A sequential naive Bayes method for music genre classification based on transitional information from pitch and beat

摘要

Due to the rapid development of digital music market, online music websites are widely available in our daily life. There is a practical need to develop automatic music genre classification algorithms to manage a huge amount of music. In this regard, the transitional information contained in pitches and beats should be very useful. Particularly, the transition in pitches produces a melody, and the transition in beats produces a rhythm. They both decide the music genre. To take these valuable information into consideration, we propose here a sequential naive Bayes method for music genre classification. This method can be viewed as an novel extension of the classical naive Bayes classifier, but takes the transitional information between pitches and beats into consideration. To reduce the number of estimated parameters, we propose a BIC-type criterion and develop a computationally efficient algorithm for model selection. The selection consistency of the BIC method is theoretically proved and numerically investigated. The finite sample performance of the proposed methods are assessed through both simulations and a real music dataset.