The UCD System for the 2018 FEMH Voice Data Challenge

We are all exposed to a risk of voice disorders at some point in our life, as we use our voice daily for speaking, singing, laughing and more. Voice disorders are often characterized by difficulty when breathing, a vocal fatigue, a sore throat, changes in voice such as hoarseness, repetitive cough, and difficulty when swallowing. When these disorders occur, a change in the acoustic characteristics can be noticed. It is then possible to diagnose these diseases based on the automatic analysis of the acoustic quality of the voice signal. In this paper, we present the National School of Applied Sciences of El Jadida - University of Chouaib Doukkali (UCD) system for automatic voice disorder classification. The system was developed for participation in the 2018 FEMH Voice Data Challenge. For this task, we used audio recorded data from the patients of the Far Eastern Memorial Hospital (FEMH). Different types of features are investigated; the Mel Frequency Cepstral Coefficients, the zero-crossing rate, the roll-off frequency, the spectral centroid, the spectral bandwidth and the spectral contrast. All of these features are normalized using summary statistics such as mean, standard deviation, skewness, maximum, minimum and the median. Experimental results show that the Support Vector Machine outperforms the other classifiers and that the best performance of our classifier is reached when using MFCC with spectral features.