Speech contains different paralinguistic aspects especially pathologies may affect speaker’s communication. Patients with neurodegenerative diseases such as Parkinson’s disease (PD) usually have hypokinetic dysarthria, and there will be clinical manifestations such as unclear expression and vague voice will appear when speaking. At present, the conventional medical clinical diagnosis mainly depends on the experience judgment of doctors such as static tremor and slow motion, etc. Developing automatic assessment of pathological speech will improve the efficiency and accuracy of diagnosis and treatment if we make use of advanced technology to assist doctors. This paper develops deep learning methods in speech emotion analysis with medical disease diagnosis for early detection of pathological speech. Speech audio features are extracted according to unsupervised learning approach and utterance-level features are constructed for comparison. Meanwhile, we provide feature importance analysis for further medical diagnosis. Deep neural networks (DNNs) and support vector machines (SVMs) are introduced for identifying PD patients and health control (HC) subjects, which in the further study allows to support medical diagnosis and disease severity evaluation.

Neurodegenerative Diseases,Automatic Speech Recognition,Deep Learning,Autoencoders,deep neural networks