SPEECH RECOGNITION BY STATISTICAL METHODS.

Abstract

Summary form only given, as follows. The Speech Recognition Group at the IBM Thomas J. Watson Research Center has recently completed the implementation of a real-time, IBM-PC based, large vocabulary speech recognition system that can be used for dictating office correspondence. To make real-time performance possible, the vocabulary is restricted to 5000 words that must be spoken with short pauses between them. The system is based on information theoretical (rather than expert system) formulation of the recognition problem. The algorithms used are statistical and all the parameters of the system are estimated directly from data. The basic theory behind the recognizer is presented with special emphasis on the language model component which provides the probability of the next word, given the past hypothesis.