Detection of hypernasality from speech signal using group delay and wavelet transform

Abstract

One of the most common disorders in children with cleft palate is hypernasality that survives also after operation. To solve this problem, it is required to set many speech therapy sessions. Therefore, assessment of hypernasality is fundamental for speech therapists and could be done either by a nasometer equipment or an expert speech therapist. Recently speech processing methods are introduced as an efficient alternative tool. In this study, vowels (/a/) extracted from 392 utterances of disyllables (/pamap/) that were uttered by 22 normal subjects and 13 subjects with cleft palate have been used and are recorded by nasal and oral microphones. Some analyses are performed on Group Delay parameters as well as features of wavelet transform. The results show that extracted parameters from Group Delay spectrum of second (/a/) in (/pamap/) context, obtained from both nasal and oral signals, are better than that of the first (/a/), and in the best outcomes an accuracy of 94.1 % is achieved. In wavelet transform, statistical features are calculated from 5 sub-bands of Daubechies4 coefficients of two (la/) vowels and their transients. In the best results an accuracy of 97.1 % for transient (lma/) from combination of nasal and oral features is obtained.