APPARATUS FOR CORRECTING ERROR OF SPEECH RECOGNITION

摘要

A voice recognition error correction device of the present invention comprises a feature vector extraction part extracting a feature vector from received voice; a voice recognition part recognizing the received voice as word strings on the basis of the extracted feature vector; a phoneme weighted finite state transducer (WFST) conversion part converting the recognized word strings by the voice recognition part into a phoneme WFST; and a voice recognition error correction part correcting errors for the converted phoneme WFST. The voice recognition error correction part comprises a WFST composition part modeling pronunciation variation on the basis of a kullback-leibler (KL) distance matrix for the phoneme WFST transmitted from the phoneme WFST conversion part. [Reference numerals] (110) Feature vector extracting unit;(120) Sound recognition unit;(130) Phoneme WFST converting unit;(200) Sound recognition error compensating unit;(310) Sound model;(320) Pronunciation dictionary;(330) Language model;(340) Pronunciation modification model;(AA) Sound input signal;(BB) Feature vector;(CC) Recognized word line;(DD) Phoneme WFST;(EE) Error corrected word line