Audio Noise Reduction (WebGPU)
Load an audio file and this page quiets the background hiss between words with a noise gate and a light smoothing filter, using WebGPU where your browser has it and plain JavaScript where it does not. Play the original and the result side by side. The audio never leaves your browser.
Drop an audio file (WAV, MP3, etc.) here,
or click to choose one
About this tool
The tool listens for the quietest stretches of your recording to estimate the noise floor, then turns those stretches down with a noise gate. A short moving-average filter then takes the edge off high-frequency hiss. The smoothing step runs as a WebGPU compute shader when your browser supports it, and as the same algorithm in JavaScript when it does not or when anything goes wrong.
This is a simple gate plus smoothing, not an AI noise-removal model, so it cannot separate a voice from noise that happens at the same time. Everything runs in your browser and the file is not uploaded. To keep processing quick, only the first 3 minutes are processed.
How it works
- Noise floor: the recording is cut into 20 ms windows and the RMS level of each is measured. The 20th-percentile level is taken as the noise floor.
- Gate: windows at or below 1.5 times the noise floor are turned down. At strength 100 they are reduced to about 10% of their level (−20 dB); at strength 0 nothing changes. The gain is interpolated between windows so it does not click.
- Smoothing: a moving average with a radius of 0 to 3 samples (strength 0 to 100) is applied to every channel. At low strength it is off.
- Output: 16-bit PCM WAV at the sample rate the browser decoded the file to (usually 44.1 or 48 kHz), with the same number of channels. Each channel is gated on its own level.
FAQ
Q. Can it remove a steady hum or a fan?
A. Only partly. A gate turns down the quiet parts between sounds, but it cannot remove noise that continues underneath speech. Use a spectral noise-reduction tool for that.
Q. Why does speech sound slightly dull at high strength?
A. The moving average is a low-pass filter, so it also softens the highest frequencies of the voice. Lower the strength until the hiss is acceptable.
Q. Does WebGPU change the result?
A. No. Both paths run the same algorithm, so the output should be the same apart from tiny floating-point differences. WebGPU only changes how fast the smoothing step runs.
Q. Why only 3 minutes?
A. Longer files take a long time to process in the browser and use a lot of memory. Trim the file first if you need a different part.