ملخص: We examine key research issues in adaptively modifying the conventional plug-in MAP decision rules in order to improve the robustness of the classification and decision strategies used in automatic speech recognition (ASR) systems. It is well known that the commonly adopted plug-in MAP decoder does not achieve the minimum error rate desired in ASR because the joint probability distribution of speech and language is usually not known exactly. The optimality issue becomes even more serious when there exists acoustic mismatch between training and testing conditions. We review in detail two recently proposed classification rules, namely minimax classification and Bayesian predictive classification. Both of them model classifier parameter uncertainty and modify the classification rules to satisfy some desired robustness properties. We also present an overview on a number of related techniques and discuss how these algorithms can be used to improve the robustness of speech recognizers.

لغة: إنجليزية