Some recent advances in speech recognition with potential applications in other statistical pattern recognition areas

Summary form only given. The paper reviews some recent developments in the area of statistical speech recognition, which could also be potentially useful to other statistical pattern recognition applications. Among other issues, the author discusses the use of new forms of expert mixtures, for example, the one based on the minimization of the product of error probabilities. This rule, sometimes referred to as "product-of-errors rule" has recently been used quite successfully in multi-channel (multi-modal) processing. In speech recognition, this rule was also used to implement automatically noise robust speech recognition approaches (based on frequency subband processing), which do not require noise adaptation or explicit noise models. In a related framework, he introduces the theory of "missing data", yielding significantly improved noise robustness in the case of classification of multidimensional feature vectors prone to noise in some (unknown) components. Finally, as a further generalization, he also discusses a new hidden Markov model (HMM), where the HMM emission probabilities are themselves estimated state-dependent HMMs.