Keyword recognition system and method using template concantenation model

Claim 1

A system for automatically recognizing a keyword in a speech utterance comprising: (a) means for inputting a speech utterance in the form of parametric representations of input speech in sequential frames; (b) filler-template storing means for storing filler templates each indicative of a short speech sound which is not a keyword and is generally shorter than a keyword; (c)keyword-template storing means for storing keyword templates each indicative of a keyword to be recognized; (d) template-concatenation means for statistically matching a sequence of frames of input speech with the filler templates in said filler-template storing means, on the one hand, and the keyword templates in said keyword-template storing means, on the other, and for concatenating a plurality of candidate strings of filler templates, on the one hand, and keyword templates, on the other, each acceptably matching the sequence of input speech, said plurality including candidate strings containing different sequences of filler and keyword templates which test both the hypothesis of a keyword being present and the alternate hypothesis of the input speech more closely matching an acoustic pattern of speech that is similar to the keyword, but does not actually contain the keyword; (e) scoring means for accumulating dissimilarity scores for each of the respective candidate strings representing the dissimilarity of a candidate string from the input speech; and (f) recognition means for determining when an optimal candidate string has a total dissimilarity score which is less than the total dissimilarity scores of the other candidate strings, and for recognizing a keyword in the input speech based upon a corresponding keyword template being contained in the optimal candidate string.

Claim 2

A system for automatically recognizing a keyword in a speech utterance comprising: (a) means for inputting a speech utterance in the form of parametric representations of input speech in sequential frames; (b) filler-template storing means for storing filler templates each indicative of a short speech sound which is not a keyword and is generally shorter than a keyword; (c)keyword-template storing means for storing keyword templates each indicative of a keyword to be recognized; (d) template-concatenation means for statistically matching a sequence of frames of input speech with the filler templates in said filler-template storing means, on the one hand, and the keyword templates in said keyword-template storing means, on the other, and for concatenating a plurality of candidate strings of filler templates, on the one hand, and keyword templates, on the other, each acceptably matching the sequence of input speech, said plurality including candidate strings containing different sequences of filler and keyword templates which test both the hypothesis of a keyword being present and the alternate hypothesis of the input speech more closely matching an acoustic pattern of speech that is similar to the keyword, but does not actually contain the keyword; (e) scoring means for accumulating dissimilarity scores for each of the respective candidate strings representing the dissimilarity of a candidate string from the input speech; and (f) recognition means for determining when an optimal candidate string has a total dissimilarity score which is less than the total dissimilarity scores of the other candidate strings, and for recognizing a keyword in the input speech based upon a corresponding keyword template being contained in the optimal candidate string.

Claim 7

A system according to claim wherein said scoring means includes means for adding a penalty to the partial string scores of the candidate strings for each template added to a respective string in order to bias the score in favor of a string having a lower number of templates.

Claim 12

A method for automatically recognizing a keyword in a speech utterance comprising the steps of: (a) inputting a speech utterance in the form of parametric representations of input speech in sequential frames; (b) storing filler templates each indicative of a short speech sound which is not a keyword and is generally shorter than a keyword; (c) storing keyword templates each indicative of a keyword to be recognized; (d) statistically matching a sequence of frames of input speech with the stored filler templates, on the one hand, and the stored keyword templates, on the other and concatenating a plurality of candidate strings of filler templates, on the one hand, and keyword templates, on the other, each acceptably matching the sequence of input speech, said plurality including candidate strings containing different sequences of filler and keyword templates which test both the hypothesis of a keyword being present and the alternate hypothesis of the input speech more closely matching an acoustic pattern of speech that is similar to the keyword, but does not actually contain the keyword; (e) accumlating dissimilarity scores for each of the respective candidate strings representing the dissimilarity of a candidate string from the input speech; and (f) determining when an optimal candidate string has a total dissimilarity score which is less than the total dissimilarity scores of the other candidate strings, and recognizing a keyword in the input speech based upon a corresponding keyword template being contained in the optimal candidate string.

Claim 13

A method for automatically recognizing a keyword in a speech utterance comprising the steps of: (a) inputting a speech utterance in the form of parametric representations of input speech in sequential frames; (b) storing filler templates each indicative of a short speech sound which is not a keyword and is generally shorter than a keyword; (c) storing keyword templates each indicative of a keyword to be recognized; (d) statistically matching a sequence of frames of input speech with the stored filler templates, on the one hand, and the stored keyword templates, on the other and concatenating a plurality of candidate strings of filler templates, on the one hand, and keyword templates, on the other, each acceptably matching the sequence of input speech, said plurality including candidate strings containing different sequences of filler and keyword templates which test both the hypothesis of a keyword being present and the alternate hypothesis of the input speech more closely matching an acoustic pattern of speech that is similar to the keyword, but does not actually contain the keyword; (e) accumlating dissimilarity scores for each of the respective candidate strings representing the dissimilarity of a candidate string from the input speech; and (f) determining when an optimal candidate string has a total dissimilarity score which is less than the total dissimilarity scores of the other candidate strings, and recognizing a keyword in the input speech based upon a corresponding keyword template being contained in the optimal candidate string.