A linguistic processor for Japanese continuous speech recognition

This paper presents a linguistic processor for Japanese continuous speech recognition with a large vocabulary. The linguistic processor receives a syllable lattice which is produced by an acoustic processor. It searches the lattice for an optimal word sequence which satisfies the linguistic constraints. The linguistic knowledge prepared is task-unspecified and consists of a 5000 word dictionary, a "part of speech" network and a word connection table. A best-first search method is used to search the lattice more efficiently. The evaluation experiments were made using 200 phrase utterances made by four talkers, totaling 800 phrases, spoken continuously. The results showed that the phrase recognition score for the first ranking was 71.5% on an average. The cumulated score up to the 10th ranking was 90.1% on an average.