| Abstract [eng] |
This master thesis investigates the detection of Parkinson's disease using techniques of speech signal analysis. Existing researches, detection problems, used features, applied methods and achieved results are discussed in the literature review section. The process of speech signal analysis is divided into stages, analysis of the speech feature extraction is made, the process of compression of the received attributes using statistical functionals and GMM by using the random classification of the forest is done in the research methods part. Two databases of speech recordings were used during the experiments. The first database consists of 369 recordings that are recorded by acoustic cardioid microphone and the second database consists of 98 recordings that are recorded by smart phone. Signals for both databases are recorded simultaneously, under the same conditions, but the second database has fewer subjects than the first one. Both databases contain a single pronunciation sentence “turėjo senelė žilą oželį”. The equal error rate (EER) was used to evaluate the goodness of detection. The experimental investigation has shown that using the GMM for the first database resulted in EER 13.41% and EER 13.43% was achieved by using statistical functions. After the first dataset was reduced to the second size (both datasets had same subjects), the lowest EER 21.32% was achieved by using statistical functions of the acoustic microphone. Meanwhile, the smart phone achieved the lowest EER of 26.65% by using the GMM. Variable importance analysis from the random forest indicated that Mel-frequency cepstral coefficients (MFCC) and various spectral descriptors from Essentia library are the most important descriptors for detection accuracy. Furthermore, statistical functions of the upper quartile and standard deviation are the most important for compression of the short-term audio features. A prototype of the Parkinson's disease detection expert system was created during the research, which allows the system user to access it using the internet browser. This solution was created to share the results and to detect the disease at an early stage phase using a non-invasive tool. It is proposed to combine from speech and sustained voice modalities in the future in order to improve the accuracy of Parkinson's disease detection. |
