Clean background noise off a voice recording with AI, or strip hiss and mains hum from music with spectral processing. Two modes, one tool.
The noise reducer has two modes, because voice and music are genuinely different problems. Voice mode runs a speech model trained to tell a talking voice apart from the hum, hiss, and room noise around it, which is what makes it effective on a recording made somewhere imperfect. Music mode uses spectral processing and hum notching instead — no model guessing at what's noise, just targeted removal of the tones and hiss you don't want.
Use voice mode for interviews, podcasts, and calls; use music mode to clean hiss off a transfer or pull mains hum out of a guitar take. Both run on your own device, which is also why aggressive settings should be checked before you commit — noise reduction always trades some signal for silence, and pushing it hard leaves a processed, underwater quality behind. Recording cleanly in the first place beats fixing it later.
Voice mode is an AI model trained on one job: keep human speech, remove everything else. To that model, guitars, pads, drums, and reverb tails are "everything else" — run a song through it and the instruments get eaten along with the noise. That's why the tool shows a warning in Voice mode and offers a separate Music mode, which uses spectral processing (not AI) designed to lower hiss, hum, and steady background noise while leaving the actual music intact. Match the mode to the content and both work well; cross them and you'll get exactly what the warning says.
It blends the processed audio with your untouched original — 100% is the fully denoised version, 0% is the original, and the default sits at 75%, which in our testing keeps nearly all of the noise reduction while letting just enough of the original through to avoid the "pumping" or underwater artifacts aggressive denoising can cause. The blend applies live during playback, so you can find the sweet spot by ear, and the export uses exactly what you hear. If a result ever sounds worse than the original, pull the slider down — that's what it's for.
Everything runs on your own device — there's no server farm doing the work. Voice mode pushes every 16 milliseconds of audio through a neural network step by step, which on a typical laptop takes roughly half the clip's length (a 10-minute recording ≈ 5–6 minutes; slower on phones). Music mode is much faster. The progress bar shows real percentage and a time estimate, and your device stays responsive because the processing runs in a background worker. Files longer than 12 minutes are declined up front with a suggestion to split them with the Audio Cutter.
Voice mode's AI model (GTCRN, an open-source speech-enhancement network) isn't built into browsers, so the first run fetches it from our own CDN — about half a megabyte, roughly a tenth the size of a typical photo. Your browser caches it, so later runs start instantly, and once it's cached the tool works fully offline: your audio never goes anywhere in either mode, in the same way as every other tool on this site. Music mode downloads nothing at all, ever.
Music mode works on the full frequency range at your file's original sample rate and keeps stereo — quality loss beyond the intended noise removal is minimal by design, and the conservative settings cap how much any frequency can be cut. Voice mode is optimized for speech: it processes the speech band (16 kHz) and outputs mono, which is ideal for podcasts, voice memos, and interviews but is a real trade-off for anything musical — another reason songs belong in Music mode. In both modes, exporting as WAV adds no further loss; MP3/Opus exports apply normal compression.