ABSTRACT
This chapter provides an introduction to statistical methods for automatic speech recognition. Certain symbols have come to be conventionally associated with particular quantities, although there is still some variation in the details of the notation that is used. The probability of the observations, given the model, is made up of contributions from a very large number of alternative state sequences. However, the probability distributions associated with the states will be such that the probability of the observed feature vectors having been produced by many of the state sequences will be microscopically small compared with the probabilities associated with other state sequences. The probabilities of occupying the states can then be taken into account when gathering the statistics of state sequences and of feature vectors associated with the states, in order to obtain new estimates for the transition probabilities and for the emission probabilities respectively.
