核心要点 (BLUF)： Noise suppression removes steady background noise from a recording — air-conditioner hum, microphone hiss, street rumble — right in your browser. Upload audio or a video with sound, press the button and get a cleaned track: you can play it on the page and download it as a WAV file. The file is never uploaded, everything runs on your device, no tokens are spent and there is no paid API here.
It cleans a recording from constant background noise. The input is audio or a video with sound, the output is the same recording without hum, hiss and rumble. It helps when an interview was recorded in a cafe, a lecture in an air-conditioned hall, or a voice message by a window on a busy street.
A cleaned track is useful for editing too: a podcast, a channel video, a course or a webinar recording sounds tighter when the background does not compete with the voice.
The recording is split into short frames of 480 samples — exactly 10 milliseconds at 48 kHz. For every frame the model estimates where the voice sits in the spectrum and where the noise is steady, then attenuates the noisy part. The frames run back to back, so speech does not break up or drift in loudness.
The result is delivered as a 48 kHz mono WAV file — you can play it on the page and download it.
The recording never leaves the browser: the model is loaded once and the computation runs on your processor. The tool has no paid API, so no tokens are spent and the source file is not sent to a server.
The model module is about 110 KB and is cached by the browser — on your next visit it comes from the cache instead of the network.
It handles steady background best: hum, hiss, fan noise, street rumble. Speech stays intelligible and pauses become quieter.
It does not clean music: if music plays in the recording, the tool will attenuate it together with the noise — the section has a separate tool for splitting tracks. Heavy reverb and echo are not removed either: that is a different job.
Upload the whole recording rather than a short cut: the model needs a few seconds to understand what counts as noise in this particular recording.
If the recording is quiet, raise the level in an editor first and clean afterwards: on very quiet material noise and voice are harder to tell apart. The finished file downloads as WAV, ready for editing.
Yes. The model and its weights are loaded into the browser and the computation runs on your device. Neither the recording nor the result is sent anywhere — the tool has no paid API at all.
Nothing. No tokens are spent on this tool: it does not call paid services, everything runs on your device.
Audio and video with sound: MP3, WAV, FLAC, OGG, M4A, MP4, MOV, WebM and other common formats. From a video the tool takes the audio track.
The model works with a single channel at that rate — this way it separates voice from steady background more accurately. For speech and podcasts that is enough.
No. The tool is designed for speech: it will attenuate music together with the noise. The section has a separate tool for splitting a track into vocals and instrumental.
No. The model is small and runs on the processor, without a graphics card. On a weak device a long recording simply takes more time.