Significance of Source Information in Hypernasality Detection

Abstract

This work analyzes the peak to side-lobe ratio (PSR) around each glottal closure instant (GCI) in the Hilbert envelope (HE) of linear prediction (LP) residual as an excitation source-based cue for the hypernasality detection. PSR is defined as the ratio of peak value around GCI to the mean of sample values around GCI in the 3 ms range of HE of LP residual. The coupling between nasal and oral tract occurs during the production of voiced sound in hypernasal speech. The air leakage from nasal tract affects the abruptness of glottal closure, which in turn affects the peak strength around the GCIs. The nasal tract adds zeros in the spectrum of voiced sound. Since the LP model is poor in modeling the zeros in the spectrum, the zeros get filtered in the LP residual signal. This increases the side-lobe strength around the peak in the HE of LP residual. Hence, the PSR gets affected in hypernasal speech. Classification between pre-known normal and hypernasal sound based on a threshold value of PSR gives the accuracy of 70.49, 78.19, 63.15, 60.67, and 67.27% for high vowel, low vowel, glides, liquids, and voicebar sounds, respectively.