Method and apparatus for pattern recognition employing the hidden markov model

Claim 2

A recognition apparatus comprising: function value calculating means for mapping each pair C m and y(t), C m ,y(t) into an element u(y(t),m) .di-elect cons.U=[a,b], where m=1, . . . , M, a,b.di-elect cons.R 1 for O≦a≦b, C={C 1 , C 2 , . . . , C M }, y(t).di-elect cons.R n , y(t) is a vector observed at time t and R n is n-dimensional Euclidean space, element occurrence probability memory means for storing the occurrence probability of each element of set C, weighted sum calculating means for calculating the weighted sum of logarithmic values of occurrence probabilities of elements in the set C or calculating the weighted arithmetic mean of them, where said occurrence probabilities of elements in the set C are memorized in said element occurrence probability memory means and said weighing coefficients for the m-th element is defined by u(y(t),m), parameter memory means, having the weighted sum weighted arithmetic mean as the occurrence degree of the vector y (t), for storing a parameter of each of a plurality of models estimated so that the occurrence degree of the pattern set to be modeled composed of observation vector series y(t) . . . , y(t), . . . y(T) may be maximum on the basis of said vector occurrence degree, likelihood calculating means for calculating the likelihood of each model for the observation vector series by disposing the parameter memory means in each of a plurality of recognition units, and comparing means for comparing the likelihood of each model obtained by the likelihood calculating means, and identifying, from among the plurality of recognition units, the recognition unit corresponding to the model giving the maximum value as the result of recognition.

Claim 3

A recognition apparatus comprising: function value calculating means for mapping each pair C m and y(t), (C m ,y(t)) into an element u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . , M, a,b .di-elect cons.R 1 for O≦a≦b, C={C 1 , C 2 , . . . , C M }, (y)t.di-elect cons.R n , y(t) is a vector observed at time t and R n is n-dimensional Euclidean space, element occurrence probability memory means for storing the occurrence probability of each element of set C, power product calculating means for calculating the product of powers of occurrence probabilities of elements in the set C or calculating the weighted geometric mean of occurrence probabilities of elements in the set C, where said occurrence probabilities of elements in the set C are memorized in said element occurrence probability memory means and said power or weighing coefficient for the m-th element is defined by u(y(t),m), parameter memory means, having the product of power or the weighted geometric mean as the occurrence degree of the vector y(t), for storing a parameter of each of a plurality of models estimated so that the occurrence degree of the pattern set to be modeled composed of observation vector series y(t) . . ., y(t), . . . y(T) may be maximum on the basis of the vector occurrence degree, likelihood calculating means for calculating the likelihood of each model for the observation vector series by disposing the parameter memory means in each of a plurality of recognition units, and comparing means for comparing the likelihood of each model obtained by the likelihood calculating means, and identifying, from among the plurality of recognition units, the recognition unit corresponding to the model giving the maximum value as the result of recognition.

Claim 4

A recognition apparatus comprising: function value calculating means for mapping each pair C m and y(t), (C m ,y(t)) into an element u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . , M, a,b .di-elect cons.R 1 for O≦a≦b, C={C 1 , C 2 , . . . , C M }, (y)t.di-elect cons.R n , y(t) is a vector observed at time t and R n is n-dimensional Euclidean space, element occurrence probability memory means for storing the occurrence probability of each element of set C, power product calculating means for calculating the product of powers of occurrence probabilities of elements in the set C or calculating the weighted geometric mean of occurrence probabilities of elements in the set C, where said occurrence probabilities of elements in the set C are memorized in said element occurrence probability memory means and said power or weighing coefficient for the m-th element is defined by u(y(t),m), parameter memory means, having the product of power or the weighted geometric mean as the occurrence degree of the vector y(t), for storing a parameter of each of a plurality of models estimated so that the occurrence degree of the pattern set to be modeled composed of observation vector series y(t) . . ., y(t), . . . y(T) may be maximum on the basis of the vector occurrence degree, likelihood calculating means for calculating the likelihood of each model for the observation vector series by disposing the parameter memory means in each of a plurality of recognition units, and comparing means for comparing the likelihood of each model obtained by the likelihood calculating means, and identifying, from among the plurality of recognition units, the recognition unit corresponding to the model giving the maximum value as the result of recognition.

Claim 5

A speech recognition apparatus comprising: a feature extraction part for converting speech signals to be recognized into feature vector series Y=y(t), . . . ,y(T), a centroid memory part for storing centroids of clusters F 1 , . . ., F M obtained by clustering the set of feature vector y(t), a parameter memory part for storing the parameters of hidden Markov model each of which is corresponding to each recognition unit w(=1, . . ., W), a membership value calculating part for calculating the membership value of said feature vector y(t) to said clusters F 1 , . . . , F M defined by each corresponding centroid, vector occurrence degree calculating means for calculating one of the group consisting of ##EQU66## for the occurrence probability b w im in state i of hidden Markov model λ w read out from said parameter memory part, as the occurrence degree ω w 1 (y w (r) (t)) of feature vector y(t) in state i of the hidden Markov model λ w corresponding to the recognition unit w, u(y,1), u(y,2), . . ., u(y,M) are function values corresponding to weighting coefficients, wherein u(y,m) .di-elect cons.U, a likelihood calculating part for calculating the likelihood L 1 , . . . , L w corresponding to feature vector series Y=(y(1), y(2), . . . y (T)) of the hidden Markov model λ w by using the occurrence degree by said vector occurrence degree calculating means, and a decision part for comparing the likelihood L 1 , . . . , L w and deciding which is the maximum of L w (w=1, . . . , W) calculated by said likelihood calculating part, and producing the w giving the maximum as the recognition result.

Claim 6

A speech recognition apparatus comprising: a feature extraction part for converting speech signals to be recognized into feature vector series Y=y(t), . . . ,y(T), a centroid memory part for storing centroids of clusters F 1 , . . ., F M obtained by clustering the set of feature vector y(t), a parameter memory part for storing the parameters of hidden Markov model each of which is corresponding to each recognition unit w(=1, . . ., W), a membership value calculating part for calculating the membership value of said feature vector y(t) to said clusters F 1 , . . . , F M defined by each corresponding centroid, vector occurrence degree calculating means for calculating one of the group consisting of ##EQU66## for the occurrence probability b w im in state i of hidden Markov model λ w read out from said parameter memory part, as the occurrence degree ω w 1 (y w (r) (t)) of feature vector y(t) in state i of the hidden Markov model λ w corresponding to the recognition unit w, u(y,1), u(y,2), . . ., u(y,M) are function values corresponding to weighting coefficients, wherein u(y,m) .di-elect cons.U, a likelihood calculating part for calculating the likelihood L 1 , . . . , L w corresponding to feature vector series Y=(y(1), y(2), . . . y (T)) of the hidden Markov model λ w by using the occurrence degree by said vector occurrence degree calculating means, and a decision part for comparing the likelihood L 1 , . . . , L w and deciding which is the maximum of L w (w=1, . . . , W) calculated by said likelihood calculating part, and producing the w giving the maximum as the recognition result.