Remove background noise from audio

How Audio Denoiser works

The file is read with a FileReader, rewritten in memory and offered as a download. This is why large PDFs are limited by your device's RAM rather than by an upload size limit. A 200 MB scan never crosses the network, so the slowest part is your own disk.

That is the whole architecture: static files, one script, no upload endpoint to leak or log. Close the tab and the work is gone.

Noise is steady; speech is not

A fan, a hum, a microphone's self-noise: all of them have a spectrum that barely changes from second to second. A voice changes constantly — syllables, pauses, plosives. That difference is the entire basis of noise reduction, and it is why the tool can estimate the noise without you recording a sample of it first.

The signal is cut into 1024-sample frames with a quarter-frame hop, each frame windowed and transformed into the frequency domain. Frames are ranked by total energy, and the quietest 15% are averaged into a noise profile. If someone is speaking continuously with no gaps, that profile is contaminated by speech and the result sounds over-processed — lower the strength in that case.

Subtraction, and the artefacts it creates

For each frame, the noise magnitude is multiplied by a strength factor and subtracted from the frame's magnitude, with a floor that keeps a small fraction of the original so the signal does not collapse to silence. The phase is left untouched, because phase errors are far more audible than magnitude errors.

The characteristic artefact is “musical noise”: isolated bins survive subtraction and ring like tiny random tones. It is why professional tools use spectral smoothing and psychoacoustic masking. Here the voice floor slider is the control — raising it keeps more of the original signal and suppresses the ringing, at the cost of leaving some hiss.

What the two sliders do

Strength is the oversubtraction factor: how many times the estimated noise spectrum is removed. Around 2 is a good start for a fan or air-conditioning; 3 or 4 for a genuinely noisy room, with more artefacts. The voice floor is the minimum fraction of the original magnitude that always survives, which keeps the result from sounding underwater.

The honest ceiling: this removes stationary noise. It will not remove a dog barking, a keyboard, or traffic passing — those change too fast to be modelled by an average. For that class of problem, a trained speech-enhancement model is the right tool, and the paid services are charging for exactly that.

Ten minutes, and why WAV

The maths is per frame and per sample, in plain JavaScript: a ten-minute stereo file is roughly 60,000 frames and about 26 million samples per channel. Beyond that the tab starts to feel frozen, so the tool refuses rather than appearing to hang.

The output is 16-bit PCM WAV because a browser can write that with no codec at all. MP3 would need a licensed encoder shipped to the page, which is the same reason the audio trimmer returns WAV.

Good to know

• Works offline once loaded, including on flights and in locked-down networks.
• No file size limit beyond your device's own memory.
• Free to use, with no account and no usage counter.
• Results are produced locally, so they are not logged or shared with anyone.

Questions