Dialogue speech recognition using syntactic rules based on subtrees and preterminal bigrams

Abstract

We are studying spontaneous speech recognition in building speech translation and/or spoken dialogue systems. It is desirable for speech recognizers of such types to be portable for some domains and/or tasks. Therefore, we are examining a CFG-based speech parser that outputs scored subtree sequences. First, we design and develop syntactic rules that represent pause units, i.e., segments separated by pause information, as subtrees. In order to deal with spontaneous speech, the syntactic rules are written to parse pause units rather than sentences. The developed grammar is sufficient for dealing with almost all utterances in our travel conversation database. We also explain the bigrams of preterminal symbols. Next, we propose a dialogue speech recognition method with subtree-based syntactic rules and preterminal bigrams. Speech recognition experiments using our travel conversation database confirm that our method yields good results. It has better performance in using both syntactic rules and preterminal bigrams than that obtained by using syntactic CFG rules alone or preterminal bigrams alone.