AI Audio Denoising Online: How It Works and How to Try It Free

AI audio denoising is a model that has learned what speech looks like and uses that knowledge to decide which parts of a recording to keep. Older tools followed fixed rules: cut everything below 80 Hz, gate anything that is too quiet, pull down a few narrow bands. A model can tell the difference between a voice and a refrigerator even when they overlap in frequency, because it has heard both thousands of times.

This article explains what happens between upload and download, shows a before and after on a hard recording, and points to the fixes for specific problems like hiss and hum.

What happens to your file

Almost every speech model works on a spectrogram, a picture of how energy in each frequency band changes over time. The model produces a mask for that picture: keep this part, drop that part. A voice and a running engine may sit in the same bands, but the engine holds steady while the voice flickers with each syllable, and the mask follows the flicker.

Because the answer depends on patterns rather than rules, results vary with the material. A steady fan or a mains hum is close to ideal, because the noise is predictable. A noisy street with horns and passing cars is harder. Speech that overlaps another voice, music, or heavy reverb is hardest of all, and no model on the market solves it.

How it differs from a noise gate or an EQ

A gate mutes everything below a threshold, so the room goes silent the instant the speaker stops. That silence is what people notice as "processed". An EQ removes the same frequencies always, including the parts of the voice that live there, which is why a high pass can make a voice sound thin.

A model does neither. It removes noise while the speaker talks, keeps the room tone natural, and leaves the voice body alone. Our earlier guide compares background noise removal in Audacity against a learned model, and the difference is audible on low frequencies.

The test: rumble under a voice

This recording was made near a train, so a low rumble sits under everything the speaker says. Here is the raw file, then the same file after processing:

Measured on this file with a simple band energy check, energy below 150 Hz fell by about 9 dB while the speech band between 500 Hz and 2 kHz changed by less than a decibel. That is the shape of a good result: a large drop where the noise lives, almost nothing where the voice lives.

Running it online

The workflow is the same everywhere. Upload the file, wait for the model to run, listen, download. Browser tools need no install, and the heavy computation happens on a server, so a laptop or phone is enough. Fullband accepts common audio formats including wav, mp3, flac, ogg, and m4a, works at 48 kHz, and keeps processed files available for download for 24 hours.

Two practical notes. Fix the source when you can: a hum from a ground loop is better solved at the socket than in software, as we describe in this guide to microphone hiss and hum. And if you want to compare many models before picking one, our head to head listening test of ten denoisers covers the trade offs.

FAQ

The sample above comes from our own demo library and was recorded before any processing. Band energy was measured with ffmpeg on the original and processed files.