Parkinson's disease (PD) and essential tremor (ET) are prevalent movement disorders primarily affecting the elderly population. Accurate diagnosis is challenging due to shared clinical features, such as tremor. While PD is characterized by hypokinetic dysarthria (monoloudness, monotonicity, imprecise pronunciation) and ET by hyperkinetic dysarthria (phonatory and prosodic disturbances due to tremor), the potential of speech analysis for reliable differentiation has not been extensively explored. Developing automatic discrimination technology could significantly improve early diagnosis and continuous monitoring. However, a major hurdle is the limited availability of suitable speech data and the significant phonetic variability across different languages. This research proposes a novel method to overcome these limitations by using a cross-language approach, adapting models trained on German and Spanish speech data to classify PD and ET in Czech speakers. The aim is to determine the effectiveness of speech signals in distinguishing PD and ET, identify the most affected speech dimensions (articulation, phonation, prosody), and assess the compatibility of the methodology across different languages. The hypothesis is that a UBM trained on data from speakers of specific languages can accurately model speech impairments in patients speaking a different language.

Previous studies have explored differentiating PD and ET using various sources like video-recorded neurological examinations, hand tremor signals, gait signals, electromyograms, handwriting, and medical images. While research has successfully differentiated hypokinetic dysarthria in PD from hyperkinetic dysarthria in other disorders (e.g., Huntington's disease), a direct comparison between PD and ET using speech analysis was absent. This gap highlights the need for speech-based differentiation, particularly in early disease stages when clinical signs are subtle. The lack of sufficiently large and standardized speech databases for neurodegenerative diseases, exacerbated by cross-language phonetic variations, has hindered the development of effective machine learning frameworks. While some studies suggest that language differences may not significantly impact clinical assessments of disease phenotypes, the impact on speech-based automatic classification requires further investigation. Therefore, creating cross-language models is crucial for robust and generalizable voice-based pathology classification and monitoring. This study addresses this need by focusing on the differentiation between PD and ET, which serves as a valuable theoretical model for testing a cross-language speech analysis framework.

The study employed a GMM-UBM approach combined with SVM for domain adaptation across languages and pathologies. The GMM-UBM framework allows for knowledge transfer from models trained on one language (German and Spanish) to classify Czech speakers. This approach is particularly suitable for scenarios with limited data, which is common in the study of rare diseases. SVMs were chosen for their robustness in handling small datasets. The study used speech recordings from 50 PD patients, 50 ET patients, and 50 healthy controls (HC) who were all native Czech speakers. Recordings were made using a standardized protocol. Three speech dimensions were modeled: articulation (features extracted from transitions between voiced and unvoiced segments), phonation (features characterizing voiced sounds), and prosody (features modeling intonation, timing, and loudness). UBM models were trained using data from Spanish (CIEMPIESS) and German (Verbmobil) datasets, as well as a combination of both. The MAP adaptation method was used to adapt these UBMs to individual Czech speakers. GMM supervectors were generated for each speech dimension and fused using principal component analysis (PCA) for dimensionality reduction. Bi-class (PD vs. ET) and tri-class (PD vs. ET vs. HC) classifications were performed using SVM classifiers. Model performance was evaluated using accuracy, sensitivity, and specificity. A stratified 10-fold cross-validation strategy was used. Hyperparameter optimization for the SVM classifiers involved a grid search across different values of C and γ.

In bi-class classification (PD vs. ET), the fusion of the three speech dimensions achieved the best results. Accuracies of 81.4% (monologue task) and 86.2% (/pa-ta-ka/ task) were obtained. For the /pa-ta-ka/ task, the model adapted from the German UBM performed best. In the /pa-ta-ka/ task, prosody showed the strongest contribution to accuracy, while articulation was more important in the monologue task. Tri-class classification (PD vs. ET vs. HC) yielded lower accuracies. The fusion of the three speech dimensions reached 63.3% accuracy for the monologue task, while the prosody dimension alone achieved 71.6% for the /pa-ta-ka/ task. The confusion matrices revealed that misclassifications were more frequent for PD patients being misclassified as healthy controls, potentially due to the overlap between speech changes associated with hypokinetic dysarthria and healthy aging. Analysis of the data distribution using linear discriminant analysis (LDA) visually confirmed the better separability of ET patients, particularly in the /pa-ta-ka/ task. No significant correlation was found between age, disease severity, and LDA components.

The high accuracy achieved in the bi-class classification demonstrates the potential of automatic speech analysis for differentiating PD and ET. The complementary nature of articulation, phonation, and prosody features is evident, with articulation being more impactful in spontaneous speech and prosody in controlled tasks. This suggests that different aspects of speech production are affected differentially in PD and ET. The lower accuracy in tri-class classification highlights the challenge of distinguishing subtle speech changes associated with early-stage PD from healthy aging. The superior performance of models trained on German data compared to Spanish may reflect the closer linguistic relationship between Czech and German. The results suggest that the /pa-ta-ka/ task, a language-independent test, is slightly better for differentiating between PD and ET than spontaneous speech, likely due to cerebellar involvement in ET affecting sequential motor planning. The robustness of the GMM-UBM approach across different recording conditions is shown.

This study demonstrates the feasibility of using automatic speech analysis to distinguish between PD and ET, even with cross-language adaptation. The high accuracy obtained in bi-class classification, particularly with the /pa-ta-ka/ task, underlines the potential of this method as a valuable clinical tool. Further research should focus on validating and expanding this approach to encompass earlier disease stages, employing deep learning techniques, and exploring transfer learning strategies across various languages and dysarthria types. The findings highlight prosody and articulation as promising biomarkers for differential diagnosis.

While the study demonstrated high accuracy, some limitations exist. The UBM models trained with patient data showed lower performance, possibly due to the high variability of dysarthric speech patterns. The study lacked sufficient ET data in other languages to create UBMs that included both pathologies. The better recording quality of the Czech data compared to the training data might have influenced the results. Finally, the potential impact of cognitive impairment on speech was not explicitly investigated. Future research should address these limitations to further improve the reliability and generalizability of the proposed method.