Speaker Independent Phonetic Transcription of Fluent Speech for Large Vocabulary Speech Recognition

SPEAKER INDEPENDENT PHONETIC TRANSCRIPTION OF FLUENT SPEECH FOR LARGE VOCABULARY SPEECH RECOGNITION S. E. Levinson M. Y. Liberman A. Ljolje and L. G. Miller AT&T Bell Laboratories Murray Hill, New Jersey 07974 ABSTRACT Speaker independent phonetic Iranscription of fluent speech is performed using an ergodic continuously variable duration hidden Markov model (CVDHMM) to represent the acoustic, phonetic and phonotactic structure of speech. An important property of the model is that each of its fifty-one states is uniquely identified with a single phonetic unit. Thus, for any spoken utterance, a phonetic transcription is obtained from a dynamic programming (DP) procedure for finding the state sequence of maximum likelihood. A model has been constructed based on 4020 sentences from the TIMIT database. When tested on 180 different sentences from this database, phonetic accuracy was observed to be 56% with 9% insertions. A speaker dependent version of the model was also constructed. The transcription algorithm was then combined with lexical access and parsing routines to form a complete recognition system. When tested on sentences from the DARPA resource management task spoken over the local switched telephone network, phonetic accuracy of 64% with 8% insertions and word accuracy of 87% with 3% insertions was measured. This system is presently operating in an on-line mode over the local switched telephone network in less than ten times real time on an Alliant FX-80. INTRODUCTION Though rarely explicitly stated, a fundamental assumption on which many speech recognition systems are implicitly based is that speech is literate. That is, it is a code for communication having a small number of discrete phonetic symbols in its alphabet. These symbols are, however, merely mental constructs and, as such, are not directly accessible but are, instead, observable only in their highly variable acoustic manifestation. It is also well-known but equally seldom expressed that a hidden Markov model comprises a finite set of discrete inaccessible states observable only via a set of random processes, one associated with each hidden state. When these two simple ideas are juxtaposed, it seems to us inescapable that the most natural representation of speech by a hidden Markov model is one in which the hypothetical phonetic symbols are identified with the hidden states of the Markov chain and the variability of the measurable acoustic signal is captured by the observable, state-dependent random processes. 75 The mathematical details of just such a model are given in [6]. Its application to a small- vocabulary continuous speech recognition system and a large-vocabulary isolated word recognition system are described in [7] and [8], respectively. Here we present a brief overview of the use of this approach in large vocabulary continuous speech recognition and some preliminary results of two experiments performed with it on the TIMIT [4] and DARPA [9] databases. THE MODEL We have constructed two models, a 51 state model on which the speaker-independent phonetic transcription results are based, and a 43 state model on which the speaker-dependent recognition of sentences from the DARPA resource management task are founded. The …

Content truncated to 500 words