LSTM Siamese Network for Parkinson's Disease Detection from Speech

Parkinson's disease (PD) is a neurodegenerative disorder that severely affects motor functions. Symptoms include dysarthria and this fact has been the basis for PD detection from speech in several works. Machine learning-based technologies have made significant strides in automatic speech recognition, but their use is fairly limited for clinical diagnosis due to the lack of large corpora containing parkinsonian speech. In this work, we propose a two-step strategy to use machine learning methods for PD detection. In the first step, we use Long Short-Term Memory (LSTM)-based siamese networks to learn feature representations that highlight the information related to speech articulation and prosody relevant for PD detection. Siamese networks are trained on data pairs employing a Spanish corpus containing 52 patients and 56 control subjects. In the second step, we train a classifier to make decisions about the presence or absence of PD employing the features provided by the LSTM networks. We achieve an EER of 1.9% in the detection by combining the scores of different text-dependent models. Preliminary experiments show the efficacy of the proposed method and prove the usefulness of LSTM for PD detection from speech.