Predominant Instrument Recognition in Polyphonic Music Using GMM-DNN Framework

摘要

In this paper, the predominant instrument recognition in polyphonic music is addressed using timbral descriptors in three frameworks-Gaussian mixture model (GMM), deep neural network (DNN), and hybrid GMM-DNN. Three sets of features, namely, mel-frequency cepstral coefficient (MFCC) features, modified group delay features (MODGDF), and lowlevel timbral features are computed, and the experiments are conducted with individual set and its early integration. Performance is systematically evaluated using IRMAS dataset. The results obtained for GMM, DNN, and GMM-DNN are 65.60%, 85.60%, and 93.20%, respectively on timbral feature fusion. Architectural choice of DNN using GMM derived features on the feature fusion paradigm showed improvement in the system performance. Thus, the proposed experiments demonstrate the potential of timbral descriptors and DNN based systems in recognizing predominant instrument in polyphonic music.