Integrated Processing of Linguistic and Speech Information
Hozumi TANAKA, Takenobu TOKUNAGA and Hui LI
Department of Computer Science, Tokyo Institute of Technology
2-12-1 Ookayama Meguro Tokyo 152 JAPAN
e-mail: tanaka@cs.titech.ac.jp
It is obvious that successful speech recognition requires the use of
linguistic information. For this purpose, a generalized LR (GLR)
parser provides an exceptionally competent and flexible framework to
combine linguistic information with phonological information.
The combination of a GLR parser and allophone models is considered
very effective for enhancing the recognition accuracy in a large
vocabulary continuous speech recognition. The main problem of
integrating GLR parsing into an allophone-based recognition system is
how to solve the word juncture problem, that is, how to express
the phones at a word boundary with allophone models.
This paper proposes a new method called CPM ( Constraint Propagation
Method ) to generate an allophone-based LR table, which can
effectively solve the word juncture problem. In our method, by
introducing the allophone rules into the CFG and lexical rules, an
LR table is generated, then the LR table is modified on the
basis of an allophone connection matrix by applying the constraint
propagation method. With this modified LR table, precise allophone
predictions for speech recognition can be obtained.
Keywords: canonical LR table, GLR, allophone, constraints propagation, speech recognition