Evaluating Various Feature Extraction Methods and Classification Algorithms for Music Genres Classification

Mohammadreza Bakhtyari, Saeedeh Davoudi, Sayeh Mirzaei

Abstract: Thousands of songs are released monthly through each of the music streaming services. This amount of data requires careful data management and analysis. Toward this aim, Music Genres Classification (MGC), one of the most prominent research fields of Music Information Retrieval (MIR), has been introduced. While the end-to-end deep learning approach has been successful in many classification tasks, the two-steps approach is more efficient in some practical applications. The first step is to extract appropriate features by a feature extraction technique. Then, they will serve as an input for the classification model in the next step. In this study, we intend to use a two-step approach. We examine the harmonic-percussion source separation (HPSS), the Mel-spectrogram and the modulation spectrogram for the feature extraction stage, and different versions of an inception block for nonlinear features. In this paper, we develop several models by combining various feature extraction methods and classification algorithms and assess the impact of each presence. The model based on the attention mechanism and a convolutional-recurrent architecture with the input of Melspectrogram has shown the best accuracy compared to others.

External IDs:dblp:conf/csicc/BakhtyariDM22