CONTINUOUS VOICE RECOGNITION SYSTEM AND METHOD USING PROBABILISTIC DEPENDENCY GRAMMAR AS BACK-ORIENTED LANGUAGE MODEL

摘要

PURPOSE: A continuous voice recognition system and method using probabilistic dependency grammar as backward oriented language model is to decide a ranking of a document recognition candidate. CONSTITUTION: A characteristic extracting part(11) extracts a characteristic of an input voice. A word recognizing part(12) recognizes the input voice as a word unit from the characteristic extracted by the characteristic extracting part. A sentence recognizing part(13) searches the word recognized by the word recognizing part from the forward direction using a forward directed language model and recognizes words having a lattice structure in which a transfer probability between words is recorded. A sentence re-recognizing part(14) searches the recognized lattice structured word using a probabilistic dependency grammar in which a probability between a modifier and a modifier is recorded as backward oriented language model and re-recognizes the sentence.