A continuous Putonghua recognizer

Abstract

A multi-speaker continuous Putonghua recognizer has been developed composing of 20 speaker-dependent recognizer as sub-systems. Each sub-system is a network of hidden Markov models modeling triphones as the fundamental speech units. Over 3 GB of speech data have been collected for training from twenty native Putonghua speakers reading carefully designed tests trying to include all phone-to-phone transitions in Putonghua. A Viterbi path search yields the best speech unit sequence over the HMMnet for each unknown input utterance which is then passed down to a language model for post-processing. The most suitable word sequence is determined by means of the bigram statistics of 470 word classes covering a vocabulary of over 80,000 words. An enrollment process is required for each new user to select the most suitable speaker-dependent system among the 20 sub-systems according to their recognition performance on a small quantity of speech data collected from the user.