Remove background noise from audio in your browser
This audio noise remover cleans up voice recordings: the hum of a fan or air conditioner, computer noise, traffic outside the window, tape or microphone hiss, and the 50 or 60 Hz buzz of mains electricity. It is made for podcasts, interviews, lectures, voice-overs, online-class recordings and voice notes that were captured in less-than-ideal rooms.
Three tools work together. RNNoise, a small neural network trained to tell speech from noise, removes changing background sounds around a voice. A spectral noise gate learns the fingerprint of a steady noise — automatically from the quietest moments, or from a noise-only part you select — and turns it down across the whole file, which also works on music. A hum filter notches out mains hum and its harmonics, and an optional low cut removes rumble.
The recording never leaves your device. The noise model is downloaded once and runs in your browser, so you can safely clean confidential meetings and interviews.
How to use the Audio Noise Remover
- Drop a recording onto the box or click to choose one; you can open MP3, WAV, M4A, OGG, Opus, FLAC or WebM audio.
- Leave AI speech denoise on for voices and set the Strength; lower it if the voice starts to sound thin.
- For steady hiss or noise on music, switch on the Spectral noise gate. Let it learn from the quietest parts, or drag across a part of the waveform that contains only noise.
- If you hear a low buzz, choose the Hum filter for your country’s mains frequency, and tick Low cut for rumble.
- Press Preview and Original to compare eight seconds, adjust, then press Remove noise and Download.
Audio Noise Remover features
RNNoise speech enhancement
Xiph’s recurrent-network denoiser (BSD licence, WebAssembly build by Shiguredo under Apache-2.0) processes 10 ms frames at 48 kHz; your audio is resampled to 48 kHz and back automatically.
Spectral noise gate
Learns a noise profile and lowers it by 6–24 dB per frequency band, with smoothing that avoids the “underwater” chirps of crude noise gates.
Select the noise yourself
Drag across a pause in the recording to teach the gate exactly what the noise sounds like, or type the times.
50 / 60 Hz hum filter
Narrow notches at the mains frequency and its first eight harmonics remove buzz without touching the voice.
Strength control and preview
Blend the cleaned and original sound, and hear the result on a short excerpt before processing the whole file.
Speech detection report
After processing, see in how much of the recording RNNoise detected speech — a hint when a file is mostly music.
Which filter for which noise?
| Noise | Best tool | Notes |
|---|---|---|
| Fan, air conditioner, computer, traffic around speech | AI speech denoise | Handles noise that changes over time |
| Steady hiss (tape, cheap microphone, high gain) | Spectral noise gate | Learn from a pause in the recording |
| Low buzz from cables or lights | Hum filter 50 or 60 Hz | 50 Hz in Europe, India, Australia and most of Asia and Africa; 60 Hz in North America, Brazil, South Korea, Taiwan, the Philippines and western Japan |
| Wind, handling bumps, desk knocks | Low cut below 80 Hz | Removes rumble under the voice |
| Noise on music | Spectral noise gate only | AI denoise is trained on speech and can remove instruments |
How RNNoise and the noise gate work
RNNoise, created by Jean-Marc Valin at Xiph.Org, splits each 10 ms of audio into frequency bands and lets a small recurrent neural network estimate how much of each band is voice. Bands dominated by noise are turned down, bands carrying speech are kept, and a pitch filter tidies the gaps between the voice’s harmonics. Because it was trained on speech, it is excellent on voices and can be harsh on anything else.
The spectral gate compares every short slice of the sound with the noise profile. Where a frequency is barely above the noise, it is lowered by the chosen reduction; where the wanted sound is clearly stronger, it passes untouched. Gains open instantly and close over about 60 ms, and neighbouring frequencies are smoothed, which keeps the result natural.
Honest limits
After cleaning, the Audio Normalizer brings a podcast to −16 LUFS, and the Audio Spectrogram shows exactly where the noise sat. Recording again? Check your setup with the Voice Recorder first.
- Noise removal cannot recover a voice that is buried: when the noise is as loud as the speech — a crowded café, strong wind, a loud TV — expect a cleaner but slightly robotic sound.
- Echo and room reverberation are not removed; record closer to the microphone if the room sounds hollow.
- On music, use only the spectral gate at a modest reduction (6–12 dB); RNNoise may treat instruments as noise.
- Too much reduction makes voices thin or watery — lower the Strength or the Noise reduction and compare with the preview.
Audio Noise Remover FAQ
How do I remove background noise from a voice recording?
Open the recording, keep AI speech denoise on at 100 % strength, press Preview to check it, then press Remove noise and download. Add the spectral gate for steady hiss and the hum filter for mains buzz.
Is this audio noise remover free and private?
Yes. It is free, needs no account and processes your audio on your own device — nothing is uploaded.
Can it remove noise from music?
Partly. Use the spectral noise gate with a gentle reduction for hiss or hum under music. The AI denoiser is trained on speech and can remove parts of the instruments, so switch it off for songs.
Should I choose 50 Hz or 60 Hz for hum?
Use the mains frequency of the country where the recording was made: 50 Hz in Europe, India, Australia and most of Asia and Africa; 60 Hz in the USA, Canada, Brazil and some other countries. If unsure, preview both and keep the one that removes the buzz.
Why does my voice sound robotic after noise removal?
The noise was very strong compared with the voice, or the reduction was too high. Lower the Strength slider, or use a smaller Noise reduction, and compare again.
Does it work on long recordings?
Yes, up to 60 minutes. RNNoise runs at many times real-time speed, and a progress bar with Cancel shows how far it has got.