摘要 |
PROBLEM TO BE SOLVED: To attain the shortening of a recognition time by determining a recognized result at the time when a linguistic model expresses a single sentence by a linguistic restriction in a stage in which likelihoods with which the series of acoustic modes appear are calculated to reduce the computational complexity required for the recognition without deteriorating the recognition rate of a voice recognition. SOLUTION: An acoustic analysing part 1 performs the acoustic analysis of an input voice A, that is, converts it into the time series, etc., of acoustic parameters such as cepstrum,Δcepstrum and so forth, to output the converted time series of the acoustic parameters, etc., to a recognition processing part 2. The recognition processing part 2 calculates likelihoods (probabilities) with which the series of acoustic models B from the time series of the acoustic parameters of the input voice A and judges whether the series of the acoustic model B having the highest likelihood is a single sentence or not in linguistic models C which are stipulated by linguistic restrictions D (a disctionaly, a grammer) and when series of the acoustic series B is the single sentence and the likelihood of the B is highest at the point of time when several frame are continued, he processing part 2 outputs the linguistic models C as a recognized result E.
|