Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Large Vocabulary Continuous Speech Recognition Using a Hybrid Connectionist-HMM System

Mike M. Hochberg, Steve J. Renals, A. J. Robinson, D. J. Kershaw

Cambridge University Engineering Department, Cambridge, England, UK

Abbot is a hybrid connectionist-hidden Markov model (HMM) system for large vocabulary speech recognition which participated in the November 1993 ARPA Wall Street Journal benchmark tests. This system uses a recurrent network to estimate the acoustic observation probabilities within the HMM framework. Since the 1993 benchmark tests, a number of improvements have been made to the ABBOT system. These improvements have been gained through better phone-duration modeling and connectionist model combination. In addition, ABBOT has been extended to handle large vocabulary tasks with a trigram language model. Fast decoding is obtained using a pruning strategy particularly well-suited for the hybrid approach. This paper describes the recent modifications to the system and experimental results are reported for various test and development sets from the November 1993 ARPA evaluations.

Full Paper

Bibliographic reference.  Hochberg, Mike M. / Renals, Steve J. / Robinson, A. J. / Kershaw, D. J. (1994): "Large vocabulary continuous speech recognition using a hybrid connectionist-HMM system", In ICSLP-1994, 1499-1502.