Statistical and Neural Classifiers: Application for Singer and Music Discrimination in Polyphonic Music Context

摘要

The problem of identifying sections of singer voice and instruments is investigated in this paper. Three classification techniques: Linde-Buzo-Gray algorithm (LBG), Gaussian Mixture Models (GMM) and feed-forward Multi-Layer Perception (MLP) are presented and compared in this paper. All techniques are based on Mel frequency Cepstral Coefficients (MFCC), which commonly used in the speech and speaker recognition domains. All the proposed approaches yield a decision at every 125 ms only. Particularly, a large experimental data is extracted from the music genre database RWC including various style (68 pieces, 25 subcategories). The recognition scores are evaluated on data used in the training session and others never seen by proposed systems. The best results are obtained with the GMM (94% with train data and 80.5% with test data).