What Causes Noise Reducer Artifacts (and Fixes)

Amber waveform surrounded by stray fragments, illustrating the artifacts a noise reducer can leave behind

Every noise reducer, including the free Noise Reducer on this site, creates artifacts for the same basic reason: it has to guess which parts of the sound are noise, and some of its guesses are wrong. A noise reducer splits audio into many narrow frequency slices, called bins, over short slices of time, called frames. For each bin in each frame it decides how far to turn that slice down. Wrong calls, cuts that go too deep, and the limits of how finely it can see time and frequency at once each leave their own audible fingerprint.

How a noise reducer decides what to remove

Most noise reduction starts from a noise estimate: a picture of how loud the background is in each frequency bin. A classic spectral reducer builds it from the quietest stretches of the file, on the assumption that those moments are mostly noise. Then, frame by frame, it compares each bin to that estimate. A bin well above the noise is probably signal and is left alone. A bin close to the noise level is turned down.

To make the cut more effective, many reducers over-subtract, removing a bit more than the estimate says is there. That clears more noise, but it also pushes more borderline bins toward silence, and borderline bins are where artifacts come from.

The estimate itself can be wrong, too. A pad, drone or held organ note sits in every frame, so a naive "quietest level per bin" profile learns it as background and the reducer carves it out. Vocal Cut's Music mode guards against this by smoothing the profile across frequency: real noise floors are broad and even, while musical tones are narrow peaks, so the peaks don't get counted as noise.

AI speech models learn what a voice looks like instead of measuring the noise, but they still output a gain per bin per frame, so they fail in related ways. How AI Noise Reduction Works covers gates, spectral subtraction and AI models in more depth.

Musical noise: the chirpy, watery warble

Musical noise is the best-known artifact, and it comes straight from how noise behaves. Background hiss isn't perfectly steady. In any single bin, its level jumps around randomly from one frame to the next. The noise estimate is a typical level, so in some frames a few bins land above it by chance.

The reducer sees those bins poking above the line and treats them as signal. It lets them through while turning down everything around them. What survives are isolated scraps of noise, each one a narrow band lasting only a frame or two, scattered at random across the spectrum. Your ear hears them as short tones, sometimes called birdies. Together they make the bubbling, chirpy, watery texture people describe.

It gets worse as the recording gets noisier. When the voice is only a little louder than the background, more bins sit near the decision line, so more of them flip back and forth between kept and cut.

Two design choices fight it:

  • A gain floor. Instead of switching a bin all the way off, the reducer only turns it down by a set amount. The leftover noise stays as a soft, continuous hiss under everything, which hides the isolated chirps.
  • Gain smoothing. If a bin's gain can't jump wildly from frame to frame, or differ sharply from its neighbors, a single random spike can't open a bin by itself.

Vocal Cut's Music mode uses both. Its maximum cut is about 16 dB, and its gains are smoothed across time and across neighboring frequencies, specifically to keep musical noise down. The cost is that some noise remains.

The underwater, swirly sound

An underwater or swirly voice has a different cause: too much subtraction, usually combined with heavy smoothing. Here the reducer isn't letting noise through. It's taking voice out.

Many speech sounds are noise-like by nature. An "s", "f", "sh", "t" or a breath is a short burst of broadband hiss, physically similar to room noise. The high-frequency detail that makes a voice sound present and close also sits at a low level, right where hiss lives. When the subtraction is strong, those parts of the voice fall below the line along with the background.

What's left is the louder, tonal core of the voice with its edges trimmed. Heavy smoothing then blurs the gains across neighboring bins and frames, so the remaining sound moves in soft, shifting patches instead of crisp detail. It comes out hollow, a bit phasey, as if heard through a tube or water. Noise Reducer: How Much Reduction Is Too Much? describes how it sounds at different settings.

Smeared transients and soft attacks

Some artifacts come from the analysis itself, before any decision is made. A spectral reducer looks at the audio through a window of fixed length. A long window sees frequency finely, which helps separate noise from tones. But everything inside one window gets treated as if it happened at once.

That's the time-frequency tradeoff, and it hits sharp sounds. A drum hit, a plosive "p" or "b", or the start of a consonant happens in a few milliseconds. Inside a longer window, its energy is spread across the whole frame. When the reducer applies a gain to that frame, the attack gets softened, and a little of the sound can show up just before the actual hit, as a faint pre-echo.

Music mode uses a 2048-sample window, which at 44.1 kHz is about 46 ms, stepped forward every 512 samples so frames overlap heavily. That's a common, music-friendly choice. Short windows would keep attacks sharp but blur frequency detail, which makes the noise estimate less precise. No window length avoids both problems.

Pumping and breathing

Pumping is a time problem rather than a frequency one. While someone is talking, the voice covers most of the background, and the reducer keeps those bins open because they're well above the noise. In the gaps between phrases, it clamps down and the background drops away.

So the noise floor isn't constant. It's low in pauses and higher under speech, and it rises and falls with every phrase. A gentle reducer keeps that swing small enough that you don't notice it. An aggressive one behaves more like a noise gate, and you hear the background breathe in and out. Music mode limits it by opening quickly when signal arrives and closing slowly when it stops, so the floor eases down instead of snapping.

Artifacts specific to AI voice models

An AI speech model trades the classic artifacts for a different set. Vocal Cut's Voice mode runs GTCRN, an AI speech-enhancement model, in your browser. Because it learned what speech looks like, it handles noise that changes over time much better than a fixed noise estimate can. That's also why its failures look different.

  • Everything that isn't a voice counts as noise. Instruments, pads, reverb tails and cymbals all get removed. That's the right behavior for a podcast and the wrong one for a song, which is why the tool warns you not to run music through Voice mode.
  • Singing and overlapping voices can warble. A model trained mostly on speech has less to go on with sustained sung notes, or with two people talking over each other, and the output can waver.
  • The top end is duller. Voice mode processes audio at 16 kHz, and a 16 kHz signal can only hold frequencies up to 8 kHz. The air above that is gone, so the voice sounds a little less bright than the original. The output is also mono and focused on the speech band.

Underneath, it's still a wrong call about what counts as noise. The model just gets it wrong in different places. AI vs Spectral Noise Reducer: Which to Use matches each mode to the recordings it suits.

When the recording makes artifacts unavoidable

Some recordings will produce artifacts at any setting, because the noise and the signal can't be told apart.

  • Low signal-to-noise ratio (how much louder the voice is than the background). When the voice is barely above the background, nearly every bin is a close call.
  • Noise that overlaps the voice. A fan's broadband rush sits in the same frequencies as consonants, so cutting one cuts the other.
  • Noise that changes. Traffic, chatter or music in the room don't match a single noise profile. A spectral reducer either misses them or cuts too broadly. An AI model does better with speech, but it still can't pull a voice cleanly out from under another voice.

Steady mains hum is a separate case: a narrow notch handles it better than broadband reduction, as How to Remove Hum and Buzz from Audio explains.

The same limits apply if you want to remove background noise from video. Pull the soundtrack out first with Video to Audio, clean it in the Noise Reducer, then put it back in your video editor in place of the original track. The step-by-step version is in How to Remove Background Noise from Audio.

How to back off before artifacts appear

Artifacts are wrong guesses applied too hard. So the fixes come down to two things: give the reducer easier calls, and make its cuts shallower.

  • Use the mode that fits the recording. Voice mode for speech, Music mode for songs and anything with instruments.
  • Blend with the mix slider. Vocal Cut processes the file once, then the slider (75% by default) remixes the original back in without re-processing. Mixing some original back fills the gaps that cause musical noise and restores consonant detail. Slide to 0% to hear the untouched file for comparison.
  • Leave some noise in. A low, even background sounds natural. Nobody notices a little hiss under a clear voice, but most listeners notice a swirly one.
  • Fix it at the source next time. A closer mic raises the voice above the room, which means less reduction and fewer close calls. Mic Placement for Clean Voice Recordings covers distance and angle.
  • Listen on headphones. Musical noise and pumping are much easier to hear on headphones than on laptop speakers.

To check your result visually, compare the before and after files in the Spectrogram. Musical noise shows up as scattered bright specks where the even haze used to be.

Processing runs in your browser, so your audio never leaves your device. Voice mode downloads its small model once (about 0.5 MB) and caches it. Files can be up to 100 MB and 12 minutes long.

Frequently asked questions

Why does noise reduction make my voice sound underwater?

Too much subtraction. Parts of a voice, like "s", "f" and "sh" sounds, breaths and high-frequency detail, are noise-like and quiet, so a strong noise reducer removes them along with the hiss. Heavy smoothing then blurs what's left into a hollow, phasey sound. Lower the reduction, or blend more of the original back in with the mix slider.

What is musical noise in audio?

Musical noise is a chirpy, watery residue left by spectral noise reduction. Random peaks in the background noise briefly rise above the noise estimate in individual frequency bins, so the reducer keeps them while cutting everything around them. The surviving scraps sound like short, scattered tones. A gain floor and gain smoothing reduce it.

Can you remove noise reduction artifacts after the fact?

Generally no. Once a noise reducer has removed parts of the voice or left musical noise behind, that information is gone, and running the file through another pass usually adds more artifacts. Go back to the original recording and process it more gently, with a lower setting or more of the original blended back in.

Does AI noise reduction cause artifacts too?

Yes, but different ones. An AI speech model usually avoids the classic chirpy musical noise, but it removes anything that isn't a voice, including instruments and reverb. It can also warble on singing or overlapping voices. And a model that runs at 16 kHz loses everything above 8 kHz, so the voice can sound duller.