Robust Automatic Speech Recognition

Chapter 8

Joint model training

Abstract

In this chapter, we use the final, fifth attribute—use of the same process at the training and decoding stages for removing the same sources of variability consistently—to categorize and analyze many noise-robust ASR techniques in the literature. We call such techniques joint model training (with decoding). More specifically, the feature compensation or model adaptation technique used in the decoding stage is also used in the training stage so that a pseudo-clean acoustic model can be established after the training. Such joint training can be carried out in the feature enhanced domain, where noisy training data is first cleaned by feature compensation methods, and subsequently the enhanced features are ...