A method based on the production mechanism of the vocals in the composite vocal polyphonic music signal is proposed for vocal melody extraction. In the proposed method, initially the non-pitched percussive source is suppressed by observing its wideband spectral characteristics to emphasise the harmonic content in the mixture signal. Further, the harmonic enhanced signal is segmented into vocal and non-vocal regions by thresholding the salience energy contour. The vocal regions are further divided into vocal note like regions by their spectral transition cues in the frequency domain. The melody contour in each vocal note is extracted by detecting the locations of instant of significant excitation by passing it through adaptive zero frequency filtering (ZFF) in the time domain. The experimental results showed that the proposed method is indeed comparable to the state-of-the-art saliency based melody extraction method.
Cite as: M., G.R., Rao, K.S. (2016) Enhanced Harmonic Content and Vocal Note Based Predominant Melody Extraction from Vocal Polyphonic Music Signals. Proc. Interspeech 2016, 3309-3313, doi: 10.21437/Interspeech.2016-856
@inproceedings{m16_interspeech,
author={Gurunath Reddy M. and K. Sreenivasa Rao},
title={{Enhanced Harmonic Content and Vocal Note Based Predominant Melody Extraction from Vocal Polyphonic Music Signals}},
year=2016,
booktitle={Proc. Interspeech 2016},
pages={3309--3313},
doi={10.21437/Interspeech.2016-856}
}