Study of Indian classical music by singing voice analysis and music source separation

The pitch variations in classical singing are known for their pleasant appeal. But these also add to the challenges in acoustic analysis of singing voice signals. The characteristics of singing voice signal such as variable base frequency, inter-tonal gaps, and intense and rapid changes within each pitch-period need to be examined, along with changes in the background music signal. In this paper, we first separate and then characterize the singing voice and music signal, by analyzing the changes in production features. The pitch, formants and energy features are examined for two prominent compositions of classical singing, Alaap and Lyrical compositions. The production features are derived from the acoustic signal, using the signal processing methods such as short-time Fourier transform (STFT), linear-prediction analysis, and zero-frequency filtering. The music and vocal regions are separated by music-source separation, using STFT with different types of windowing techniques such as Blackman, Hamming and Kaiser windows. The background music is separated from the music mixture involves Similarity Matrix-based technique to model the background music. Results of the experiments indicate that Alaap regions have higher pitch in case of female singers, whereas Lyrics compositions have higher pitch frequency for male singers. Result so fusing different windowing techniques also give decent performance for music-source separation.