Sentence understanding of spoken Japanese using phrase spotting and dependency grammar

Abstract

In order to recognize spoken Japanese sentences, a speech understanding system that consists of four stages was constructed. First, the system hypothesizes syllable candidates using an HMM‐based syllable spotter. Second, it concatenates the syllables to obtain word candidates, checking the tree‐structured dictionary. Third, it connects word candidates to get phrase candidates using a Japanese phrase spotter. At this stage, a finite state automaton that represents the Japanese intra‐phrase grammar is used. Finally, the system selects the best phrase sequence as a sentence recognition result using a backward parsing algorithm. This algorithm makes it possible to analyze the syntactic structure of a phrase sequence or a partial sentence using Japanese dependency (kakari‐uke) grammar. Japanese sentences for the recognition task were inquiries on the UNIX system and spoken by six male speakers at natural speed. Syllables extracted from the continuous speech were used as training data for the HMM syllable spotter. The system's performance was evaluated from these speech data. The semantic analysis of this system is also described.