NoCloud Media

Audio tool

Audio Denoiser — Remove Background Noise

Pick an FFT-based strength or the RNN voice model. Same family Krisp and Adobe Podcast Enhance use. Your audio never leaves your browser.

How it works

  1. 1

    Drop your audio file

    MP3, WAV, OGG, M4A, AAC, or FLAC. The file stays on your device.

  2. 2

    Pick a denoiser

    Light / Medium / Strong run FFmpeg's afftdn filter (FFT-based; works on voice, music, and ambience). RNN voice runs FFmpeg's arnndn filter with the rnnoise model — voice-trained, cleanest on speech under heavy background.

  3. 3

    Process and download

    FFmpeg.wasm runs the denoiser in your browser. The model file (~300 KB for RNN voice) is lazy-loaded the first time you pick that mode and cached after — subsequent runs are instant.

Why use Audio denoiser?

Strip HVAC drone, laptop fan whir, computer hum, and street rumble from voice recordings without re-recording.

Same family of denoiser Krisp ($12/mo) and Adobe Podcast Enhance (cloud + signup) charge for. Here it runs on your machine, on your file, with nothing uploaded.

Two algorithms in one tool — FFT-based for any content (voice, music, ambient) and RNN-based purpose-trained for speech. Pick whichever fits.

Lossless option — pick WAV or FLAC output to keep the denoised audio bit-perfect for further editing.

Private — your audio never touches our servers, which matters when the recording is personal or pre-release.

Common use cases

  • Clean up a Zoom recording with laptop-fan whir before editing into a podcast
  • Strip HVAC drone from a remote interview that you can't re-record
  • Take street rumble out of a voice memo recorded outside
  • Suppress mic-stand hum in a music demo before sending to a producer
  • Remove ambient air-conditioner noise from a YouTube voiceover recorded at home
  • Clean up a voice note recorded in a busy café before uploading to a podcast (RNN voice, heavy background)
  • Salvage a phone interview recorded in a windy environment
  • Polish a screen-recording's narration for a tutorial without re-recording

About MP3 and MP3

Two algorithm families ship in one tool. **Light / Medium / Strong** use FFmpeg's `afftdn` filter — a frequency-domain Wiener-style denoiser that estimates the noise floor per frequency band and subtracts it from the signal. Same family iZotope RX and Adobe Audition's Adaptive Noise Reduction use. Light preserves room tone (~6 dB cut), Medium is FFmpeg's documented default (~12 dB), Strong is aggressive (~24 dB). The afftdn variants work on any content — speech, music, or ambient field recordings. **RNN voice** uses `arnndn` with the rnnoise pretrained model (xiph.org's open-source neural denoiser, BSD-3-Clause licensed, ~300 KB). The network classifies each short window's likelihood of being voice and attenuates the rest. Bench-measured (`npm run bench:denoise-comparison`): RNN voice produces output noticeably closer to the clean reference than any afftdn strength when the noise is louder than the voice (-10 dB SNR — the worst case, like a phone mic in a windy café). At more moderate noise levels (voice clearly audible above the hum), RNN and afftdn perform comparably, so the choice is content-driven: speech with intentional background sound (room tone, music) → afftdn; speech under heavy background → RNN voice. Voice-trained RNN means it'll mangle music and ambient soundscapes; pick an FFT variant for those. The rnnoise model file is lazy-loaded on first use and cached — subsequent runs are instant.

Frequently asked questions

Is my audio uploaded to a server?
No. NoCloud Media runs the denoiser entirely in your browser using WebAssembly. Your file never leaves this tab. Krisp and Adobe Podcast Enhance both upload your audio to a server; this doesn't.
Which strength should I pick?
For voice with HVAC / fan / mild crowd noise: try **RNN voice** first — it's purpose-trained for speech. For music or recordings with intentional ambience (room tone, atmospheric pads): pick a **Light / Medium / Strong** afftdn variant — those preserve the surrounding sound rather than treating it as noise. Among the FFT variants: Medium is FFmpeg's documented default; Light keeps room tone; Strong is for heavy background but voice may sound a touch processed.
How does the RNN voice option compare to Krisp / Adobe Podcast Enhance?
Same family of algorithm — a small neural network trained on speech-vs-non-speech. Krisp ships its own proprietary model; Adobe ships theirs; we ship the open-source rnnoise model from xiph.org (BSD-3-Clause). On heavy background noise the bench (`npm run bench:denoise-comparison`) shows RNN voice noticeably outperforms FFT-based denoising. For typical podcast recording conditions (voice clearly above the noise floor), all three produce subjectively comparable output. The differentiator here is privacy + price: this runs locally with nothing uploaded, no signup, free.
Will RNN voice damage music or ambient sounds?
Yes. The rnnoise model was trained to recognize speech and attenuate everything else. On music it'll mangle the audio (treating instruments as 'noise'). On ambient field recordings (rain, wind, birdsong) it'll silence the very sound you wanted. For non-speech content, pick Light / Medium / Strong instead — those use afftdn, which is content-agnostic.
Why does the RNN voice option take longer the first time?
The rnnoise model file (~300 KB) downloads on first use and caches in your browser. Subsequent runs skip the download. The model is BSD-3-Clause licensed from github.com/GregorR/rnnoise-models, served from this site — there's no third-party fetch, nothing tracked.
Can I combine denoising with loudness normalization?
Not in a single click here — this page locks the widget to denoise mode. Run the audio through the **volume changer** for loudness, then bring it here for denoise (or vice versa). For a single-pass workflow, use the volume changer's three-mode picker.
Is the output re-encoded?
Yes — denoising modifies every sample, which requires re-encoding. Lossy formats (MP3, OGG, M4A) use a 192 kbps default. Pick WAV or FLAC for lossless output if you'll do more editing downstream.
What's the maximum file size?
It depends on your browser's available memory. Denoising is memory-heavy because it holds frequency-domain data in RAM. Files under 200 MB are reliable; very large files may run out of memory.
Which browsers are supported?
Chrome, Edge, Firefox, and Safari 15+. We require WebAssembly and SharedArrayBuffer, both standard in modern browsers.

Related tools