Remove Noise From Audio

Drop the file above, mark a second of pure background noise with the handles, and press Remove noise. Then compare Original against Cleaned before downloading. There is no sign-up, no watermark and no file size limit, and the recording is never uploaded — the processing runs on your own machine.

No. It is decoded and processed entirely in your browser, and nothing is sent to any server. This matters here more than almost anywhere else: the recordings people clean up are interviews, lectures, medical appointments, therapy sessions and confidential meetings.

Steady, continuous noise — tape hiss, computer fans, air conditioning, traffic rumble, room tone, mains hum, general static. These have a roughly constant spectrum, which is what makes them separable from speech and music. It will not remove a door slamming, a dog barking, a chair scraping or someone coughing: those are one-off events that look like signal rather than noise, and no tool that works this way can tell them apart. Anything claiming otherwise in a browser is overselling.

Spectral gating, the same approach Audacity uses. The recording is cut into short overlapping frames and each is transformed into a spectrum. From the noise you marked, the tool learns how loud each narrow band of frequency is when only noise is present. Then, in every frame, each band is compared against that measurement: louder than the noise and it is kept, quieter and it is turned down by the amount you set. The result is transformed back into sound.

Yes, and it is the fastest way to get the settings right. Each stage keeps the sound it took out and shows it separately: rumble, mains hum, background noise, and silenced pauses, each with its own level, waveform and play button. Play them. If a layer is nothing but steady hiss, the reduction is doing exactly its job. If you can hear words, breaths or music in it, that material is being removed from your recording and the strength should come down. Judging the cleaned track alone will never tell you this, because what is missing from it is by definition not there to hear. The layers are exact — the cleaned track plus every layer adds back up to the original sample for sample.

That artefact is called musical noise, and it comes from each frequency band being switched on and off independently between frames. The residual noise stops being a smooth hiss and becomes a shifting pattern of isolated tones, which the ear finds far more objectionable than the hiss it replaced. The Smoothing controls prevent it by blurring the decision across neighbouring bands and neighbouring frames, so nothing flickers. If your result sounds watery, raise Smoothing first and lower Strength second.

No, but it is better. Without one, the tool estimates the noise floor from the quietest quarter of each frequency band across the whole recording, which works because in most recordings most frames of any given band contain only noise. Marking a genuine second of silence is more accurate, and it is the single biggest thing you can do to improve the result.

It sets how far the noise is pushed down, in decibels. 12 dB is gentle and nearly always safe. 18 dB, the default, handles most recordings. Past about 24 dB the processing usually starts to be audible on the voice itself — a slight hollowness or a loss of breath and air. More is not better; the aim is the least reduction that makes the recording comfortable to listen to.

It decides how far above the measured noise floor a sound has to be before it counts as something you want to keep. Raise it if hiss is surviving the process. Lower it if quiet consonants, the ends of words, or soft passages are being eaten. It is the control to reach for when Strength alone is not getting you there.

Use the Mains hum setting rather than the main noise reduction — a notch filter is far more surgical for this. Choose 50 Hz if the recording was made in Europe, Asia, Africa or Australia, and 60 Hz in the Americas or Japan. Then raise the harmonics setting: a mains hum is never only its fundamental, and it is the harmonics stacked above it that make the sound buzz rather than hum, so notching only the fundamental removes surprisingly little of the problem.

For speech, almost always. Traffic, air conditioning, footsteps and microphone handling noise all sit below about 80 Hz, and the human voice has essentially nothing down there — so removing it costs you nothing and cleans up a lot. For music, be careful: bass guitar, kick drum and the low notes of a piano live in exactly that range, so turn it down or off.

Yes — drop the video file in directly and it will decode and clean the audio track. What you get back is an audio file, not a video with the soundtrack replaced. To put it back you would need a video editor.

Noise reduction is inherently a trade. It cannot know exactly which parts of the sound are noise, so every setting is a balance between removing background and leaving the wanted sound untouched. Used gently it is close to free; pushed hard it audibly thins the voice. This is exactly why the A/B comparison is built in rather than tucked away — judge it by ear against the original, not by how much the number moved.

Because it is real signal processing rather than a volume adjustment. Every frame of audio — and there are roughly ninety per second — is transformed into the frequency domain, examined band by band, and transformed back. A few minutes of audio means tens of thousands of transforms. It still runs many times faster than the recording plays.

None imposed, since nothing is uploaded. The practical limit is memory: decoding turns compressed audio into raw samples, and the analysis holds several arrays of that size at once. An hour of audio is fine on a normal computer; phones have far less room, so trim long recordings first.