● a real neural denoiser (RNNoise), running entirely in your browser
AI Noise Remover & Dialogue Booster
Strip wind rumble, hiss, hum, and background noise out of an audio clip using RNNoise — a small neural network, not an EQ preset — then push voices forward with a dialogue boost that targets the model's own speech-detection signal, and hand-mute anything it can't fix automatically, like a single seagull call. Nothing is uploaded anywhere.
Drop an audio file here, or click to choose one
Click-drag on the waveform to select a moment to mute or attenuate.
How it works
Your file is decoded with the Web Audio API, downmixed to mono and resampled to 48kHz (what the model expects), then run 10 milliseconds at a time through RNNoise — a small recurrent neural network compiled to WebAssembly — which was trained to separate speech from background noise. The mix slider blends the cleaned signal back with the original so you can dial back how aggressive it is. The dialogue boost slider reuses RNNoise's own per-frame voice-probability output to add extra gain specifically on the frames it judges speech-like, then runs the result through a presence EQ (a lift around 3kHz, where speech intelligibility lives) and a compressor/limiter so voices sit louder and more forward without clipping. For a single unwanted sound the model won't target on its own — a seagull, a cough, a door — click-drag over it on the waveform and mute or attenuate just that stretch.
FAQ
Is this actually AI, or just an EQ filter?
It's a real small neural network called RNNoise, compiled to WebAssembly and run frame-by-frame directly in your browser — the same style of model used in production by tools like Discord's and Jitsi's built-in noise suppression. It's not an EQ preset.
Will this remove a specific one-off sound, like a seagull call or a door slam?
Not automatically. RNNoise is very good at continuous background noise — wind rumble, hiss, hum, fan/AC noise, crowd murmur — but it isn't designed to detect and cut a single distinct event. For that, use the waveform above: click-drag over the moment you want gone, then mute or attenuate just that region.
How does the dialogue boost actually work?
RNNoise returns a voice-probability estimate for every 10ms frame it processes (that's how it decides what to suppress) — the boost slider reuses that same signal as a gain map, so frames it's confident are speech get pushed up more than frames it isn't. That's combined with a speech-presence EQ lift and a compressor/limiter so the result gets louder without distorting. It's not a flat volume increase across the whole clip.
Why does removing noise also make dialogue sound clearer?
Suppressing everything that isn't speech already makes the remaining speech more prominent on its own — the dialogue boost slider goes a step further and actively pushes the voice up, rather than only removing what's around it.
Does this work on video files?
Not yet — this handles audio files only (wav, mp3, m4a, and anything your browser can decode). If you have a video, extract the audio track first, clean it here, then remux it back.
Free, unlimited, no account. If this saved you a paid tool, a small thanks is appreciated.