Only Trusted Audio Noise Remover
v3Remove background noise from speech, on your device. The length stays the same; only the noise goes.
Drop an audio file here
Voice memos, lectures, podcasts, interviews. MP3, WAV, M4A, OGG, Opus, FLAC and WebM all decode in your browser.
Listen: original vs cleaned
Cleaning engine
Strength
Strong removes more noise but can thin out voices. Medium is the safe default. Fast blends the cleaned sound with the original; Stronger raises its noise-suppression limit.
Export the cleaned audio
Check the result
DownloadProcessed on this device: 0 · Uploaded: 0
How it works
Your file is cleaned frame by frame, on your device
Your browser decodes the recording, resamples it to 48 kHz (the rate both engines were trained at), and feeds it through the engine you picked in 10-millisecond frames. Fast (RNNoise) predicts, for each frame, how much of the sound is voice and how much is noise, and keeps the voice part; the strength setting blends the cleaned sound with the original, so Light keeps more of the room and Strong keeps only what the engine is sure is voice. Stronger (DeepFilterNet3) goes further: a small neural network rebuilds the speech frequencies frame by frame, and the strength setting raises the limit on how much noise it is allowed to remove. The cleaned frames are joined back into a full recording at the same length, and the export is written at 48 kHz as MP3 or WAV.
Guide
Getting a clean result
- Start with Fast at Medium, then listen. Play the cleaned version against the original. If noise is still audible, try Strong, or switch to the Stronger engine.
- Thin or robotic voice means too much cleaning. Step the strength down one notch rather than switching engines; the voice usually comes back before the noise does.
- Music and performances: both engines are trained on speech and will partly suppress music. Do not use this tool on a song and expect the music to survive.
- Noisy rooms: steady noise (fans, air conditioning, hum) cleans up best. Sudden sounds (doors, coughs, overlapping voices) improve less, because they look like speech to the engine.
- Pair it with the Silence Remover. Clean the noise first, then cut the dead air: the Silence Remover measures loudness, and a noise-free recording gives it a cleaner floor to measure against.
- Long lectures: export MP3 at 192 kbps for sharing (about 1.4 MB per minute). WAV only if an editor needs it next.
Questions
Will it also remove music or other sounds?
Both engines are trained on speech, not music. They are good at steady background noise behind a voice: fans, hum, hiss, air conditioning, traffic murmur. Music, applause and overlapping voices get partly suppressed too, so do not use this tool on a song or a performance and expect the music to survive.
Which engine should I pick, Fast or Stronger?
Start with Fast (RNNoise). It loads in about 122 KB and handles everyday noise. Pick Stronger (DeepFilterNet3) when Fast leaves noise behind: it cleans deeper, but downloads about 23 MB of model files on first use from this site, then runs offline. Stronger is also the better pick for very noisy recordings, and the worse pick when you want the lightest possible touch.
Why does cleaning run at 48 kHz?
Both engines were trained at 48 kHz and only accept that rate, so every file is resampled to 48 kHz before cleaning and exported at 48 kHz. The length of the recording does not change, and 48 kHz is a standard rate for published audio.
Is my audio uploaded anywhere?
No. Decoding, resampling, cleaning and encoding all run in your browser, on your device. The Stronger engine's model files come from this site on first use, like any other page asset; your recordings never leave the device. Load the page, turn off your internet after the engine has loaded once, and the tool still works.
Can it handle very long recordings?
The whole recording is decoded into memory, so very long or very large files can strain a phone. Files over about 350 MB get a warning before decoding. For long lectures on a phone, split them first with the Audio Trimmer, or keep them under an hour.