Deep acoustic embeddings for identifying parkinsonian speech

Abstract

—Parkinson’s disease is a serious neurological impairment which adversely affects the quality of life in individuals. While there currently does not exist any cure for this disease, it is well known that early diagnosis can be used to improve the quality of life of affected individuals through various types of therapy. Speech based screening of Parkinson’s disease is an active area of research intending to offer a non-invasive and passive tool for clinicians to monitor changes in voice that arise due to Parkinson’s disease. Whereas traditional methods for speech based identification rely on domain-knowledge based hand-crafted features, in this paper, we investigate the efficacy of and propose the deep acoustic embeddings for identification of Parkinsonian speech. To this end, we conduct several experiments to benchmark deep acoustic embeddings against handcrafted features for differentiating between speech from individuals with Parkinson’s disease and those who are healthy. We report that deep acoustic embeddings consistently perform better than domain-knowledge features. We also report on the usefulness of decision-level fusion for improving the classification performance of a model trained on these embeddings.