Effective Detection of Voice Dysfunction Using Glottic Flow Descriptors

The presence of various vocal pathologies seriously affects the quality of the speech. These pathologies can treat better if they are diagnosed in primary stage. In this work, for early detection, we conceived non-intrusive automatic vocal fold pathologies recognition system. The sustained vowel /ah:/ with normal intonation for both healthy and pathologic subjects are extracted from PdA corpus. Glottal Inverse Filtering (GIF) is used to estimate glottal pulseform from frame of voiced speech signal. Various time and frequency domain descriptors are extracted from glottal pulseform and used for detection of voice disorder. For inverse filtering, Iterative Adaptive Inverse Filtering (IAIF) algorithm with Discrete All-Pole (DAP) model for vocal tract is used. The extracted descriptors are fed to classifier to separate healthy and pathologic subjects. The artificial neural network (ANN), support vector machine (SVM) and k-nearest neighbour (kNN) were used for classification. We have used box and density plots to investigate the discrimination ability of extracted glottal descriptors. To observe the discrimination ability of descriptors quantitatively, analysis of variance (ANOVA) and information gain feature scoring method is used. The time domain descriptors were found very rich in discrimination compared to frequency domain. The best classification rate achieved were 99.85%, 99.90% and 99.95% with kNN, SVM and ANN respectively.