Speech-recognition circuitry employing phoneme estimation

Claim 2

For determining a modeling matrix to be used in identifying a predetermined speech event, the speech event being a speech element or a class of speech elements in received speech, a method comprising the steps of: A. dividing speech in a development base of recorded speech into speech segments and labeling the speech segments in accordance with whether the speech event occurred in the speech segment, thereby associating original labels with speech segments; B. computing a preliminary modeling matrix by correlating the speech segments with the original labels associated therewith; C. employing the preliminary matrix to associate predicted labels with the speech signal; D. adjusting the association of at least some of the original labels from the speech elements to which they were previously associated to nearby speech elements so that the original labels whose associations have been adjusted coincide with predicted labels; and E. computing a new modeling matrix by correlating the speech segments with the labels with which they are associated after adjustment.

Claim 5

A method as defined in claim 2 wherein the modeling step comprises generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 6

A method as defined in claim 2 wherein the vector output of the sequence of vector-processing steps that is modeled in the modeling step includes the second-nonlinear-step output vector.

Claim 7

A method as defined in claim 4 wherein the modeling step comprises generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 8

A method as defined in claim 2 wherein the input of at least one of the nonlinear steps includes a plurality of time-interval vectors, each element of one of the time-interval vectors representing a speech time interval the same as that represented by the other elements of the same time-interval vector but different form the speech time interval represented by the elements of a different one of the time-interval vectors, at least one nonlinear step comprising computing nonlinear combinations of elements of different ones of the time-interval vectors to include in its output a separate element representing the result of each such computed combination.

Claim 9

A method as defined in claim 6 wherein the modeling step comprises generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 10

A method as defined in claim 6 wherein the vector output of the sequence of vector-processing steps that is modeled in the modeling step includes the second-nonlinear-step output vector.

Claim 11

A method as defined in claim 8 wherein the modeling step comprises generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 12

A method as defined in claim 6 wherein the second nonlinear step comprises computing the nonlinear combinations of different elements of the first-nonlinear-step output vector by: A. computing linear combinations of the first-nonlinear-step elements; and B. computing nonlinear combinations of different ones of the linear combinations, thereby computing nonlinear combinations of different elements of the first-nonlinear-step output vector.

Claim 13

A method as defined in claim 10 wherein: A. the step of computing linear combinations of the first-nonlinear-step elements comprises computing highest-variance components as the linear combinations, the highest-variance components being the components of the first-nonlinear-step input vector in the directions of the eigenvectors of the first-nonlinear-step input vector's covariance matrix that are associated with the greatest eigenvalues; and B. the step of computing nonlinear combinations of different ones of the linear combinations comprises computing nonlinear combinations of the highest-variance components but omitting computation of the nonlinear combinations of at least some of the components of the first-nonlinear-step input vector in the direction of the eigenvectors of the first-nonlinear-step input vector that are associated with the lowest eigenvalues.

Claim 16

A method as defined in claim 12 wherein the modeling step comprises generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 17

A method as defined in claim 12 wherein the vector output of the sequence of vector-processing steps that is modeled in the modeling step includes the nonlinear-step output vector.

Claim 18

A method as defined in claim 14 wherein the modeling step comprises generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 19

A speech processor adapted for reception of a speech signal and comprising: A. means for converting the speech signal to a sequence of digital quantities; B. means for subjecting the digital quantities to a sequence of digital processing steps including a sequence of vector-processing steps whose outputs are multi-element vectors and whose inputs include vector outputs of previous processing steps, the sequence of vector-processing steps including a first nonlinear step, which comprises computing nonlinear combinations of different ones of its input elements to produce a first-nonlinear-step output vector that includes a separate first-nonlinear-step element representing the result of each such computed combination, and a second nonlinear step, which comprises computing nonlinear combinations of different elements of the first-nonlinear-step output vector to produce a second-nonlinear-step output vector that includes a separate second-nonlinear-step element representing the result of each such computed combination; and C. means for modeling the vector output of the sequence of vector-processing steps to generate speech-element-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective speech elements.

Claim 20

A speech processor adapted for reception of a speech signal and comprising: A. means for converting the speech signal to a sequence of digital quantities; B. means for subjecting the digital quantities to a sequence of digital processing steps including a sequence of vector-processing steps whose outputs are multi-element vectors and whose inputs include vector outputs of previous processing steps, the sequence of vector-processing steps including a first nonlinear step, which comprises computing nonlinear combinations of different ones of its input elements to produce a first-nonlinear-step output vector that includes a separate first-nonlinear-step element representing the result of each such computed combination, and a second nonlinear step, which comprises computing nonlinear combinations of different elements of the first-nonlinear-step output vector to produce a second-nonlinear-step output vector that includes a separate second-nonlinear-step element representing the result of each such computed combination; and C. means for modeling the vector output of the sequence of vector-processing steps to generate speech-element-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective speech elements.

Claim 21

A speech processor as defined in claim 16 wherein the modeling means comprises means for generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 22

A speech processor as defined in claim 16 wherein the vector output of the sequence of vector-processing steps that is modeled in the modeling means includes the second-nonlinear-step output vector.

Claim 23

A speech processor as defined in claim 18 wherein the modeling means comprises means for generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 24

A speech processor as defined in claim 16 wherein the input of at least one of the nonlinear steps includes a plurality of time-interval vectors, each element of one of the time-interval vectors representing a speech time interval the same as that represented by the other elements of the same time-interval vector but different from the speech time interval represented by the elements of a different one of the time-interval vectors, at least one nonlinear step comprising computing nonlinear combinations of elements of different ones of the time-interval vectors to include in its output a separate element representing the result of each such computed combination.

Claim 25

A speech processor as defined in claim 20 wherein the modeling means comprises means for generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 26

A speech processor as defined in claim 20 wherein the vector output of the sequence of vector-processing steps that is modeled in the modeling means includes the second-nonlinear-step output vector.

Claim 27

A speech processor as defined in claim 26 wherein the modeling means comprises means for generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 28

A speech processor as defined in claim 16 wherein the second nonlinear step comprises computing the nonlinear combinations of different elements of the first-nonlinear-step output vector by: A. computing linear combinations of the first-nonlinear-step elements; and B. computing nonlinear combinations of different ones of the linear combinations, thereby computing nonlinear combinations of different elements of the first-nonlinear-step output vector.

Claim 29

A speech processor as defined in claim 24 wherein: A. the step of computing linear combinations of the first-nonlinear-step elements comprises computing highest-variance components as the linear combinations, the highest-variance components being the components of the first-nonlinear-step input vector in the directions of the eigenvectors of the first-nonlinear-step input vector's covariance matrix that are associated with the greatest eigenvalues; and B. the step of computing nonlinear combinations of different ones of the linear combinations comprises computing nonlinear combinations of the highest-variance components but omitting computation of the nonlinear combinations of at least some of the components of the first-nonlinear-step input vector in the directions of the eigenvectors of the first-nonlinear-step input vector that are associated with the lowest eigenvalues.

Claim 30

A speech processor adapted for reception of speech and comprising: A. means for converting the speech signal to a sequence of digital quantities; means for subjecting the digital quantities to a sequence of processing steps including a sequence of vector-processing steps whose outputs are multi-element vectors and whose inputs include vector outputs of previous processing steps, the sequence of vector-processing steps including a nonlinear step whose input includes a plurality of time-interval vectors, each element of a time-interval vector representing a speech time interval the same as that represented by the other elements of the same time-interval vector but different from the speech time interval represented by the elements of a different one of the time-interval vectors, the nonlinear step comprising computing combinations of higher than second order of elements of different ones of the time-interval vectors and generating as its output a nonlinear-step output vector that includes a separate element representing the result of each such computed combination; and C. means for modeling the vector output of the sequence of vector-processing steps to generate speech-element-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective speech elements.

Claim 31

A speech processor adapted for reception of speech and comprising: A. means for converting the speech signal to a sequence of digital quantities; means for subjecting the digital quantities to a sequence of processing steps including a sequence of vector-processing steps whose outputs are multi-element vectors and whose inputs include vector outputs of previous processing steps, the sequence of vector-processing steps including a nonlinear step whose input includes a plurality of time-interval vectors, each element of a time-interval vector representing a speech time interval the same as that represented by the other elements of the same time-interval vector but different from the speech time interval represented by the elements of a different one of the time-interval vectors, the nonlinear step comprising computing combinations of higher than second order of elements of different ones of the time-interval vectors and generating as its output a nonlinear-step output vector that includes a separate element representing the result of each such computed combination; and C. means for modeling the vector output of the sequence of vector-processing steps to generate speech-element-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective speech elements.

Claim 32

A speech processor as defined in claim 26 wherein the modeling means comprises means for generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.

Claim 33

A speech processor as defined in claim 26 wherein the vector output of the sequence of vector-processing steps that is modeled in the modeling means includes the nonlinear-step output vector.

Claim 34

A speech processor as defined in claim 28 wherein the modeling means comprises means for generating as the speech-element-estimate signals phoneme-estimate signals that represent respective figures of merit related to the likelihoods that the speech contains respective phonemes.