Publication

Abstract : Cleft lip and palate is one of the main reason for speech disorders in children. The major one among these speech disorders is hypernasality caused due to abnormal nasal resonance. For the detection of hypernasality, we commonly analyse normal vowels and nasalized vowels. In this paper, we have done a comparative study on two feature extraction methods such as Mel Frequency Cepstral Coefficients (MFCC) and Bessel Frequency Cepstral Coefficients (BFCC). We have trained these features using Gaussian Mixture Model (GMM) of various orders. The classification of hypernasal speech is performed with upto 50% for 'A', 87.5% for 'E', 100% for 'I', 'O' and 'U' for MFCC with GMM of various orders. For BFCC with GMM of various orders, performance upto 87.5% for 'A', 75% for 'E', 75% for 'I', 100% for 'O' and 'U' is achieved. We have given an analysis hypernasal and normal speech vowel using its spectrogram.