ByteScope

Audio Denoise

Drop a voice recording — or a video, its audio track is used — and hear the first 15 seconds cleaned within moments. Flip between Original and Denoised without losing your place, set how much noise to remove, then download a WAV of the same length. RNNoise runs as WebAssembly in a Web Worker; the file never leaves your computer.

Your files never leave your browser — all processing is local.

Drop an audio or video file, or click to choose

One file at a time: MP3, M4A, WAV, OGG, FLAC, or a video's sound track

Before you start

  • It is trained on speech: very good against air conditioning, fans and street noise, but it damages music. Don't use it on songs.
  • It does not remove echo (reverb), repair clipped audio, or bring a distant voice closer.
  • Set too strong, voices sound watery or choppy. Turn the strength slider down.
  • Browser recordings may already have noise suppression on, so a second pass helps less.
  • This version handles up to 30 minutes mono or 15 minutes stereo per file.
  • Your file is processed in this tab and never uploaded.
  • The output is a 48 kHz, 16-bit WAV with the same channels as the input.

About this tool

A recording made in a real room carries the room with it: the laptop fan, the air conditioner, traffic behind the window, the hum of a keyboard or a fridge. This page removes that kind of steady background noise from speech using RNNoise, the recurrent-neural-network speech denoiser published by Jean-Marc Valin in 2018, compiled to WebAssembly and run in a Web Worker inside your browser. There is no server behind the page: the file is decoded, processed and written back on your own machine, which is also why there is no queue, no account and no size-based pricing. It accepts MP3, M4A, WAV, OGG and FLAC, and video files such as MP4 or MOV, from which only the audio track is read. Every channel is kept and processed on its own, so a stereo recording stays stereo.

Hear it before you wait for it

The first 15 seconds are processed first and become playable almost immediately. An Original / Denoised switch lets you compare while the rest of the file is still being worked on, and it keeps the playback position, so you can jump back and forth on the same sentence. The strength slider runs from 0 to 100 %, with a default of 85 %: leaving a little of the original noise in place sounds more natural than silence between words, which is where an over-cleaned recording starts to feel artificial. Changing the strength is a mix between the two signals, not a new pass, so it takes effect at once and never reprocesses the file.

What comes out

The result is a 48 kHz, 16-bit WAV with exactly the same length as the original and the same number of channels. RNNoise works natively at 48 kHz in 10 ms frames, so the input is resampled to that rate before processing; a 44.1 kHz MP3 therefore comes back as a 48 kHz WAV. WAV is uncompressed, which keeps the cleaned voice free of a second lossy encoding — about 5.8 MB per minute of mono, double for stereo. The whole file is held in memory while it is processed, so the current limit is 30 minutes of mono or 15 minutes of stereo; longer recordings should be cut into parts with an audio editor first.

What it will not fix

RNNoise was trained on speech, and only speech. Music is treated as noise and gets damaged — do not run it on a song. It removes steady noise, not room: echo and reverb stay, because to the model they are part of the voice. It cannot repair clipping from a microphone that was too hot, and it cannot bring a voice that was far from the microphone closer. Pushed too far, the output turns watery or choppy, especially on quiet consonants; if that happens, lower the strength rather than looking for a different file. Recordings made by video-call software or in the browser are often noise-suppressed already, and a second pass has correspondingly less to remove.

Where it fits

The page is built for the step before publishing or transcribing: an interview recorded in a café, a lecture with the projector fan running, a voice memo taken on a train. Clean the audio here, then transcribe it with Video Subtitles, or pull the audio out of a video first with Extract Audio from Video. For music files — finding the tempo and key of a track — BPM & Key Finder is the right page, not this one.

Frequently asked questions

Is my recording uploaded anywhere?

No. The file is decoded by your browser, processed by RNNoise compiled to WebAssembly inside a Web Worker, and written to a WAV in the same tab. Nothing is sent to a server — there is none behind the page — and the page works the same way whether the recording is a public podcast or a confidential interview. Nothing about the file — its name, length or content — leaves the browser.

Why does the voice sound watery or robotic?

Because the strength is too high for this recording. RNNoise estimates how much of each frequency band is speech and attenuates the rest; when the noise and the voice overlap — quiet consonants, breathy vowels, a very noisy room — a strong setting removes parts of the voice too, which sounds like underwater speech or small dropouts. Pull the slider down from the default 85 % until the artefacts go away. A little remaining noise sounds far more natural than a damaged voice.

Can I use it on music or a song?

You can, but you will not like the result. The model was trained to keep speech and remove everything else, so instruments, reverb tails and harmonies are treated as noise and get carved up. It is meant for spoken recordings: interviews, lectures, podcasts, voice memos, meeting recordings. If you want to analyse a song instead — its BPM and key — use BPM & Key Finder.

Does it remove echo or reverb?

No. Echo and reverb are reflections of the voice itself, and to the model they are part of the speech, not noise. The same goes for a voice that was recorded too far from the microphone: the page cannot bring it closer. What it removes is steady background noise — fans, air conditioning, traffic, electrical hum, keyboard clatter — which is a different problem from room acoustics.

Why is the output a 48 kHz WAV and not an MP3?

48 kHz because that is the rate RNNoise works at — it processes 10 ms frames at 48 kHz, so the input is resampled to that rate before denoising. WAV because it is uncompressed: the cleaned voice is not pushed through a second lossy encoder, and every editor, transcription tool and video program can open it. The file is larger than an MP3, about 5.8 MB per minute of mono; if you need a smaller file, convert the WAV afterwards.

Why is there a 30-minute limit?

Because the entire decoded recording is held in the browser's memory while it is processed, and 48 kHz audio adds up quickly: 30 minutes of mono or 15 minutes of stereo is roughly where it stays safe on an ordinary laptop or phone. Longer recordings should be cut into parts first. The limit is about memory, not time — a longer file would not take dramatically longer, it would risk the tab running out of memory.

My recording already went through a meeting app. Will this help?

Less than you might hope. Video-call software and browser-based recorders apply their own noise suppression before saving, so the steady noise RNNoise is good at has largely been removed already, and what is left is usually artefacts of that first pass, which no second pass can undo. Try the 15-second preview: if Original and Denoised sound nearly the same, the recording did not need it.

Related tools