Unsupervised Singing Voice Separation Using Gammatone Auditory Filterbank and Constraint Robust Principal Component Analysis

This paper presents an unsupervised singing voice separation algorithm which using an extension of robust principal component analysis (RPCA) with rank-1 constraint (CRPCA) based on gammatone auditory filterbank on cochleagram. Unlike the conventional algorithms that focus on spectrogram analysis or its variants, we develop an extension of RPCA on cochleagram using an alternative time-frequency representation based on gammatone auditory filterbank. We also apply time-frequency masking to improve the results of separated low-rank and sparse matrices by using CRPCA method. Evaluation results demonstrate that the proposed algorithm can achieve better separation performance on MIR-IK dataset.