Phoneme-in-Context Modeling for Dragon’s Continuous Speech Recognizer

Phoneme-in-Context Modeling for Dragon's Continuous Speech Recognizer Paul Bamberg and Laurence Gillick Dragon Systems, Inc. 90 Bridge SL, Newton MA 02158 Abstract For large-vocabulary continuous speech recognition, the goal of training is to model phonemes with enough precision so that from the models one could reconstruct a sequence of acoustic parameters that accurately represents the spectral characteristics of any naturally-occurring sentence, including all coarticuladon effects that arise either between phonemes in a word or across word boundaries. The aim at Dragon Systems is to collect and process enough training data to accomplish this goal for all of natural spoken English rather than for any one restricted task. The basic unit that must be trained is the "phoneme in context" (PIC), a sequence of three phonemes accompanied by a code for prepausal lengthening. At present, syllable and word boundaries are ignored in defining PICs. More than 16,000 training tokens, half isolated words and half short phrases, were phonemically labeled by a semi- . automatic procedure using hidden Markov models. To model a phoneme in a specific context, a weighted average is constructed from training data involving the desired context and acoustically similar contexts. For use in HMM continuous-speech recognition, each PIC is converted to a Markov model that is a concatenation of one to six node models. No phoneme, in all its contexts, requires more than 64 distinct nodes, and the total number of node models ("phonemic segments") required to construct all PICs is only slightly more than 2000. As a result, the entire set of PICs can be adapted to a new speaker on the basis of a couple of thousand isolated words or a few hundred sentences of connected speech. The advantage of this approach to training is that it is not task-specific. From a single training database, Dragon Systems has constructed models for use in a 30,000-word isolated-word recognizer, for connected digits, and for two different thousand-word continuous-speech tasks. 1. Introduction The nature of the training process for a speech- recognition system changes radically once the size of the vocabulary becomes larger than the number of words for which a user is willing to provide training tokens. Below this threshold, it is reasonable to make an independent model for each word in the vocabulary. Such a model, based on data from that word and no others, can in principle capture all the acoustic-phonetic subtleties of the word, even though the phonetic spelling of the word is not even used in constructing the model. For continuous speech recognition, the quantity of data required for complete training grows much more rapidly than vocabulary. In the simple case of a recognizer for three-digit strings, for example, each digit should at a minimum be trained in initial, medial, and final position, while for optimum performance all digit triples should be included in the training data. The approach to training at Dragon Systems has been to regard the recognition task as all of natural English, whether isolated words or connected speech. We have …

Content truncated to 500 words