NEURAL NETWORKS FOR INTERFERENCE REDUCTION IN MULTI-TRACK RECORDINGS

Multi-track recordings are sometimes created by simultaneously capturing several sources with several microphones. This scenario can result in the interference of undesired source(s) in the various tracks. Interference reduction aims to recover the source(s) associated with a particular track. In this paper, we present two neural networks for interference reduction. The first network uses a convolutional autoencoder-based architecture and uses time-frequency representation as input. The second network uses a truncated Unet architecture and directly estimates the interference from the time-domain multi-track representation. Our experiments indicate the effectiveness of the proposed methods, with the truncated Unet showing superior performance. Also, the audio outputs produced by the proposed methods have improved quality, resulting in better music source separation performance. Code is available at https://github.com/its- rajesh/IRMR/