Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
This paper introduces a large vocabulary isolated words speech recognition system for Chinese speech based on HMM. modelling of phonemes.Through the concatenation of phonemes a word HMM model can be formed in training stage,in this case the coarticuiation of connected utterance is preserved in the mode Lin recognition phase, Syllable String Network(SSN) without any lexical constraint outputs one or two phonetic strings with different length,then every word in library is matched and scored comparing with these owe or two strings, faking 50-100 candidates,rebuildrag lexical tree,a less pruning Viterbi Beam Search(VBS) is applied to get final resulLThe system achieves 86% recognition rate for top-1 and 94% for top 5.A concept of using pre-computcd confusion matrices is proposed for phoneme string mathcing in this paper.Also the way of estimation of these matrices is provided.
Bibliographic reference. Xu, Bo / Lin, Z. W. / Huang, Taiyi / Xu, D. X. / Gao, Y. Q. (1992): "A. 46,500 word Chinese speech recognition system", In ICSLP-1992, 1563-1566.