Title:  Bayesian Adaptive Learning of Deep Latent Variables for Acoustic Knowledge Transfer

Committee: 

Dr. Lee, Advisor

Dr. Moore, Chair

Dr. Anderson

Dr. Heck

Abstract: The objective of the proposed research is to establish a Bayesian adaptive learning framework, to bridge the performance gap between source and target domains for acoustic classification tasks. Acoustic mismatches, such as changes in speakers, recording devices and environmental noise, usually cause an unexpected and severe performance degradation. Traditional Bayesian formulations often impose uncertainties on model parameters to perform adaptation. However, for adaptation of deep neural networks (DNNs) commonly used today, the number of parameters is usually much more than the available training samples, causing the estimated DNN parameters to be inaccurate and making Bayesian adaptation less effective. To solve these problems, we propose to build a Bayesian adaptive learning framework based on the manageable latent variables of DNNs. In our preliminary research, we investigate a variational Bayesian approach. The prior distributions of the source latent variables are optimally combined, in a Bayesian sense, with a small set of data from the target domain to obtain the posterior distributions through a variational inference framework. Promising experimental results were obtained in our experiments.