Neural Generative Codes for Clinical Speech Pathology Applications
Neural Generative Codes for Clinical Speech Pathology Applications


A goal of this talk is to introduce how deep learning can be used to estimate the neural codes of speech generation known as phone-attributes. In particular, we will see their application for healthcare, namely pathological speech analysis.

Using machine learning in clinical applications is still controversial as promising results are diminished by a lack of interpretability. Phone-attributes are popular in machine learning-based speech analysis, and yet are interpretable, such as voicing and a place/manner of articulation. A range of clinical applications using the phone-attribute codes is presented: severity assessment of progressive apraxia of speech, classification of Parkinson's Disease (PD) vs. healthy control subjects from isolated and running speech, prediction of the UPDRS score of the PD patients, estimation of voice quality of the PD patients, and distinguishing ON from OFF motor state in patients with PD. (The content has been published in multiple IEEE journals of engineering as well as medical journals for speech pathology diagnosis and treatment.)


Afsaneh has more than 18 years of professional experience in applied Artificial Intelligence R&D. She completed her Ph.D. at the Swiss Federal Institute of Technology Lausanne (EPFL) on the robustness of machine learning and pattern recognition in commercial applications. Her contributions were recognized by community awards and granted several million Euros of Swiss, European & American funding. At Digital Product School of UnternehmerTUM, she supports her teams with exploiting the AI capabilities, their feasibility, and risks for problem-solution fits. In a multi-disciplinary setup, Afsaneh guides them throughout the challenges on how a data-driven solution should be developed and evaluated and how to transform the assumptions and data requirements to product vision and market opportunities