Remove background noise without upload
Remove hiss, fan hum, street noise and room reverb from voice recordings – voice memos, interviews, lectures, podcasts or the sound of a video. The AI (DeepFilterNet3) runs on your device; recording, model and result do not leave the browser. Before and after can be compared at the same position.
Runs on the processor – about 10 times faster than the recording is long
DeepFilterNet3 (MIT) · 9 MB, served by mytoolio
Pick a recording on the left – hiss, fans, street noise and background reverb are filtered out.
How it works
Choose a recording
Drop an audio or video file on the field or pick it. Processing starts right away – about ten times faster than the recording is long.
Compare & adjust
Play the recording and switch between “Original” and “Denoised”. Use the “Strength” slider to mix a little room sound back in if the voice seems too dry.
Save as WAV
Download the result as a WAV file (48 kHz, mono) – for editing, uploading or transcribing.
Frequently asked questions
Is my recording uploaded?+
No. The recording is processed in your browser and does not leave your device. The AI model and the runtime are served by mytoolio itself – no other provider is involved.
Which noises are removed?+
Steady hiss, fans and air conditioning, hum, street and wind noise, keyboard clatter, background chatter and part of the room reverb. The model is trained on speech: it treats music as noise and attenuates it – the tool is not meant for singing or instruments.
Which files can I use?+
Anything your browser can play: MP3, WAV, M4A/AAC, OGG/Opus, FLAC, WebM and the audio track of MP4 and MOV videos – up to 30 minutes long. Stereo stays stereo (up to 15 minutes; both channels are processed separately, which takes twice as long); switch off “Keep stereo” to mix down to mono. The result has 48 kHz – as WAV (16 bit), Opus or, for videos, as MP4 with the new sound.
The voice sounds muffled or artificial afterwards – what can I do?+
Set the “Strength” slider to about 70 to 85%: a little of the original room sound stays and the voice sounds more natural. With very loud noise or a distant microphone the model can no longer separate parts of the voice cleanly; a lower strength helps there too.
How long does it take?+
On a current computer about a tenth of the recording’s length – ten minutes of audio take roughly one minute; longer on a phone. The waveform is coloured in while it works; you can already listen to the original meanwhile.
Does it work offline?+
Yes. On the first visit the model (9 MB) and the runtime (14 MB) are downloaded and cached by the browser; after that the tool also works without an internet connection.
Can I save the result smaller than WAV?+
Do I get my video back with the cleaned sound?+
Can I record straight from the microphone?+
Noise reduction with DeepFilterNet3 by Hendrik Schröter et al. (MIT licence) on onnxruntime-web – entirely on your device. The result is a computed estimate of the speech; quiet sounds or sounds masked by noise can be lost.