The currently dominant speech recognition technology, hidden Markov
modeling, has long been criticized for its simplistic assumptions
about speech, and especially for the naive Bayes combination rule
inherent in it. Many sophisticated alternative models have been suggested
over the last decade. These, however, have demonstrated only modest
improvements and brought no paradigm shift in technology. The goal
of this paper is to examine why HMM performs so well in spite of its
incorrect bias due to the naive Bayes assumption. To do this we create
an algorithmic framework that allows us to experiment with alternative
combination schemes and helps us understand the factors that influence
recognition performance. From the findings we argue that the bias
peculiar to the naive Bayes rule is not really detrimental to phoneme
classification performance. Furthermore, it ensures a consistent behavior
in outlier modeling, allowing the efficient management of insertion
and deletion errors.
