Reference pattern learning system
AssigneeNEC Corporation
InventorKen-ichi Iso
Life of this patent
Estimating the term…
Abstract
A first parameter set constituting reference patterns of each category in speech recognition based on pattern matching with a reference pattern is to be determined from a plurality of learning utterance data. The first parameter set is determined so that a third evaluation function, represented by a sum of a first evaluation function and a second evaluation function is maximized. The first evaluation function represents a matching degree between all learning utterances and corresponding reference patterns. The second evaluation function represents a matching degree between elements of the first parameter set.
Claims7
A method of generating reference patterns for a new speaker in a speech recognition system including a plurality of existing speakers, where the new speaker's speech will be compared with reference patterns obtained in advance from each of said plurality of existing speakers in order to recognize the new speaker's speech, the method comprising the steps of:
- (a) generating a plurality of sets of speech recognition reference pattern data wherein each of said sets contain data representing more than one word, one set of reference pattern data is generated for each of said plurality of existing speakers by using speech utterance data from each of said plurality of existing speakers;
- (b) analyzing said plurality of sets of reference pattern data by calculating correlation between parameters of the reference patterns such that correlations are calculated for all sets of phonemes present in the plurality of sets of speech recognition reference pattern data;
- and
- (c) generating a set of reference pattern data for said new speaker based only on the results of the analysis of said analyzing step (b) and on special utterance data from said new speaker which correspond to only a portion of said one set of reference pattern data, said special utterance data being substantially less than said speech utterance data from each of said plurality of existing speakers used in said generating step (a).
Dependencies are inferred from the claim text.