Minimum error rate training of combined string models

Claim 3

The method of claim 1 wherein the step of generating the set of one or more confusable string models comprises generating N-best word string models.

Claim 4

The method of claim 1 wherein the first scoring signal reflects a measure of similarity between the training speech utterance signal and the string model for that utterance.

Claim 5

The method of claim 3 wherein the measure of similarity comprises a log-likelihood recognizer score.

Claim 6

The method of claim 1 wherein the second scoring signal reflects a measure of similarity between the training speech utterance signal and one of said confusable string models.

Claim 7

The method of claim 5 wherein the measure of similarity comprises a log-likelihood recognizer score.

Claim 8

The method of claim 1 wherein the step of generating a signal reflecting a comparison comprises forming a difference of the first scoring signal and a combination of the one or more second scoring signals.

Claim 9

The method of claim 1 wherein the step of modifying the one or more of said current speech recognition models comprises the steps of: 1. generating a recognition model modification signal reflecting a gradient of a function, the function reflecting a recognizer score of a training speech utterance based on a string model for that utterance and one or more recognizer scores of the training speech utterance based on the one or more confusable string models; and 2. adapting the one or more of said current speech recognition models based on the modification signal.

Claim 10

The method of claim 8 wherein the function reflects a difference of a recognizer score of a training speech utterance based on a string model for that utterance and a weighted sum of the one or more recognizer scores of the training speech utterance based on one or more confusable string models.

Claim 11

The method of claim 1 wherein one of said sets of current speech recognition models comprises acoustic models.

Claim 12

The method of claim 10 wherein the acoustic models comprise hidden Markov models.

Claim 13

The method of claim 1 wherein one of said sets of current speech recognition models comprises language models.

Claim 14

The method of claim 1 wherein one of said sets of current speech recognition models comprises pitch models.

Claim 15

The method of claim 1 wherein one of said sets of current speech recognition models comprises energy models.

Claim 16

The method of claim 1 wherein one of said sets of current speech recognition models comprises speaking rate models.

Claim 17

The method of claim 1 wherein one of said sets of current speech recognition models comprises duration models.

Claim 18

A speech recognizer trainer for making a speech recognition model database based on a training speech utterance signal and a plurality of sets of current speech recognition models, the trainer comprising a. means for generating a set of one or more confusable string models, a confusable string model comprising a plurality of said current speech recognition models from two or more of said sets of said current speech recognition models, wherein each of at least two of said sets of models corresponds to different characteristics of speech and wherein the Confusable string model is a model which, if selected to represent the training speech utterance, would result in a misrecognition of the training speech utterance; b. means for generating a first scoring signal based on the speech utterance signal and a string model for that utterance, the string model for the utterance comprising a plurality of said current speech recognition models from said two or more of said sets of said current speech recognition models; c. means for generating one or more second scoring signals, a second scoring signal based on the training speech utterance signal and a confusable string model; d. means for generating a signal reflecting a comparison of a likelihood of correctly recognizing the training speech utterance and a likelihood of incorrectly recognizing the training speech utterance, said likelihood of correctly recognizing the training speech utterance based on the first scoring signal and said likelihood of incorrectly recognizing the training speech utterance based on the one or more second scoring signals; and e. means, responsive to the signal reflecting the comparison of said likelihoods, for modifying one or more of said current speech recognition models to increase the probability that a string model for the utterance will have a rank order higher than the confusable string models.

Claim 19

A speech recognizer trainer for making a speech recognition model database based on a training speech utterance signal and a plurality of sets of current speech recognition models, the trainer comprising a. means for generating a set of one or more confusable string models, a confusable string model comprising a plurality of said current speech recognition models from two or more of said sets of said current speech recognition models, wherein each of at least two of said sets of models corresponds to different characteristics of speech and wherein the Confusable string model is a model which, if selected to represent the training speech utterance, would result in a misrecognition of the training speech utterance; b. means for generating a first scoring signal based on the speech utterance signal and a string model for that utterance, the string model for the utterance comprising a plurality of said current speech recognition models from said two or more of said sets of said current speech recognition models; c. means for generating one or more second scoring signals, a second scoring signal based on the training speech utterance signal and a confusable string model; d. means for generating a signal reflecting a comparison of a likelihood of correctly recognizing the training speech utterance and a likelihood of incorrectly recognizing the training speech utterance, said likelihood of correctly recognizing the training speech utterance based on the first scoring signal and said likelihood of incorrectly recognizing the training speech utterance based on the one or more second scoring signals; and e. means, responsive to the signal reflecting the comparison of said likelihoods, for modifying one or more of said current speech recognition models to increase the probability that a string model for the utterance will have a rank order higher than the confusable string models.

Claim 20

The trainer of claim 17 wherein the means for generating a signal reflecting a comparison comprises means for forming a difference between the first scoring signal and a combination of the one or more second scoring signals.

Claim 21

The trainer of claim 17 wherein the means for modifying the one or more of said current speech recognition models comprises: 1. means for generating a recognition model modification signal reflecting a gradient of a function, the function reflecting a recognizer score of a training speech utterance based on a string model for that utterance and one or more recognizer scores of the training speech utterance based on the one or more confusable string models; and 2. means for adapting the one or more of said current speech recognition models based on the modification signal.

Claim 22

A speech recognition system comprising: a. a feature extractor for receiving an unknown speech signal and generating feature signals characterizing the unknown speech signal; b. a memory for storing a plurality of sets of speech recognition models, one or more of the speech recognition models generated in accordance with a process of modifying parameters of predetermined speech recognition models to enhance the probability that a correct string model will have a rank order higher than one or more confusable string models, wherein each of at least two of said sets of speech recognition models corresponds to different characteristics of speech and wherein said confusable string models are models which, if selected to represent a training speech utterance, would result in a misrecognition of the training speech utterance, and wherein the modification of said parameters is based on a comparison of a likelihood of correctly recognizing a training speech utterance and a likelihood of incorrectly recognizing the training speech utterance, the likelihood of correctly recognizing the training speech utterance based on a first scoring signal and the likelihood of incorrectly recognizing the training speech utterance based on one or more second scoring signals, the first scoring signal generated based on the training speech utterance and a string model for that utterance, the string model for the utterance comprising a plurality of said speech recognition models from said at least two of said sets of said speech recognition models, and each of the second scoring signals generated based on the training speech utterance signal and one of said confusable string models; and c. a scoring processor, coupled to the feature extractor and the memory, for comparing a string model with features of the unknown speech signal, the string model comprising one or more of said speech recognition models from each of the plurality of said sets of speech recognition models, and for recognizing the unknown speech signal based on one of a plurality of string models which most closely compares with the features of the unknown speech signal.

Claim 23

A speech recognition system comprising: a. a feature extractor for receiving an unknown speech signal and generating feature signals characterizing the unknown speech signal; b. a memory for storing a plurality of sets of speech recognition models, one or more of the speech recognition models generated in accordance with a process of modifying parameters of predetermined speech recognition models to enhance the probability that a correct string model will have a rank order higher than one or more confusable string models, wherein each of at least two of said sets of speech recognition models corresponds to different characteristics of speech and wherein said confusable string models are models which, if selected to represent a training speech utterance, would result in a misrecognition of the training speech utterance, and wherein the modification of said parameters is based on a comparison of a likelihood of correctly recognizing a training speech utterance and a likelihood of incorrectly recognizing the training speech utterance, the likelihood of correctly recognizing the training speech utterance based on a first scoring signal and the likelihood of incorrectly recognizing the training speech utterance based on one or more second scoring signals, the first scoring signal generated based on the training speech utterance and a string model for that utterance, the string model for the utterance comprising a plurality of said speech recognition models from said at least two of said sets of said speech recognition models, and each of the second scoring signals generated based on the training speech utterance signal and one of said confusable string models; and c. a scoring processor, coupled to the feature extractor and the memory, for comparing a string model with features of the unknown speech signal, the string model comprising one or more of said speech recognition models from each of the plurality of said sets of speech recognition models, and for recognizing the unknown speech signal based on one of a plurality of string models which most closely compares with the features of the unknown speech signal.