Music component characterization in the music-speech mixture for female singing tracks

摘要

Conventional pitch-contour extraction techniques often exploit the quasi-periodic nature of speech signals. However, the task of pitch extraction for singing voice signal is challenging, especially when the music signals are mixed with multi-pitch speech sound signals. Usually, the background music signal is characterized by rhythmic and repetitive patterns in the music-speech mixture. Hence, in this paper we propose a pitch extraction method for such a music-speech mixture, by exploiting these characteristics. The signal processing techniques such as autocorrelation, zero-frequency filtering and spectral analysis are explored. But, dynamic spectral characteristics of the musical compositions make it difficult to identify the different components of the acoustic signal. Hence, variations in the F0 during lyrical singing, and repetitive beat-patterns in the background music, are distinguished using the acoustic features. Features F0 and its harmonics, strength of excitation and signal energy are derived from the acoustic signal, using different signal processing methods. The comparison of harmonics of the single-pitched and multi-pitched sound sources gives insightful results. Performance of the proposed method is better than other existing methods of music-speech sound source modelling and processing.