On Filter Generalization for Music Bandwidth Extension Using Deep Neural Networks

In this paper, we address a sub-topic of the broad domain of audioenhancement, namely musical audio bandwidth extension. We formulate thebandwidth extension problem using deep neural networks, where a band-limitedsignal is provided as input to the network, with the goal of reconstructing afull-bandwidth output. Our main contribution centers on the impact of thechoice of low pass filter when training and subsequently testing the network.For two different state of the art deep architectures, ResNet and U-Net, wedemonstrate that when the training and testing filters are matched,improvements in signal-to-noise ratio (SNR) of up to 7dB can be obtained.However, when these filters differ, the improvement falls considerably andunder some training conditions results in a lower SNR than the band-limitedinput. To circumvent this apparent overfitting to filter shape, we propose adata augmentation strategy which utilizes multiple low pass filters duringtraining and leads to improved generalization to unseen filtering conditions attest time.