| 摘要 |
PURPOSE: To attain efficient processing suitable for Japanese language structure by advancing processing from the end of a sentence to its head and finding out a clause candidate group in the case of obtaining a clause group to be a recognized result from a clause lattice obtained from a speech signal. CONSTITUTION: In order to obtain a clause group to be a recognized result from a clause lattice obtained from a speech signal, the inspection of consistency between clauses is successively executed in a direction from conjugation on the end of a sentence to the head of the sentence. The clause lattice means a set of estimated clause candidates including timewise position information. Thereby the clause lattice is processed as a graph regarding each clause candidate and adjacent relation between clauses respectively as a node and a branch and a recognition result is obtained as a sequence nodes forming a route from the start end position of a voice up to its final end position. The search of a sentence is executed from its end side to the head side and tree search or route search is executed under control based upon clause estimation validity and a language restriction condition. Consequently efficient processing suitable for language structure peculiar to Japanese can be attained. |
