Remove Background Noise from Audio

How to use

1. Give it the noisy recording under Audio file. Speech is what the model was trained on.
2. Play the original once through the little player that appears, so you know what you are comparing against later.
3. Choose the output: M4A or Opus if you are sending it to someone, WAV or FLAC if the take is going into an edit.
4. Press Remove noise. The first run pulls the three model files, roughly 8.6 MB, and keeps them cached from then on.
5. Leave the tab open while the bar moves through decoding, analysis, the model itself, then encoding. Cancel stops the run if you change your mind.
6. Listen to the cleaned version in the player that appears, then save it with Download.

About this tool

Background noise is what turns an otherwise usable recording into one nobody wants to sit through. A fridge two rooms away. An air conditioner over a call. The fan in the laptop you are recording on. To remove background noise from audio here, hand the file to DeepFilterNet, a speech-enhancement model that lifts the voice forward and presses everything behind it down.

The model arrives as three ONNX files, about 8.6 MB between them, fetched the first time you run it and cached by your browser after that. It works at 48 kHz on 20 ms frames stepped every 10 ms, and it expects a single speech channel, so whatever you give it is mixed down to mono at that rate before anything else happens. Steady noise is where it is strongest: mains hum, tape hiss, fan drone, the whir of a hard drive. Keyboard taps and passing traffic drop a long way back without disappearing.

Inference runs in a worker on one thread and the whole recording goes through in a single pass, so a long recording takes a while and the progress bar is worth watching. Cancel stops it cleanly. When it finishes, a player appears above the Download button, and playing that against the original is the only honest way to judge the result. The saved file keeps your original name with -cleaned on the end.

A six-minute voice memo recorded next to the fridge

The memo was taken on a phone in a kitchen: stereo, 44.1 kHz, with a compressor humming behind every sentence. Drop it in and the player under the field lets you hear the original. Press Remove noise and the bar reaches 15 percent as the file is resampled to 48 kHz mono, then 24 percent once it has been cut into 36,002 overlapping frames of 20 milliseconds, one every 10 milliseconds. The three model files come down on the first run only, and the bar sits at 68 percent when the encoder and both decoders have finished their pass.

The model works on two layers of that spectrum. A 32-band mask, shaped like the ear's own frequency resolution, turns the hum down across the whole range up to 24 kHz, and a five-tap filter refines the 96 lowest bins, 50 Hz apart, up to 4,750 Hz, which is where the voice lives. The rebuilt signal is trimmed by the 10 milliseconds the analysis delayed it, so it lines up sample for sample with what went in. Keep M4A selected and the cleaned take saves as kitchen-memo-cleaned.m4a, six minutes long, one channel.

Good to know

• It is a speech model, and everything else follows from that. Music run through it loses texture the model reads as noise. The output is always 48 kHz mono, so a stereo source returns single-channel. Abrupt sounds such as a door slam mid-sentence or two people talking over each other are left largely alone, where steady hum is what it takes out.
• The whole file is analysed at once rather than streamed, and the spectrum is the expensive part: the frames for a six-minute recording occupy about 139 MB, and the enhanced copy the model writes back is the same size again. A two-hour recording asks for more memory than a browser tab is usually allowed. Cut a long session into parts with the Audio Cutter and clean them one at a time.
• Nothing about the model is adjustable. There is no strength slider and no noise profile to sample, so a recording that comes back too processed can only be helped at the source, by re-recording closer to the microphone.