Continuous speech recognition with the connectionist Viterbi training procedure: a summary of recent work

Author

Franzini, Michael ; Waibel, Alex ; Lee, Kai-Fu

Author_Institution

Telefonica Investigacion y Desarrollo, Madrid, Spain

fYear

1991

fDate

18-21 Nov 1991

Firstpage

1855

Abstract

Various means by which hidden Markov models (HMMs) and neural networks (NNs) can be combined for continuous speech recognition are studied. The authors describe the connectionist Viterbi training (CVT) procedure, discuss the factors most important to its design, and report its recognition performance. Several changes made to the system are reported, including: (1) the change from recurrent to non-recurrent NNs, (2) the change from Sphinx-style phone-based HMMs to word-based HMMs, (3) the addition of a corrective training procedure, and (4) the addition of an alternate model for every word. The CVT system incorporating these changes achieved 99.1% word accuracy and 98.0% string accuracy on the TI/NBS connected digits task

Keywords

Markov processes; neural nets; speech recognition; Sphinx-style phone-based HMMs; TI/NBS connected digits task; connectionist Viterbi training; continuous speech recognition; corrective training; hidden Markov models; neural networks; nonrecurrent neural nets; recognition performance; recurrent neural nets; string accuracy; word accuracy; word-based HMMs; Computer networks; Computer science; Distributed computing; Hidden Markov models; Maximum likelihood estimation; NIST; Neural networks; Rails; Speech recognition; Viterbi algorithm;

fLanguage

English

Publisher

ieee

Conference_Titel

Neural Networks, 1991. 1991 IEEE International Joint Conference on

Print_ISBN

0-7803-0227-3

Type

conf

DOI

10.1109/IJCNN.1991.170628

Filename

170628