Improving singing voice detection in presence of pitched accompaniment
N. R Pavan Santosh, S. Ramakrishnan, Vishweshwara Rao, Preeti Rao · 2009
The paper addresses the problem of singing voice detection in Indian Classical Music where we have the presence of strongly pitched accompaniment. Visual observation of spectra suggests that the temporal fluctuation of the higher harmonics is a strong characteristic of the singing voice. Standard deviation of the frequency tracks, obtained from sinusoidal modeling of the harmonic signals, is proposed as a feature for the classification. In order to reliably compute the feature, accurate frequency estimation is required. We observe that phase based methods provide for superior sinusoid detection and estimation when compared to magnitude based methods. Further, the superiority of multi resolution spectrum analysis over conventional single resolution analysis is demonstrated for real musical signals.