (the pattern that's present the whole time, whether someone's talking or not), and subtracts it from the signal at all times, even while there's speech on top. It's the difference between "cut when nobody's talking" and "remove the constant noise even while someone is talking".
The three intensity levels
- Light (6 dB reduction): for barely noticeable background noise, minimizes the risk of affecting the voice.
- Medium (12 dB): the reasonable starting point for most recordings with moderate background noise (fan, distant traffic).
- Strong (24 dB): for noticeable background noise, at the cost of a higher risk of artifacts.
Why too much reduction sounds "metallic"
The algorithm has to decide, frequency by frequency, which part of the signal is "noise" and which is "real voice" — and that distinction is never perfect. When you force very aggressive reduction, the algorithm starts removing parts of the voice that overlap in frequency with the estimated noise too, leaving audible artifacts: a robotic or "bubbly" timbre that's itself a kind of distortion, different from the original noise but just as annoying. If you notice that effect, the right fix is almost always to lower the intensity, not raise it.
When noise reduction can't help
This technique works best with constant, relatively uniform noise (electrical hum, fans, AC, background hiss). It doesn't work well on sudden, unpredictable noises (a slamming door, a bark, a knock) because there's no stable "profile" to subtract — for those cases you need to manually edit the specific section.
Frequently asked questions
Can I remove noise completely? Not without also affecting the voice — there's always a trade-off between how much noise you remove and how much you degrade the useful signal. The realistic goal is reducing it to a non-annoying level, not eliminating it 100%.
Why does my voice sound "robotic" after strong reduction? It's a typical artifact of forcing too much reduction — lower the intensity until that effect disappears.
Does it work on sudden noises like a door slam? That's not the ideal use case; it works best on constant, uniform background noise.
Is it processed in my browser? Yes, with FFmpeg compiled to WebAssembly, with nothing uploaded to any server.
Reduce background noise from your recordings with the noise reducer, with three intensity levels or manual control, 100% in your browser.