Information-theoretic distortion measures for speech recognition: theoretical considerations and experimental results

It is shown that there is a general framework, based on information theory, underlying many currently popular distortion measures used for speech recognition. Within this framework, three general categories of information-theoretic distortion measures are introduced: the generalized Kolmogorov variational distance, the f -divergence, and the Chernoff distance. There are two major results of this investigation. First, it is found that most of the important distortion measures used by workers in speech recognition fall out as a special case of one or another of the classes of probability-distribution dissimilarity measures. Second, the information-theoretic perspective adopted makes it possible to discover new distortion measures which may display superior speech recognition performance; one measure, the clamped log (cos β) distance, has been investigated experimentally, with promising results