Remove background noise from audio, right on your phone

What counts as background noise

Background noise is everything the microphone picks up besides the voice you meant to record. For a noise-removal engine built for speech, what matters is less where a noise comes from than how it behaves over time (steady, changing, or in short bursts) and whether it sounds like a voice.

Real recordings rarely hold just one: a street interview carries traffic and wind, a voice-over at home a fan and the hum of a charger. Noise Remover handles every family it supports in a single pass, with no noise type to pick.

How the cleaning works

Noise Remover cleans with RNNoise, an open-source neural network for speech denoising from Xiph.Org, which runs on the phone itself. Each file is first converted to 48 kHz and passed through an 80 Hz low-cut filter that takes out the deepest rumble. The model then works through the audio in 10 ms frames, splits each one into 32 frequency bands, estimates how much of every band is voice, and turns down the rest.

Because it decides frame by frame, it needs no silent stretch to learn the noise first, and it keeps up with noise that changes, such as a passing car. It also reports, for every frame, how likely it is that someone is speaking, which is how the app can tell when a cleaning has turned your voice down along with the noise.

What you hear afterwards is a mix you control. The intensity slider starts at 80%, which caps the reduction near 29 dB to keep the voice natural, and the ready-to-publish level, on by default, evens the result out toward −16 LUFS. The whole pipeline, step by step, is on how it works.

What the app can’t remove

The model is trained to keep speech. Anything that sounds like speech, or that is part of the voice itself, is out of its reach:

• Other voices. A second person talking over the speaker, or the chatter of a café, is barely lowered and stays in the cleaned file, and chatter as loud as the speaker turned the main voice down a little too, by up to 3 dB in our lab test. Interviews where people take turns are fine, and the app shows a tip when a cleaning made your voice noticeably quieter.
• Music and TV. A television carries voices, which stay, and a song keeps its singer as a second voice. An instrumental background came down in our test, but music is not what the model is made for, and results vary.
• Room echo. Reverberation is your own voice bouncing off the walls, so it stays: the noise on top of it goes, but the voice still sounds distant. Echo and background noise are different problems.
• Noise as loud as the voice. The noise goes, but the voice comes out thinner.
• Recordings that are already clean. Cleaning a clean voice can colour it slightly; the default intensity and the Original/Cleaned switch are there to catch it.

How much noise it removes, noise by noise

In our lab test at full intensity, clean speech was mixed with each noise at four levels, from as loud as the voice to 20 dB below it, then cleaned by the same engine the app runs. The chart shows how far each noise dropped in the speaker’s pauses, and how much the voice’s own level changed. The mark at 29 dB is the app’s default: at 80%, it deliberately stops short of the model’s full strength.

The right-hand column matters as much as the bars. On the noises the app is made for, the voice moved by 1.6 dB at most. The rows set apart are the ones the model barely lowers: other voices, which it keeps as speech, and music. With chatter as loud as the speaker, the speaker lost up to 2.8 dB. The noises were synthetic, and a real recording adds rooms and microphones the test did not cover.