Maestrías de la Facultad de Ingeniería · 2022
Transfer learning strategies to model speech impairment in patients with neurodegenerative diseases
ABSTRACT : Nowadays, the interest in the automatic analysis of speech in different scenarios has increased. This biomarker has been explored for diagnostic support and monitoring of different neurodegenerative diseases such as Parkinson’s disease (PD) and Huntington’s disease (HD). Speech has the main benefit of being a non-invasive method, which can be captured remotely, at a low cost, and includes detailed information about each participant. Deep learning (DL) has facilitated the development of different robust computational models, their main advantages are that these systems receive the raw signal at the input and can process a large amount of information in parallel. However, obtaining large amounts of data for pathological speech, particularly in neurodegenerative diseases, is very difficult and expensive. Therefore, it is necessary to implement DL techniques to address this issue. Transfer Learning (TL) takes advantage of the experience obtained in a previously trained model to improve the learning of new target phenomena. The main aim of this work is to evaluate the suitability of using DL for pathological speech processing applications, particularly using TL methods. Different approaches are implemented including transfer knowledge in classical methods by crossing pathologies and languages for the classification of PD and HD in the Czech language. In addition, different Convolutional Neural Networks (CNNs) are trained to perform a fine-tuning strategy from cross-language and cross-pathology for the discrimination of patients with PD and HD with respect to Healthy Control (HC) subjects and also the estimation of their depression level. Subsequently, a freezing of layers strategy was performed for the classification of PD patients vs. HC subjects in different languages. Finally, a DL strategy based on embeddings is proposed to know which additional demographic information a CNN learns from pathological speech data.