VOICE RECOGNITION DEVICE AND COMPUTER PROGRAM
PROBLEM TO BE SOLVED: To provide a voice recognition device capable of increasing recognition accuracy by utilizing properties of a neural network.SOLUTION: A voice recognition device includes: an acoustic model 308 by an RNN (recurrent type neural network) calculating posterior probabilities of a s...
Saved in:
Main Author | |
---|---|
Format | Patent |
Language | English Japanese |
Published |
22.12.2016
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | PROBLEM TO BE SOLVED: To provide a voice recognition device capable of increasing recognition accuracy by utilizing properties of a neural network.SOLUTION: A voice recognition device includes: an acoustic model 308 by an RNN (recurrent type neural network) calculating posterior probabilities of a state series when an observation series composed of a prescribed voice feature amount obtained from voice signals is given; a WFST 320 by SHCLG calculating posterior probabilities of word strings when the state series is given with respect to respective word strings; and a hypothesis selection part 322 for performing voice recognition of a voice signal on the basis of scores calculated for each hypothesis of a word string corresponding to a voice signal by using the posterior probabilities calculated respectively by the acoustic model 308 and the WFST 320 with respect to input observation series.SELECTED DRAWING: Figure 7
【課題】ニューラルネットワークの特性を活かして、認識精度を高めることができる音声認識装置を提供する【解決手段】音声認識装置は、音声信号から得られた所定の音声特徴量からなる観測系列が与えられたときの状態系列の事後確率を状態系列ごとに算出するRNN(リカレント型ニューラルネットワーク)による音響モデル308と、状態系列が与えられたときの単語列の事後確率を各単語列について算出するS−1HCLGによるWFST320と、入力観測系列について音響モデル308及びWFST320がそれぞれ算出する事後確率を用いて、音声信号に対応する単語列の仮説ごとに算出されたスコアに基づいて、音声信号に対する音声認識を行うための仮説選択部322とを含む。【選択図】図7 |
---|---|
Bibliography: | Application Number: JP20150104336 |