New Feature Extraction from Electroglottographic Signals Applied to Automatic Detection of Laryngeal Pathologies

The objective of this report is to design a mechanism of classification that, through electroglottography, helps distinguishing between healthy and pathological subjects, as well as maximizing the efficiency of electroglottography through an optimal configuration of the classification parameters of SVM (Support Vector Machine). The proposed system consists in parameterizing electroglottography signals obtained in the open database, Saarbruecken Voice DataBase, and to draw the more relevant characteristics in temporary, frequency and cepstral domain. Afterwards, the samples are classified with a SVM. The study carried out contains different combinations of parameters and characteristics in order to assess the appropriate configuration considering: the recorded vowel, the type of windowing, the configured SVM percentages of training and the different values of the SVM parameters. The results obtained are compared to the real data, in this way, it is obtained the performance values of the system (precision, sensitivity and specificity) for each features configuration contemplated. The best results come from vowel I, 30 ms windowing with 50% overlapping, percentages of training around 80–90% (PES higher than PEP) and γ and σ2 values of 100 and 0.1 respectively. This study expects to provide a greater knowledge to the classification methods based on electroglottography as an aid in diagnosing laryngeal diseases.