Hypernasality severity analysis using spectral and residual features

In this work, spectral and linear prediction residual characteristics of normal, mild, and moderate-severe hypernasal /i/ vowel are analyzed for finding their nature in these three levels of hypernasality severity. The nasal tract and the oral tract coupling during vowel sound production by the children with cleft palate introduces additional nasal pole-zero pairs in these vowels. So the spectral characteristics of these vowels deviate compared to normal vowels. Further, the inappropriate modeling of this deviated spectrum by the LP model, due to the presence of zeros, gives the deviated residual characteristic also. The spectral and residual analysis is done using vocal tract constriction (VTC) evidence, power spectral entropy (PSE) and peak to side-lobe ratio (PSR) features. The first two features capture the spectral deviation in terms of low-frequency dominance and spectral flatness in the hypernasal vowel spectrum respectively whereas the third feature capture the residual deviation in terms of the presence of undesirable signal components in the residual signal. The analysis result reveals that the value of VTC evidence increases whereas the value of PSE and PSR features decreases as the severity of hypernasality increases.