This noise remover cleans up voice recordings, podcasts, interviews and video audio in your browser. Drop one file or a folder, pick the RNNoise neural model (a 4.8 MB one-time download that's excellent on speech) or the built-in spectral noise gate, and download clean WAV files with before-and-after players to compare. Nothing is uploaded.
| Recording | Best method |
|---|---|
| Voice with fan or traffic | RNNoise |
| Voice with steady hiss | Either |
| Music or instruments | Noise gate, low strength |
| Very noisy phone call | RNNoise, mix 0.8 |
The gate learns the noise from quiet moments and turns down the frequencies where it sits; RNNoise uses a small neural network trained to separate voice from noise.
The gate does, gently. RNNoise is tuned for speech and can dull music.
Anything your browser can decode, including MP3, WAV, M4A, OGG and video files; output is WAV.
No. Files and text are processed in your browser. Only the model itself is downloaded, once, when you ask for it.
Yes. The Downloaded models panel lists what this tool stored and deletes it after a warning; you'd need to download it again to use it.
Audio is decoded by the browser at 48 kHz and mixed to mono. The built-in method is a spectral gate: a 2048-point STFT with 75% overlap, a noise profile averaged from the quietest 10% of frames (or a chosen time range), and a per-bin gain of (SNR − 1) ÷ SNR limited to a maximum reduction and smoothed over neighbouring bins and frames before overlap-add resynthesis. The neural method runs Xiph's RNNoise, compiled to WebAssembly, on 480-sample frames; it is downloaded from jsDelivr on request, checked against a SHA-384 fingerprint and kept in Cache Storage until cleared. A wet/dry mix and peak normalisation follow, and results are saved as WAV.