Key takeaways
- A robotic, watery or lispy sound means the reduction is cutting into the voice; it happens most when the noise is nearly as loud as the voice.
- A capped reduction keeps a voice natural: in our lab test, speech quality peaked with the cap around 24 to 30 dB, which is why Noise Remover defaults to 80%, about 29 dB.
- Overlapping voices are the hard case: the engine barely removes the other voice, so it stays, and the main one can come out a little quieter. Lower the intensity or re-record.
- Judge on headphones, at matched loudness, switching between the original and the cleaned version at the same moment.
What “robotic” actually is
Noise reduction works by turning down, many times a second, the parts of the sound it judges to be noise. When it judges right, the voice is untouched. When it doesn’t, or when it has to cut where the voice and the noise overlap, you hear artifacts, which people describe in a handful of ways.
Over-suppression: swallowed endings and breaths
The soft parts of speech (a word fading out, an “s”, an “f”, a breath) are quiet and noise-like, and a strong reduction takes them for noise. Words end abruptly, sentences sound clipped, and the voice loses the small sounds that make it human. The silence between phrases can also feel unnaturally dead.
Musical noise: the watery, underwater sound
When the reduction leaves scattered fragments of noise that flicker from one moment to the next, they sound like faint warbling tones: “musical noise”, the bubbly, watery texture of over-processed audio. It was the classic flaw of older spectral tools. Neural engines tend to produce less of it, but push any of them hard enough and some of that texture comes back.
A thin voice when the noise is loud
When the noise is as loud as the voice, the two share the same frequencies at the same moments, and removing one takes part of the other. The noise goes, but the voice comes out thinner and slightly hollow. In our lab test, with a fan or hum as loud as the speech, the noise went while the voice lost 1.5 to 1.6 dB and some of its body. How we measured.
Overlapping voices
A speech engine is trained to keep speech, and a second voice is speech too. When someone talks over the speaker, the engine cannot keep one voice and drop the other, so it keeps both. In our lab test, overlapping voices were barely removed (0 to 11 dB), and with voices as loud as the speaker, the main voice came out up to 3 dB quieter and harder to follow. Speakers taking turns, as in most interviews, were fine.
| Noise | Noise removed during pauses, in dB | Change in voice level |
|---|---|---|
| Electrical hum | 24 to 41 dB | −0.3 to −1.6 dB |
| Traffic | 24 to 37 dB | −0.1 to −0.3 dB |
| Fan, air conditioning | 24 to 25 dB | −0.3 to −1.5 dB |
| Not what it is made for | ||
| Crowd chatter | 5 to 11 dB | −0.3 to −2.8 dB |
| Another voice | 0 to 2 dB | +0.2 to −0.6 dB |
A clean recording, slightly coloured
Cleaning a recording that had little noise to begin with does not make it better. In our lab test, a clean voice came out slightly coloured: its PESQ score (a standard measure of speech quality against a reference, where 4.64 means identical) fell to 4.10. If the noise is faint, a low intensity, or no cleaning at all, may be the right call.
How Noise Remover keeps a voice natural
The intensity is a cap, in decibels
The intensity slider sets how far, at most, the noise can be turned down: the cap rises evenly from 0 dB at 0% towards 36 dB, and at 100% it is lifted. Noise Remover does it by mixing the original back in, far below the cleaned voice, so nothing, noise or voice, ends up more than the cap below where it was. A consonant the engine mistook for noise is turned down by at most that amount, instead of erased.
In our lab test on fan noise, the speech quality score (PESQ) peaked with the cap around 24 to 30 dB, above the uncapped result, while removing everything took out only a little more noise: 23.1 dB with a 30 dB cap, against 24.8 dB with none. That is why the default sits at 80%, a cap of about 29 dB. The full table is on the how-it-works page.
The original is lined up to the millisecond
The engine works in 10 ms slices and hands back its result about 20 ms late. Mixing that with the original as it is would put two copies of the voice 20 ms apart, which sounds hollow and phasey: the delay cancels a comb of frequencies spaced every 50 Hz. Noise Remover delays the original by exactly the engine’s lag before mixing, so the two line up.
A/B at the same moment and the same loudness
The ear’s memory for fine detail fades within seconds, and the louder of two versions almost always sounds better. So the Original / Cleaned switch changes track at the same position, with no gap, and the original plays at about the loudness the cleaned version will have once exported. What changes when you switch is the noise, not the level.
A tip when your voice came out quieter
During the cleaning, Noise Remover compares the level of your voice before and after, over the moments the engine marks as speech. If the cleaned voice came out more than 6 dB quieter, a tip under the slider suggests a lower intensity. In our lab test, ordinary noises left the voice within 1.6 dB, and even overlapping voices within 2.8 dB, so the tip is kept for recordings harder than any we measured, usually voices talking over each other.
Hear the default for yourself: a voice next to a fan, cleaned at 80%. Switch while it plays, and listen to the breaths and the ends of the words.
Hear the difference
A voice over the noise of a fan, cleaned by the app’s own engine. Press play, then switch between Original and Cleaned.
Practical rules
| Intensity | Cap on the reduction | Use it for |
|---|---|---|
| 50% | about 18 dB | Light noise, a voice that sounds processed, overlapping voices |
| 70% | about 25 dB | Moderate noise, when the default thins the voice |
| 80% (default) | about 29 dB | Most recordings: fans, air conditioning, traffic, wind |
| 90% | about 32 dB | Loud, steady hum or hiss |
| 100% | no cap | The full reduction, as long as the voice still sounds like you |
- Start at 80%. It suits most voices recorded near a fan, air conditioning, traffic or wind.
- Go down to between 50 and 70% if the voice sounds processed: watery, metallic, lispy, or with clipped word endings. A trace of noise is far less noticeable than a damaged voice.
- Go up to between 90 and 100% for loud, steady hum or hiss, and stop as soon as the voice changes.
- Judge on headphones. A phone speaker hides both the low hum and the artifacts; earbuds reveal them. Then check once on the speaker, where many people will listen.
- Compare at matched loudness, at the same moment. Noise Remover’s A/B switch does both; with other tools, match the levels before you judge.
- Listen to the pauses and the word endings, not only to the middle of sentences: that is where artifacts show first.
Record cleaner in the first place
Every rule above is easier with a better recording. The closer the microphone, the louder your voice against the noise, and the less the engine has to take out: halving the distance gains about 6 dB. A softer room removes the echo no denoiser can touch, and a quiet moment keeps other voices out. The 12 recording tips cover the rest.
Try the intensity slider on your own recording, offline.
Free · Android 8.0 or later, 64-bit