Feature selection for hypernasality detection using PCA, LDA, kernel PCA and greedy kernel PCA

Cleft lip and palate, due to morphological problems, allow the passage of air through the nasal cavity, introducing inappropriate nasal resonance during speech production and resulting in hypernasality speech. This paper proposes a methodology based on spectral and cepstral features, such as Modified Group Delay Functions with Mel Frequency Cepstral Coefficients, and uses relevance analysis and redundancy elimination, allowing the automatic hypernsality detection. The methodology seeks to evaluate four kinds of selection techniques: LDA (Linear Discriminator Analysis), PCA (Principal Component Analysis), Kernel PCA and Greedy Kernel PCA which provide a lot of information in the detection process and in turn contain the lowest value of redundancy. The experiments were performed considering a database which includes the five Spanish vowels uttered by 130 children whose voices were diagnosed as hypernasal by a phoniatrics expert plus 108 healthy were analyzed.