How the cleaning works, on your phone and offline
Noise Remover removes noise with RNNoise, an open-source neural network for speech created by Xiph.Org, running entirely on your phone. Your audio is decoded, filtered, cleaned and exported without ever leaving the device.
From your file to a clean export
- Decode
Audio and video files are read with Android’s own decoders. From a video, only the soundtrack is cleaned, and the picture is never altered.
- Resample
Everything is converted to 48 kHz, the rate the noise model works at.
- Low-cut
A gentle filter removes the rumble below 80 Hz before the model hears it.
- Neural noise suppression
Every 10 ms, the network estimates how much of each frequency band is voice, and turns the rest down. A stereo file is cleaned in stereo, with both channels turned down together so the voice keeps its place.
- Intensity mix
A time-aligned copy of the original is blended back in to cap the reduction where you set it: 80% by default.
- Loudness
The final mix is measured (ITU-R BS.1770) and brought to −16 LUFS, unless the −1 dBTP ceiling comes first.
- Trim and export
The range you kept is encoded on the phone: M4A, WAV, MP3, Opus (Android 10+). With Pro, a video leaves as an MP4: the cleaned sound is encoded and placed beside the picture, which is copied as it is.
- Kept on the phone
The original and the export stay in the app’s storage, out of cloud backups, until you delete them. For a video, a copy of the video is kept too, so that it can be exported with the cleaned sound.
The AI model: RNNoise, built for speech
RNNoise combines classic signal processing with a recurrent neural network. It does not generate a new voice: 100 times a second, it computes one gain per frequency band from what it learned speech and noise sound like, and turns down the bands that hold noise. It was trained on dozens of open speech datasets covering many languages and accents, so it works with any spoken language.
Two consequences matter. It needs no noise sample: it follows the noise as it changes, a passing car as well as a steady fan. And because it only turns bands down, it cannot invent or change words. The worst it can do is turn down too much, which is what the intensity slider is for.
The model is small enough to run on a mid-range phone, faster than real time: a minute of audio takes seconds, not minutes.
What stays on the phone, and what doesn’t
Everything about your audio stays on the device: your recordings, the files you import, the soundtrack of your videos, the cleaned results and their history. The app has no server of its own, so there is nowhere to send them, and the cleaning works in airplane mode.
A free app still uses the network for a few things, and we would rather say which: ads in the free version (Google AdMob, with Google’s consent form where the law requires it), usage statistics and crash reports (Firebase), with no name, e-mail or account attached, and the purchase check if you buy Pro (RevenueCat). None of them receives your audio. The privacy policy has the details.
- No account, no sign-up, no e-mail address
- Your cleanings are excluded from Android’s cloud backup and device transfer
- Delete one cleaning, the originals or everything from the app
Measured results
We measured the engine before building the app around it. Clean speech was mixed with thirteen kinds of noise at four levels, from noise as loud as the voice (0 dB SNR) to a faint background (20 dB SNR), cleaned at full intensity, then scored. The chart shows twelve of them. The thirteenth, a song with vocals, is left out because it behaves like another voice: its singer stays in the cleaned file.
| Noise | Noise removed during pauses, in dB | Change in voice level |
|---|---|---|
| Hiss, white noise | 23 to 43 dB | −0.2 to −0.6 dB |
| Electrical hum | 24 to 41 dB | −0.3 to −1.6 dB |
| Broadband noise | 26 to 39 dB | −0.3 to −0.9 dB |
| Traffic | 24 to 37 dB | −0.1 to −0.3 dB |
| Low rumble | 25 to 34 dB | −0.3 to −1.0 dB |
| Wind gusts | 19 to 27 dB | −0.3 to −1.3 dB |
| Fan, air conditioning | 24 to 25 dB | −0.3 to −1.5 dB |
| Rain | 15 to 20 dB | −0.3 to −0.8 dB |
| Keyboard typing | 11 to 13 dB | 0.0 to −0.2 dB |
| Not what it is made for | ||
| Crowd chatter | 5 to 11 dB | −0.3 to −2.8 dB |
| Instrumental music | 8 to 9 dB | −0.2 to −0.3 dB |
| Another voice | 0 to 2 dB | +0.2 to −0.6 dB |
Two figures per noise: how much quieter the noise gets during the speaker’s pauses, and how much the voice itself changes. On steady and outdoor noise the voice stays within about 1 dB, 1.6 dB at worst, for a hum or a fan as loud as the voice. Keyboard clicks come down less, by 11 to 13 dB, but the voice does not move and measured intelligibility rises.
Why the default is 80%
Removing every last decibel of noise is not what sounds best. In our tests, speech quality peaked when the reduction was capped between 24 and 30 dB, a little below the maximum. The slider maps 0 to 100% to a cap from 0 to 36 dB, and 100% removes the cap; the default 80% sits near 29 dB.
| Reduction cap | Noise removed |
|---|---|
| 6 dB | 5.9 dB |
| 12 dB | 11.6 dB |
| 18 dB | 16.8 dB |
| 24 dB | 20.8 dB |
| 30 dB | 23.1 dB |
| No cap (100%) | 24.8 dB |
The original is delayed by the model’s 20 ms before it is mixed back in, so the two never phase against each other, and the Original / Cleaned switch compares them at the same moment and the same loudness. Keeping a voice natural goes further.
What it can’t do
- Other voices. When someone talks at the same time as you, the model hears two voices and keeps both, and can turn yours down a little. Interviews where people take turns are fine.
- Music and TV. An instrumental background is pushed down by only 8 to 9 dB, and a song keeps its singer. It is not a music tool.
- Room echo. Reverb is your own voice bouncing back, so it stays. Echo or noise?
- Noise as loud as your voice. It can be removed, but the voice comes out thinner. Record closer when you can: 12 recording tips.
- Already-clean audio. A clean recording gains nothing and can be slightly coloured. Compare, and lower the intensity.
Frequently asked questions
Does Noise Remover upload my audio to a server?
No. Noise Remover removes the noise on your phone with a model built into the app, and has no server to send audio to. Recordings, imports and results stay in the app’s private storage, are excluded from cloud backups, and the cleaning works in airplane mode.
Is Noise Remover an AI noise remover?
Yes. The cleaning is done by RNNoise, a recurrent neural network trained to tell speech from noise, running on the phone. It turns noisy frequency bands down 100 times a second; it does not generate audio or change what was said.
Do I need to record a sample of the noise first?
No. Unlike classic noise reduction, which learns a noise profile from a silent passage, the model in Noise Remover adapts on its own as the noise changes. Import or record, and the cleaning starts.
How long does a cleaning take?
It runs faster than real time on a mid-range phone: a minute of audio takes seconds. Long files take proportionally longer, and the progress is shown as a percentage with the waveform filling in.
Clean voice. Offline. One tap.
Record or import, let it clean, compare, share. Your audio never leaves your phone.
Free · Android 8.0 or later, 64-bit
An iPhone version is coming · Free and Pro