Analyze every 10 ms
The recording is converted to mono 48 kHz and cut into 10-millisecond frames. For each frame the network estimates how much of every frequency band is speech and how much is noise.
Clean hiss, fan noise and room hum out of a voice recording with an AI model — processing stays in your browser.
Drop a voice recording here
or click to choose from your computer
MP3 · WAV · FLAC · AAC · M4A · OGG
Working…
The AI Noise Reducer takes a voice recording and removes the background noise around the voice: hiss from a microphone preamp, fan and computer noise, air conditioning, traffic and the empty-room sound between sentences. It is meant for podcasts, voice-overs, interviews, lectures, voice notes and screen recordings.
Drop in a file, listen to Original and Cleaned one after the other, adjust Strength, and download the version you like. There is no account, no upload queue and no server round trip.
Noise reduction is different from simply turning the volume down or using an equalizer. An equalizer cuts a fixed frequency range for the whole file, which also dulls the voice. The AI model decides frame by frame which part of the sound is speech and which is noise, so the voice stays clear while the hiss and hum between words fade out.
Typical results: a laptop or phone recording with a constant fan or hiss drops by several decibels in the pauses, while the speech itself stays almost unchanged. Very loud, non-stationary noise such as a passing siren or a dog barking in the middle of a sentence is reduced less, so check the result by ear before you publish.
The tool uses RNNoise, a small neural network published by Xiph.Org. It was trained on speech mixed with many kinds of noise, so it learned to tell the two apart in real time.
The recording is converted to mono 48 kHz and cut into 10-millisecond frames. For each frame the network estimates how much of every frequency band is speech and how much is noise.
Bands that are mostly noise are turned down. Bands that carry voice are left alone, so consonants and the tone of the voice stay natural.
The cleaned signal is mixed with the original according to the Strength slider. Optional normalization brings the peak level back to a healthy volume before you export.
The model is trained on speech, so it works best on voice. It can also help with noisy background audio, but it is not a music tool.
Start with the highest quality source you have. Heavily compressed audio gives the model less to work with.
Compare Original and Cleaned during a pause between sentences, where the noise is easiest to hear. If the voice starts to sound thin, lower Strength by 10–20%. If you still hear noise, raise it, or cut low rumble with the Equalizer.
After cleaning, you can trim the silence with the Audio Trimmer, adjust the speed with the Speed Changer, or raise the final loudness with the Volume Booster.
Yes. It runs RNNoise, a small neural network trained to tell speech from background noise. The model is compiled to WebAssembly and runs inside your browser tab, so no audio leaves your device.
Broadband noise: hiss from a preamp or tape, computer and fan noise, air conditioning, traffic and room tone. Steady mains hum at 50 or 60 Hz is only partly removed. For hum, use the Equalizer to cut the lowest frequencies.
It is built for speech. On a full mix it can make cymbals and reverb sound dull or watery, so use the Light setting and check the result on the chorus before you export.
It blends the original with the cleaned version. 0% is the untouched recording and 100% is the full model output. Light, Standard and Strong are quick presets. Standard keeps most of the natural sound of the voice while removing the worst noise.
Full strength removes more of the noise but can also take small pieces of quiet speech with it, especially breaths and soft consonants. Drop to Standard or Light and the voice usually sounds more natural.
No. The file is decoded, resampled and cleaned on your device. The cleaned file is encoded in the browser as MP3 or WAV and downloaded directly. Nothing is sent to Sound Booster servers.
The model runs faster than real time: in our tests a 10-second clip took under a second on a desktop browser. Longer files take proportionally longer, and the progress bar shows how far along the process is.
MP3 at 192 or 320 kbps is fine for podcasts, voice notes and video. Choose WAV if you will keep editing the file in a DAW or another audio editor, so the audio is not compressed twice.
Keep editing in the browser after you clean the audio.