
With any noise reducer, including the free Noise Reducer on this site, you've gone too far once you start hearing the processing instead of the noise. The signs are a watery or chirpy residue, a voice that turns dull or lispy, and a background that breathes in and out between phrases. Every extra decibel of reduction swaps a bit of noise for a bit of artifact. The target is noise sitting well below the voice, where nobody notices it. Getting it to zero isn't the goal.
That runs against what most people expect. The waveform looks cleaner as the gaps flatten out, so turning the reduction up feels like progress. But a light hiss under a clear voice sounds fine to most listeners, and a silent background under a voice that seems to come through a drain pipe does not.
Over-processing has a few recognizable sounds. Once you can name them, you'll know when to stop.
This is the classic one. When a noise reducer cuts hard, the noise doesn't disappear evenly. Small random pieces of it survive in scattered frequency bands, and they pop in and out as short tonal blips. Together they make a bubbling, warbling texture people describe as "watery", "chirpy" or "underwater". It's often more distracting than the hiss it replaced, because it moves. What Causes Noise Reducer Artifacts explains where each of these sounds comes from.
Consonants are noise-like by nature. An "s", "f", "sh" or "t" is a short burst of broadband hiss, physically not far from the room noise you're trying to remove. Push the reduction hard enough and the processor starts treating those sounds as noise as well. Sibilants go soft or smeared, the top end of the voice dulls, and speech gets a slight lisp. The voice can also lose the low-level detail that gives it body and sound thin, even when no single word is obviously damaged.
When a person talks, the voice covers the noise. When they pause, the processor clamps down, and the background drops to near silence. Then the next word starts, the gate opens, and the noise floor swells back up underneath it. That rise and fall is called pumping or breathing, and it's very noticeable on headphones.
Songs have quiet, noise-like content on purpose. Reverb tails, cymbal decay, brushed snares, breathy vocals and the faint sound of the room all sit at low levels with a wide, hissy spectrum. Aggressive reduction can't tell those apart from tape hiss, so it trims them along with the noise. The mix ends up drier and smaller, and notes that used to ring out stop short.
A low, steady noise floor is part of what makes a recording sound real. Every room has one, called room tone, and film and podcast editors record a stretch of it on purpose to fill gaps, because dead digital silence between lines sounds edited and slightly robotic.
The voice also masks noise for you. While someone is speaking, the louder speech covers quieter sound in the same frequency range, so a hiss that's obvious in a pause mostly disappears under words. So you don't need the noise gone everywhere, just low enough that the pauses don't draw attention.
Consistency matters more than how quiet the background is. If you edit takes together, a steady, low background across every cut sounds smooth, while a background that's silent in one sentence and hissy in the next makes every edit audible. A little noise, applied evenly, beats none at all applied unevenly.
The reliable way to set noise reduction is to compare against the untouched original and back off as soon as the processing becomes audible. The Noise Reducer is built for that. It processes the file once, then gives you a live Noise reduction slider that blends the original with the cleaned version as it plays. Moving it doesn't re-run anything, and the export uses whatever mix you settle on.
If you're used to dB-based editors, a common rule of thumb there is to take off around 6–10 dB at a time and stop when you hear artifacts. The mix slider gets at the same thing differently: you choose how much of the cleaned signal to keep, and your ears decide where it stops.
To see what you're hearing, load the before and after files into the Spectrogram. Noise shows up as a haze across the whole picture. You want that haze clearly darker after processing, not wiped out, with the bright detail of the voice or the cymbals still intact.
How hard you can push depends on what's in the recording, and the two modes fail in different ways. AI vs Spectral Noise Reducer: Which to Use covers the choice in depth.
Use Voice mode. It runs an AI speech model in your browser, so it can follow noise that changes over time, like traffic, a fan cycling on, or keyboard clatter. The first run downloads the small model once, and a long file takes a while because it processes on your own machine. Voice mode keeps the voice and removes nearly everything else, so the usual over-processing symptom is speech that sounds thin or colored, not musical noise. If that happens, blend some original back in with the slider. For the inner workings, see How AI Noise Reduction Works.
Use Music mode, and never run a song through Voice mode. Voice mode treats instruments, pads, reverb tails and cymbals as noise and removes them. Music mode is spectral gating, not AI. It keeps stereo and the full frequency range, builds a noise profile from the quietest moments in the file, and caps its maximum cut at about 16 dB on purpose. So it can't hollow out a mix, but some noise will remain. That's expected, and usually the right result. If the leftover is a mains drone, tick the 50/60 Hz hum option; How to Remove Hum and Buzz from Audio explains why hum needs different treatment from hiss.
The same rules apply to a video soundtrack, with one extra step. Extract the audio first with Video to Audio, clean it in the Noise Reducer, export it as WAV, then put it back in your video editor in place of the original track. On video, a natural voice matters more than a silent background, so err toward a lower setting. The full walkthrough is in How to Remove Background Noise from Audio.
Some noise can't be reduced cleanly at any setting. When a loud noise sits right on top of the voice, in the same frequencies at the same moment, like a siren under a sentence or a crowd talking over a speaker, there's no way to remove one without damaging the other. The slider will show you that no setting sounds good.
In those cases the real fix comes earlier. Getting the mic closer to the source raises the voice relative to the room, which reduces the noise you have to deal with later without adding any artifacts. Mic Placement for Clean Voice Recordings covers distance and angle.
Echo is a different problem too. A boomy, reflective room isn't background noise, and turning noise reduction up won't fix it. See How to Remove Echo and Reverb from Audio for that.
Can you remove background noise completely?
Not without side effects. Noise and the quiet parts of a voice or song share frequencies, so removing every trace of noise also removes consonants, breaths, reverb tails and room ambience. The result sounds watery or robotic. A better target is noise sitting well below the voice, low enough that listeners stop noticing it. That reads as clean, and it keeps the recording sounding natural.
Why does my voice sound underwater after noise reduction?
That watery or underwater sound is over-processing. When reduction is pushed hard, small fragments of noise survive in scattered frequency bands and flicker in and out, and parts of the voice get treated as noise too. Lower the reduction. In Vocal Cut's Noise Reducer, pull the Noise reduction slider down from the 75% default and compare against 0% until the voice sounds natural again.
Is it better to do several gentle passes or one strong one?
In dB-based editors, a common rule of thumb is several gentle passes of around 6–10 dB, stopping when artifacts appear, because each pass is easier to judge. Vocal Cut works differently: it processes the file once, and the mix slider sets how much of the cleaned version you keep. Feeding an already-cleaned export back through compounds the artifacts, so adjust the slider instead.
Does Vocal Cut upload my audio to remove noise?
No. Noise reduction runs in a Web Worker in your browser, so your audio never leaves your device. Voice mode downloads a small AI speech model the first time you use it, and it's cached after that. Music mode is spectral gating and needs no download. Either way, the cleaned file is created on your own computer and you export it as WAV, MP3, or Opus where the browser supports it.