AI Audio Noise Remover

RNNoise is a small recurrent neural network trained on thousands of hours of speech mixed with real-world noise. It processes audio in short frames of about 10 milliseconds, estimating for each frequency band how much energy is voice and how much is noise, then suppressing the noisy bands. Because it is a speech model, it preserves formants and consonants — the parts that make words intelligible — while pushing steady background sounds down 15 to 25 dB. The whole network compiles to roughly 1.5 MB of WebAssembly, small enough to run in real time on a phone.

Denoising works best on a decent recording, so: 1. Record in the quietest room available — a closet full of clothes beats a big empty “treated” room. 2. Keep the mic 4–6 inches from your mouth; every doubling of distance roughly halves the voice-to-noise ratio. 3. Use a directional mic if you have one. 4. Record WAV at 44.1 kHz or higher — heavy MP3 compression bakes in artifacts no denoiser can fix. 5. Leave two seconds of room tone at the start so you can A/B the before and after honestly.

A 42-minute podcast recorded next to a humming refrigerator. In the original, voice peaks hit −12 dB while the fridge hum sits at −38 dB — clearly audible in every pause. After processing, the hum drops to about −58 dB, a 20 dB reduction that makes it effectively inaudible, while voice clarity is unchanged. The 8.4 MB MP3 goes in; a 440 MB WAV comes out (uncompressed 42-minute stereo). Compress the result to 128 kbps MP3 — about 40 MB — for publishing.

RNNoise is trained for speech: music, singing and sound effects get distorted, so keep it to voice recordings. Intermittent sounds — a door slam, a dog bark, someone talking over you — are only partially reduced by any denoiser. Heavy room echo stays; the model removes additive noise, not reflections. Output is uncompressed WAV, so long recordings produce large downloads, and files over about 100 MB may strain mobile browsers.