Singing Pitch Extraction from Monaural Polyphonic Songs by Contextual Audio Modeling and Singing Harmonic Enhancement

This paper proposes a novel approach to extract the pitches of singing voices from monaural polyphonic songs. The hidden Markov model (HMM) is adopted to model the transition between adjacent singing pitches in time, and the relationships between melody and its chord, which is implicitly represented by features extracted from the spectrum. Moreover, another set of features which represents the energy distribution of the enhanced singing harmonic structure is proposed by applying a normalized sub-harmonic summation technique. By using these two feature sets with complementary characteristics, a 2stream HMM is constructed for singing pitch extraction. Quantitative evaluation shows that the proposed system outperforms the compared approaches for singing pitch extraction from polyphonic songs.