The aim of this master’s thesis is to evaluate the applicability of interpretable speech
biomarkers and deep speech representations for the automatic classification of individuals
with Parkinson’s disease and healthy controls. The thesis is based on the assumption
that Parkinson’s disease affects speech production, particularly in relation to hypokinetic
dysarthria, and that these changes can be captured using acoustic and deep speech
features.
The experimental part is based on a database of spontaneous monologues. Adjusted
interpretable speech biomarkers and x-vector representations extracted using a pretrained
ECAPA-TDNN model are used. Three approaches for combining both types of features
were designed and compared: late fusion, early fusion, and hybrid fusion. The biomarkerbased models as well as the early and hybrid fusion models use the XGBoost algorithm
with hyperparameter tuning performed using Optuna, whereas the x-vector branch and
the late-fusion meta-model are based on logistic regression. The final evaluation was
performed using the leave-one-out method.
The results show that x-vector representations provide stronger discriminative information than standalone interpretable biomarkers. The standalone x-vector model based on
logistic regression achieved an accuracy of 0.7519 and an AUC value of 0.8045, while
the biomarker-based XGBoost model achieved an accuracy of 0.6589 and an AUC value
of 0.6878. Fusion approaches made it possible to combine the performance of deep representations with the information contained in interpretable biomarkers; however, their
contribution depended on the specific feature combination strategy. Among the fusion
approaches, hybrid fusion achieved the best results, with an accuracy of 0.7597 and an
AUC value of 0.8117. The results confirm the potential of automatic speech analysis
for assessing speech manifestations of Parkinson’s disease; however, they must be interpreted with respect to the limited size of the dataset, the validation protocol used, and
the possible influence of individual differences between speakers.