The paper presents a new method for sound source separation in classical music recordings. This approach uses spatial information obtained during the recording setup to train a spatial mixing filter. This filter is integrated into a multichannel non-negative matrix factorization (MNMF) scheme to better capture the variances of different sound sources. The recording setup used in the experiments is typical for orchestra recordings. The results show that this method improves performance compared to conventional MNMF methods in separating individual sound sources.
Publication date: 10 Oct 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2310.05821
