A probabilistic model for the performance of word recognizers

This paper develops a probabilistic model to account for the error-rate behavior of isolated-word speech-recognition systems. It examines two kinds of errors, confusion error, an a priori characterization of a recognizer, which measures differences between words, and recognition (rank) error, an a posteriori characterization, which, in addition to taking into account differences between words, accounts for differences between different tokens of the same word. It is shown that these kinds of errors can be modeled by describing recognition trials as Bernoulli trials. Good models of error-rate behavior as a function of vocabulary size can be obtained if the distributions of confusion and recognition (rank) number are considered to be mixtures of binomial distributions. The data obtained from a recent experiment in isolated-word recognition with a large vocabulary (1109 words) are used to evaluate the model. Experimental error-rate functions obtained from each of six talkers and three partitions of the vocabulary are fit by means of an optimization algorithm to model functions based on mixture distributions. The results indicate that two-way mixture distributions account quite well for the experimental performance results.