Melody Extraction From Singing Voice Of Polyphonic Music

Melody extraction from singing voice of polyphonic music is one of the most important and difficult issues in the area of the Music Information Retrieval (MIR). Singing melody extraction has many applications including singing voice separation, singer identification and query-by-humming. The main job of singing melody extraction is to extract singing fundamental frequency from polyphonic music. The main purpose of this thesis is to improve the accuracy of singing melody extraction. The main contents are shown as follows:Frequency and quefrency features are used to perform singing voice detection. Mel Frequency Cepstrum Coefficient (MFCC) and LPc-derived Mel Cepstrum Coefficient (LPMCC) are examined in the singing voice detection experiments. It is found that the performance of LPMCC are superior to MFCC. Comparing Log Frequency Power Coefficient (LFPC), Mel Frequency Power Coefficient (MFPC), and Gammatone Frequency Power Coefficient (GFPC) are compares via the singing voice detection experiments as well. It is found that the 25 dimension LFPC brings the highest accuracy rate.The Iterative Spectral Subtraction and Modified Harmonic Product Spectrum (MHPS) are used to perform multiple pitch estimation in the polyphonic music and the performance of the two algorithms is compared. The former can solve the problem of harmonic overlaps and the latter can solve the problem of fundamental frequency missing and octave error in some degree. The results show that the number of estimated fundamental frequency in each frame used by Iterative Spectral Subtraction algorithm less than MHPS algorithm, and Iterative Spectral Subtraction Algorithm can get better raw pitch accuracy.According to the time and pitch continuous, the thesis creates pitch trajectory from multiple pitch set. The experiments show that after the pitch trajectory creation, raw pitch tracking accuracy increases slightly, and some outliers are also excluded. The thesis proposes the method based on harmonic energy ratio to determine singing melody. And summaries the method based on frequency modulation and method based on feature combination. Comparing the performance of the three methods, it is found that method based on feature combination can get better result in the singing melody determination, and can improve the accuracy of singing pitch extraction.