Markov Models for Phonetic Text

Abstract

The classical statistical models for phonetic text view phonemes as states of a Markov chain of some low order. There is a model-building method which produces models that lie between these classical models and the combinatorial model of O'Connor and Trimm. To produce the new models, one hypothesizes the existence of an underlying Markov chain of low order with some small number n of states, and assumes that the probability distribution of phonetic symbols is a function of the state of the underlying chain. Given n, the model-building method involves a computer hill-climbing procedure on a corpus of phonetic text to find the maximum-likelihood model for that text and that n. This paper describes the method and the results of a large number of ascents on a certain corpus of text. The models developed show strong linguistic features; for example, for text segmented only by silences between phoneme strings, the vowels, consonants, and boundary symbols are statistically identified with 15% errors. With text segmented into syllables, identification is essentially perfect.