Voice segmentation for early identification of neurodegenerative diseases

In the context of neurodegenerative diseases, vocal analysis is an easy and inexpensive way to extract useful elements for follow-up, prognostic prediction and rehabilitation of patients. Datasets with recordings of patients and healthy controls reading prompts have been collected to allow the development and evaluation of automatic vocal analysis systems. This work deals with the tasks of speech recognition and forced alignment, important building blocks for such systems, specifically on speech from Parkinson's Disease patients and healthy controls. A system has been designed and developed to perform automatic speech recognition and forced alignment, based on fine tuned state-of-the-art models, it can leverage unlabeled data and it can be used to align spontaneous speech with no prompts. The output of the system are the words and phonemes identified in a recording and the time-alignment of each single word and phoneme; this enables the automatic segmentation of vocal data and adds other data points for subsequent steps of analysis and correlation with the clinical parameters of the patients. The performance of the system is evaluated on “normal” speech datasets and disordered speech.