Speech recognition device, the weight vector learning device, voice recognition method, the weight vector learning methods, program

摘要

PROBLEM TO BE SOLVED: To optimize a speech recognition parameter using a WFST network expression.;SOLUTION: A speech recognition device includes a recording part, a WFST synthesis part, a feature quantity extraction part, a WFST-type log-linear decoder, and an output symbol extraction part. The recording part records a pronunciation dictionary model, a language model, a sound model, and a weight vector α. The WFST synthesis part synthesizes the pronunciation dictionary model, the language model, and the sound model, and outputs a WFST network. The WFST-type log-linear decoder expresses a score W (X, A) of an arc series A in the log domain when a time series of a feature quantity vector is given, by linear representation of a feature vector ϕ (X, A) acquired from the time series X of the feature quantity vector and the arc series A and the weight vector α, and outputs an arc series of the highest score. The output symbol extraction part determines a word sequence to the arc series and outputs it.;COPYRIGHT: (C)2011,JPO&INPIT