Title:  Bayesian Adaptation and Combination of Deep Models for Automatic Speech Recognition

Committee: 

Dr. C-H. Lee, Advisor  

Dr. Juang, Chair

Dr. Clements

Abstract:

The objective of the proposed research is to deploy a Bayesian adaptation framework for deep model based ASR systems to combat the degradation of the recognition accuracy, which is typically observed under potential mismatched conditions between training and testing. The problem is tackled from both a direct and indirect standpoint. For directly adaptation of DNN, maximum a posteriori estimation and multi-task learning techniques are employed in the manner of regularization in the DNN updating formula. Attempts have also been made to cast DNN into a generative framework to facilitate the deploy of classical ML and MAP techniques. Indirect Bayesian adaptation approach is designed by performing the classic structured maximum a posteriori adaption on transitional Gaussian mixture models with bottleneck features derived from deep neural networks. Leveraging the complementary nature of the direct and indirect adapted ASR models, hierarchical Bayesian system combination technique is employed to further enhance the adaptation performance.