Robust Automatic Speech Recognition

Chapter 4

Processing in the feature and model domains

Abstract

In this chapter, we discuss the feature- and model-space approaches to noise-robust ASR, and analyze their strengths and weaknesses. Feature-space approaches usually do not change the parameters of the acoustic model (e.g., those in HMMs). They either rely on auditory features that are inherently robust to noise or modify the test features to match the training features. Because they are not related to the back-end, the computational cost of these methods is usually low. In contrast, model-domain methods modify the acoustic model parameters to incorporate the effects of noise. While typically achieving higher accuracy than feature-domain methods, they usually incur significantly ...