Abstract
Parkinson's disease patients develop different speech impairments that affecttheir communication capabilities. The automatic assessment of the speech of thepatients allows the development of computer aided tools to support thediagnosis and the evaluation of the disease severity. This paper introduces amethodology to classify Parkinson's disease from speech in three differentlanguages: Spanish, German, and Czech. The proposed approach considersconvolutional neural networks trained with time frequency representations and atransfer learning strategy among the three languages. The transfer learningscheme aims to improve the accuracy of the models when the weights of theneural network are initialized with utterances from a different language thanthe used for the test set. The results suggest that the proposed strategyimproves the accuracy of the models in up to 8\% when the base model used toinitialize the weights of the classifier is robust enough. In addition, theresults obtained after the transfer learning are in most cases more balanced interms of specificity-sensitivity than those trained without the transferlearning strategy.