Melody Extraction From Polyphonic Music Using Particle Filter and Dynamic Programming

Abstract

Melody extraction from polyphonic music is one important but challenging task in the music information retrieval community. In this paper, a new melody extraction method based on the particle filter and dynamic programming is proposed. The constant-Q transform is first introduced for multiresolution spectral analysis of polyphonic music. Then, the melody extraction is modeled in the Bayesian filtering framework, and the particle filter is used to get a rough melody contour. Specially, the pitch transition probability of adjacent frames is approximated according to the statistical analysis based on one publicly available dataset, and the likelihood of frame-wise pitches is defined by considering pitch salience, spectral smoothness, and timbre similarity. After that, the preliminary melodic contour obtained by particle filter is smoothed to achieve the frame-wise pitch range limitation. Finally, the dynamic programming is used to accurately track the final melodic contour. The proposed method requires no prior information, and is suitable for both instrumental and vocal melodies. The experimental results show that the performances of the proposed method is robust among four publicly available datasets comparing with the state-of-the-art methods, and it achieves the highest averaged raw pitch accuracy and raw chroma accuracy performances with lower octave errors.