Sparse Modeling for Artist Identification: Exploiting Phase Information and Vocal Separation

As artist identification deals with the vocal part of music, techniques such as vocal sound separation and speech feature extraction has been found relevant. In this paper, we argue that the phase information, which is usually overlooked in the literature, is also informative in modeling the voice timbre of a singer, given the necessary processing techniques. Specifically, instead of directly using the raw phase spectrum as features, we show that significantly better performance can be obtained by learning sparse features from the negative derivative of phase with respect to frequency (i.e., group delay function) using unsupervised feature learning algorithms. Moreover, better performance is achieved by using singing voice separation as a pre-processing step, and then learning features from both the magnitude spectrum and the group delay function. The proposed system achieves 66% accuracy in identifying 20 artists from the artist20 dataset, which is better than a prior art by 7%.