Melody extraction based on dynamic bayesian network

Abstract

This thesis considers a melody extraction algorithm based on the state-space equation of the parameters that define melody. In this thesis, melody is defined to be the singing voice pitch sequence in the vocal part and the pitch sequence of leading instrument in non-vocal part of music. The main idea of the considered algorithm is that the parameters that consist of melody pitch and their harmonic amplitudes are assumed to follow two uncoupled first-order Markov processes, and the polyphonic audio is related to the parameters such that the current framed segment of the polyphonic audio is conditionally independent of other framed segments given the parameters. From this perspective, a dynamic Bayesian network (DBN) for melody extraction can be constructed. And the posterior probability is estimated from this DBN, and it is used to estimate the parameters for melody extraction. To obtain the posterior probability, the likelihood and transition probabilities need to be defined. In defining the likelihood, the accompaniment which is considered the difference between polyphonic audio and melody is assumed to follow a multivariate Gaussian distribution. The transition probability of the melody pitch is obtained based on the statistical characteristics of music that account for small and large variation in melody, and the transition probability of the harmonic amplitudes is assumed to be a Gaussian for reasons of mathematical tractability. To estimate the parameters, the sequential Monte Carlo (SMC) method is utilized. The SMC method relies on a so-called sequential importance density, and this density is designed using multiple-pitches which are estimated by a simple multiple-pitch extraction algorithm. Experimental results show that the performance of the considered algorithm is better than or comparable to those of other well known melody extraction algorithms in terms of the raw pitch accuracy and the raw chroma accuracy.