Music Genre Classification Using Timbral Feature Fusion on i-vector Framework

Abstract

A method for automatic music genre classification based on the fusion of high-level and low-level timbral descriptors is proposed. High-level features namely, i-vectors are computed from mel-frequency cepstral coefficient (MFCC)-GMM framework. Low-level timbral descriptors namely MFCC, modified group delay features (MODGDF) and timbral feature set are also computed from the audio files. Initially, the experiment is performed using i-vectors alone. Later, low-level timbral features are appended with high-level i-vector features to form a high dimensional feature vector (55 dim). Support vector machine (SVM) and deep neural network (DNN) based classifiers are employed for the experiment. The performance is evaluated using GTZAN dataset on 5 genres. With high-level i-vector features, the baseline-SVM and DNN-based classifiers report average classification accuracies (in %) of 79.30 and 80.67, respectively. A further improvement (9\%) in performance was observed when low-level timbral descriptors are fused with the i-vectors in both SVM and DNN frameworks. The results demonstrate the potential of the timbral feature fusion in the music genre classification task.