Updated October 2026 · AI audio explained
Neural noise reduction uses a trained network to tell speech from background sound. The network has learned from many examples, and it uses that learning to reduce the parts it judges to be noise. It works well on everyday recordings, but it is an estimate, so it can remove too much or leave some noise behind.
The words “neural network” can sound like a mystery, but the idea behind neural noise reduction is simple to explain. The tool has learned what speech tends to look like, and what noise tends to look like, from many examples. It then applies that learning to your recording.
During training, the network is shown many pairs of clean speech and speech mixed with noise. It learns to predict the clean speech from the noisy version. Over many examples, it builds up a sense of which sounds are voice and which are background.
When you use the tool, it looks at short slices of your audio and estimates how much of each slice is speech. It then reduces the parts it judges to be noise. Because it works on patterns, it can handle many kinds of everyday noise, such as fans, traffic and room hum.
The network makes estimates, and every estimate can be wrong. It can remove part of the voice, leave a little noise, or treat a sound that is part of the performance as background. This is why the setting matters, and why listening is essential.
It struggles with distant voices, clipped audio and several people talking at once, as covered in the limits post. A good recording is still the best input.
Explaining the method does not tell you how a particular product performs. Test any tool on your own recordings, and judge the result by ear, on the devices your listeners use.
Record voice-overs with live noise removal, or clean an existing file. Runs in your browser, nothing is uploaded.
No. It estimates speech and noise from patterns in the sound.
Different. A filter removes fixed frequencies, while a network adapts to the sound it hears.
The estimate removed some of the voice along with the noise. Use a lighter setting.
No. Listening to the result and comparing with the original is enough to judge it.
