Bayesian approaches in speech recognition

This paper focuses on applications of Bayesian approaches to speech recognition. Bayesian approaches have been widely studied in statistics and machine learning fields, and one of the advantages of the Bayesian approaches is to improve generalization ability compared to maximum likelihood approaches. The effectiveness for speech recognition is shown experimentally in speaker adaptation tasks by using Maximum A Posterior (MAP) and model complexity control by using Bayesian Information Criterion (BIC). This paper introduces the variational Bayesian approaches, in addition to the MAP, BIC and other Bayesian techniques, for speech recognition. VBEC (Variational Bayesian Estimation and Clustering for speech recognition) is a fully Bayesian speech recognition framework, and achieves robust acoustic modeling and speech classification. This paper explains the formulation and experimental effectiveness of these Bayesian approaches for speech recognition.