How to use
- Drop or click to open an audio or video file. MP3, WAV, M4A, OGG and FLAC work, and so does the sound track of a video such as MP4.
- Set the strength. Start with 80%. Every time you change an option, the first 10 seconds are processed again.
- Press Play original and Play processed in turn to compare. If speech sounds damaged or watery, lower the strength.
- Pick a format (MP3, WAV, M4A or OGG) and press Remove noise and save. A new file such as
name_denoised.mp3is downloaded. The original is not changed.
Options
| Option | What it does | Default |
|---|---|---|
| Strength (%) | How much processed sound is mixed with the original. 0% is the original, 100% is processed only | 80% |
| Reduce quiet gaps further | Smoothly lowers parts RNNoise does not consider speech | Off |
| Amount | How much extra reduction (−6 / −12 / −18 / −24 dB) | −12 dB |
| Match loudness afterwards | Sets the result to −16 LUFS (removing noise can make everything sound quieter) | Off |
How it works
The tool runs RNNoise, a small neural-network noise suppressor, inside your browser. Audio is brought to 48 kHz and cut into frames of 480 samples (10 ms). For each frame RNNoise returns the cleaned samples and a probability that the frame is speech. Files at another sample rate are converted for processing and returned to their own rate and length afterwards. Stereo channels are processed separately.
RNNoise output comes out two frames (960 samples, 20 ms) later than the input. The tool shifts it back so it lines up exactly with the original before mixing, which is why a strength below 100% does not produce an echo. At the end of the input it feeds enough zeros to flush the last sound and then trims back to the original length.
Reduce quiet gaps further lowers the gain toward −12 dB (default) for frames whose speech probability is below 0.5. The gain does not drop in one step: it moves with a time constant of about 1.5 frames when opening and about 6 frames when closing. In a silent stretch the gain is 0.885 after 1 frame, 0.393 (about −8.1 dB) after 10 frames and about −11.8 dB after 30. Inside each frame the gain is also ramped in a straight line from the previous frame’s value, so frame boundaries do not click. For stereo, both channels share one gain taken from the higher probability, so the stereo image stays steady.
Measured results
An 8-second synthetic voice-like signal (a harmonic tone with a wobbling pitch) mixed with white noise was processed. In the first 1.5 seconds with no speech, energy fell from −31.5 dB to −76.8 dB, about 45 dB. In the speech section it went from −25.8 dB to −27.5 dB, nearly unchanged. These come from a synthetic signal; real recordings depend on the kind of noise.
At strength 80%, a sample that is 1.0 in the original and 0.2 after processing becomes 1.0 × 0.2 + 0.2 × 0.8 = 0.36. Processing 8 seconds of mono took about 1.2 seconds in node on an N100-class PC (first run included); speed in the browser depends on the device. Stereo takes twice as long because each channel is processed separately.
What it suits, and its limits
- Works well on: speech recordings (conversations, lectures, meetings, podcasts) with air conditioners, fans, keyboards or distant murmur.
- Not for: music. RNNoise is built around human voices, so instruments and backing tracks can be cut like noise. When speech and noise sit at the same pitch, speech may sound slightly thinner.
- Recordings that are already clipped, or where noise is far louder than speech, cannot be fully restored. Lower the strength to trade noise removal for a natural sound.
- Files over 30 minutes take a long time and a lot of memory. The limits are 1 hour of audio and a 500 MB file.
When you open a video, only the sound is processed and saved as an audio file; the picture is untouched. To turn the cleaned speech into text, use Speech to Text; to only even out the volume, use Normalize Audio Volume; for format changes and trimming, use Audio Editor & Converter.
Your file is processed only in your browser and is never sent to a server.