Robust Automatic Speech Recognition

Chapter 6

Explicit distortion modeling

Abstract

In this chapter, we use the third attribute—with the use of explicit models of acoustic distortion or otherwise—to categorize and analyze a host of noise-robust ASR techniques in the literature. Without using explicit models of acoustic distortion, general model adaptation methods for robust ASR usually make use of a set of linear transformations to compensate for the mismatch between the training and test conditions. This involves many parameters and thus typically requires a large amount of data for parameter estimation. This difficulty can be overcome when we exploit explicit distortion models of acoustic distortion that take into account the way in which distorted speech features are produced. ...