Voice researchers and phoniatric doctors have been concerned for decades, how the nature of phonation disorders can be detected from the acoustic parameters of voices.
Purpose of my diploma work is to develop a system, which is applicable for a medical decision support system with separation of the healthy and pathological voice patterns. I searched for an engineering solution, which could be the most applicable one in the viewpoint of this task. I demonstrated on a theoretical level and with the help of test reults the possibilities of SVM (Support Vector Machine), and the fuzzy based classification. A well-designed and labelled database, which contains pathologic voices, is necessary for this process. The built-up of this database, and principles of the classification methods were defined on the basis of the phoniatric practice.
I developed a fuzzy classification system in MATLAB, which is able to make a whole cross-evaluation, and a normal testing, and also to find the best parameter settings automatically. I implemented a visualisation system, and with the help of this I could gain information from the results of training the fuzzy system related to the classification and to the database, too.
I made a leave-one-out crossvalidation, and a normal testing either for fuzzy, or for SVM systems. On the basis of test results I compared SVM and a fuzzy solution applying three different teaching algorithms (Fuzzy C-Means, Gustafson-Kessel, Gath-Geva).
Comparison of SVM and fuzzy-based classification system was made on a reference voice database generated by myself. I found out, that that there are either advantages, or disadvantages of the two different types of classifiers from the viewpoint of the medical decision support. The SVM-based solution is able to achieve a major accuracy in the field of classification, however, it can provide very few information about the problem. There is essentially more information on the output of the fuzzy-based system at our disposal, and there is a possibility for a kind of transformation of this output, which has significant surplus of information, that is very important for the doctors.