Robust Automatic Speech Recognition

Chapter 11

Summary and future directions

Abstract

In this book, we presented robust ASR techniques guided by a unified mathematical framework. To offer insight into the distinct capabilities of these techniques and their connections, we have used a taxonomy-oriented approach with five key attributes: feature vs. model domain processing, explicit vs. implicit distortion modeling, use vs. absence of prior knowledge about the distortion, deterministic vs. uncertainty processing, and joint vs. disjoint training, to organize the vast amount of material and to demonstrate the commonalities and differences among the plethora of robust ASR methods presented in this book. The representative technologies for both GMMs and DNNs are described. Given the ...