Rapid Match Training for Large Vocabularies

RAPID MATCH TRAINING FOR LARGE VOCABULARIES Larry Gillick, Barbara Peskin, and Robert Roth Dragon Systems, Inc. 320 Nevada Street Newton, Massachusetts 02160 ABSTRACT This paper describes a new algorithm for building rapid match models for use in Dragon's continuous speech recog- nizer. Rather than working from a single representative token for each word, the new procedure works directly from a set of trained hidden Markov models. By simulated traversals of the HMMs, we generate a collection of sample tokens for each word which are then averaged together to build new rapid match models. This method enables us to construct models which better reflect the true variation in word occur- rences and which no longer require the extensive adaptation needed in our original method. In this preliminary report, we outline this new procedure for building rapid match mod- els and report results from initial testing on the Wall Street Journal recognition task. 1. INTRODUCTION In this paper, we report on a new algorithm for build- ing rapid match (prefiltering) models for Dragon's con- tinuous speech recognizer. The rapid match module is intended to supply the recognizer with a relatively short list of word candidates at every point where the recog- nizer hypothesizes a new word may begin. To accomplish this, the rapid match module performs a quick but very approximate calculation using a short interval of acoustic data- usually no more than 240 milliseconds of speech - and passes on to the recognizer a list of word candidates which can then be analyzed in detail. When the rapid match module for Dragon's continuous speech recognizer was first presented nearly two years ago [1], we evaluated its performance on a test corpus of mammography reports involving a vocabulary of under 1,000 words. At that time, the performance of the mod- ule was more than adequate to meet the demands of this recognition task. But as we move to larger vocabularies, the demands on rapid match have become greater at the same time that its role in recognition has become more crucial: if we hope to approach anything like real-time recognition on a large-vocabulary task using moderately priced personal computers, the recognizer can entertain *This work was sponsored by the Defense Advanced Research Projects Agency and was monitored by the Space and Naval War- fare Systems Command under contract N00039-86-C-0307. word hypotheses for only a tiny fraction of its complete vocabulary. Thus, not only must prefiltering provide models for more words, but those models must be capa- ble of making finer distinctions. Until now, we had been generating rapid match models based on a single artificially constructed token represent- ing the "average" behavior of each word. But working from a single token made it impossible to adequately model potential variability, and extensive adaptation of the models was necessary both to estimate variances and to adjust model parameters to new speakers. In our new training procedure, we instead build word models directly from hidden Markov models for each speaker's vocabulary. As reported below, these …

Content truncated to 500 words