AI Applications in Audio Source Separation and Noise Removal

A. Swantner, G. Vance
The Audio Project, Texas, United States

Keywords: ai, neural network, communications, noise, signal

We have developed an Artificial Neural Network which removes noise from audio signals. Source separation is a highly sought after tool in Audio Signal Processing, and the network we have developed uses pairs of clean and noisy audio to learn and separate noises from its signal. We've conditioned our neural net to not only remove ambient white noise and static, common in long range telecommunications, but also naturally occurring noises from voices to engines - whether it's a coffee shop or loud stadium to a passing car or a table saw. Compared to current public and private products our AI is more precise - maintaining the integrity of the underlying source, effective - removing a wide array of noises, and fast - processing on commercially available hardware solutions.