Investigation of glottal flow parameters for voice pathology detection on SVD and MEEI databases

This paper investigated the performance of glottal flow features for the voice pathology detection particularly begnin and malignant tumors in two distinct databases. Glottal features have been widely used over the last years in pattern recognition process. The purpose of this work was to find out the most relevant glottal flow features for detecting voice disorders from normal ones. In order to choose the discriminative features, two different selection measures were applied in this work. The experiments were carried out using two different databases, “MEEI” and “SVD”, American and German databases, respectively. These databases included normal and pathalogical utterances pronounced by male and female speakers. Only the sustained vowel /a/ was used in classification task. Artificial Neuron Network (ANN) and Support Vector Machines (SVM) were used to perform the classification of normal-pathological voice. The experimental results prove that there is clear difference in performance of these glottal features independently of the used databases. The top-features selected were also varied from one database to another. There is a high accuracies using the SVM classifier, but it remains less important compared to those obtained using the ANN. The best classification rates achieved are 99.27% and 93.66% for SVD and MEEI databases, respectively.