SPEECH RECOGNITION DEVICE, METHOD THEREOF, AND PROGRAM

PROBLEM TO BE SOLVED: To provide a speech recognition device which re-calculates recognition scores by using long-distance contextual information being not across a topic boundary.SOLUTION: A topic boundary detection unit receives N best candidate strings to extract a recognition result candidate in...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: MASATAKI HIROKAZU, ASAMI TAICHI, YAMAGUCHI YOSHIKAZU, KOBASHIGAWA SATORU, NOMOTO NARIHISA
Format: Patent
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:PROBLEM TO BE SOLVED: To provide a speech recognition device which re-calculates recognition scores by using long-distance contextual information being not across a topic boundary.SOLUTION: A topic boundary detection unit receives N best candidate strings to extract a recognition result candidate in the first place located in a heading direction of an audio document, as a past spoken word set from a current speaking section in the N best candidate strings with the current speaking section as the center and to extract a recognition result candidate in the first place located in a trailing direction of the audio document, as a future spoken word set from the current speaking section with the current speaking section as the center, and compares an average value of degrees of association between past and future with a prescribed threshold to determine a topic boundary of the audio document. A unit for re-calculating scores of the N best candidates re-calculates recognition scores of the current speaking section which have not determined to be a topic boundary by the topic boundary detection unit, into a value using an average value of a total value of degrees of association between past and present and a total value of degrees of association between present and future, and performs processing of rearranging the N best candidates in the current speaking section.