Parkinsoni tõve tuvastamine eestikeelsete hääleklippide analüüsi abil kasutades masinõppe meetodeid

Abstract

This Bachelor’s thesis investigates the applicability of machine learning methods for Parkinson’s Disease (PD) detection using Estonian voice clips. The research focuses on three main questions: firstly, evaluating the generalizability of an acoustic feature-based model trained on the Spanish PC-GITA dataset to Estonian data; secondly, examining whether combining Spanish and Estonian data during training improves model performance; and thirdly, testing the direct applicability of a state-of-the-art self-supervised learning (SSL) based WavLM Base model, fine-tuned elsewhere, on Estonian data. The results indicate that the direct cross-lingual transferability of acoustic feature-based models is limited, but combining datasets significantly improves performance up to 0.7893. The direct application of a pre-fine-tuned SSL model on short Estonian audio segments without further adaptation was not successful. The thesis highlights the need for language-specific adaptation and the use of multilingual datasets in voice-based PD detection.