Off-campus University of Richmond users: To download campus access theses, please use the following link to log in to our proxy server with your university username and password.
Date of Award
Restricted Thesis: Campus only access
Bachelor of Science
Dr. Doug Szajda
The prevalence of automated speech processing systems has led to a corresponding growth in adversarial attacks on those systems. In order to better understand, and hopefully prevent those attacks, we need to better understand how voice processing systems “reason” when generating classifications. Though the mathematics that govern machine learning models is well understood, the complexity of modern deep learning models has resulted in a lack of transparency about the specific factors that drive classification of individual model inputs. The current research aims to address this issue in the speech processing domain by leveraging the reasoning behind speech processing model predictions by leveraging the Local Interpretable Model-agnostic Explanation (LIME) method, which provides human-understandable explanations for why target inputs are classified as they are. We are integrating the voice processing pipeline with LIME to generate explanations that illuminate the audio signal characteristics that drive specific character outputs. Implementing LIME within the speech processing pipeline raises significant theoretical and practical issues. We are currently designing and running experiments in order to integrate LIME in such a way that it produces reasonable and illuminating explanations.
Hu, Xiaodi, "Understanding Model Reasoning in Automated Speech Processing Systems with LIME" (2021). Honors Theses. 1575.