At a glance
- Accepts
- MP3, WAV, M4A, AAC, OGG, Opus, FLAC, and the audio track of a video
- Outputs
- A 16 kHz mono WAV with background noise suppressed
- Where it runs
- Entirely in your browser, on your device
- What leaves your device
- Nothing - the file never leaves your device
- Limits
- Output is 16 kHz mono, which is the model's own rate; tuned for speech, so background music is suppressed rather than preserved; roughly 6-10 dB of improvement depending on how noisy the input is
- Account
- Not required - there is no sign-up
- Price
- Free, with no watermark and no usage cap
- Made by
- Visrut
- Last updated
- 12 October 2026
This runs GTCRN, a speech enhancement model small enough to be unremarkable: 535 KB and about 48,000 parameters. It reads your recording one short frame at a time, decides which parts of each frame are speech, and attenuates the rest. Because it is so small it runs around fifteen times faster than playback on an ordinary laptop, with no download step and nothing sent to a server.
What the numbers look like
Tested against speech mixed with white noise at three levels, the improvement in signal-to-noise ratio was about 10 dB when the noise was as loud as the voice, 8 dB when the voice was slightly ahead, and 6 dB when it was already fairly clear. The pattern is the useful part: the worse the recording, the more there is to gain, and a recording that is already clean has little room to improve.
What it is not for
- Music. The model keeps speech and discards everything else, so a song comes back gutted. To separate a voice from music and keep both, use the voice and music splitter.
- Mastering. It removes noise; it does not make a voice sound better recorded. Clipping, heavy room echo and a bad microphone all survive.
- Full-band audio. Output is 16 kHz mono, because that is where the model works. Anything above 8 kHz in your original is gone.
Frequently asked questions
Is my recording uploaded anywhere?
No. The model is served from this site as a 535 KB file and runs in your browser tab. Your audio is decoded, processed and encoded locally - it never goes to a server.
How much noise does it actually remove?
Measured on speech mixed with white noise: about 10 dB of improvement when the noise is as loud as the voice, around 6 dB when the voice is already clearer. In plain terms, a noticeably cleaner recording rather than a perfectly silent one.
Why is the output 16 kHz mono?
That is the rate the model works at. Upsampling the result back to 44.1 kHz would make the file bigger without adding anything the model produced, so the download is left at the rate it was actually computed.
Will it work on music?
No. It is trained to keep speech and discard everything else, so music in the background is treated as noise and suppressed. To split a voice from music while keeping both, use the voice and music splitter instead.
What is the blend slider for?
Full suppression can sound airless, because a room with no ambience at all reads as unnatural. The slider mixes some of your original back in, so you can trade a little noise for a more natural result.
How fast is it?
Roughly fifteen times quicker than playback, so a ten-minute recording takes under a minute. The model is small enough that this holds on a laptop CPU.