Method and apparatus for pattern recognition employing the Hidden Markov Model

• US 5,638,489 A
• Filed: 06/07/1995
• Issued: 06/10/1997
• Est. Priority Date: 06/03/1992
• Status: Expired due to Fees

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Tsuboka, Eiichi
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US08/486,695
• Time in Patent Office: 734 Days
• Field of Search: 381/43, 395/2, 395/2.65, 395/2.54, 395/2.45, 395/2.49, 395/2.48, 395/3, 395/2.03, 395/51
• US Class Current: 704/256
• CPC Class Codes: G06F 18/2321 using statistics or function optimisation, e.g. modelling of probability density functions
• CPC Class Codes: G06F 18/295 Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

A method and apparatus for estimating parameters of a new hidden Markov model (HMM) applicable to pattern recognition such as recognition of speech signals which are time series signals, and a method and apparatus for pattern recognition employing this HMM are provided. It is easily applicable to speech and other time series signals. In particular, the pattern recognition degree of a time series observation vector signaly, received from an information source, is calculated by using function values u(y,1), u(y,2), . . . , u(y,M) and occurrence probabilities of signals C₁, C₂, . . . , C_M which are composed of set C, where u(y,m) is the image into which the pair (C_m,Y) is mapped, u(y,m).di-elect cons.U(U=[a,b], a,b.di-elect cons.R¹ and 0≦a≦b) and y.di-elect cons.Rⁿ (Rⁿ : n-dimensional Euclidean space). C is a set of signals against which the observation vector signal y is compared.

Claims

1. A computer implemented hidden Markov model creating apparatus for processing an observation vector signal y(t) representing a sound input signal, comprising:

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y(t), (C_m,y(t)) into a signal u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . ,M, a,b.di-elect cons.R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M }, y(t).di-elect cons.Rⁿ, Rⁿ is an n-dimensional Euclidean space,

signal occurrence probability memory means for storing the occurrence probability of each signal of set C, where said occurrence probabilities of signals in the set C are received and stored, and

weighted sum calculating means for generating a second calculation signal representing the weighted sum of logarithmic values of occurrence probabilities of signals in the set C or generating a third calculation signal representing the weighted arithmetic mean of said logarithmic values of occurrence probabilities of signals in the set C, where said occurrence probabilities of signals in the set C are stored in said signal occurrence probability memory means and said weighting coefficient for the m-th signal is defined by u(y(t),m),

wherein the signal representing the weighted sum or the weighted arithmetic mean is the occurrence degree of the vector signal y(t) at time t, and parameter estimating means is provided for estimating the parameter of the model so that the occurrence degree of the pattern to be modeled, composed of observation vector series y(1) . . . ,y(t), . . . y(T), may be maximum on the basis of said occurrence degree of the vector signal y(t).

2. A computer implemented hidden Markov model creating apparatus for processing an observation vector signal y(t) representing a sound input signal, comprising:

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y(t), (C_m,y(t)) into a signal u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . ,M, a,b.di-elect cons.R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M }, y(t).di-elect cons.Rⁿ, Rⁿ is an n-dimensional Euclidean space,

signal occurrence probability memory means for storing the occurrence probability of each signal of set C, where said occurrence probabilities of signals in the set C are received and stored, and

power product calculating means for generating a second calculation signal representing the product of powers of occurrence probabilities of signals in the set C or generating a third calculation signal representing the weighted geometric mean of occurrence probabilities of signals in the set C, where said occurrence probabilities of signals in the set C are stored in said signal occurrence probability memory means and said power or weighting coefficient for the m-th signal is defined by u(y(t),m),

wherein the signal representing the product of power or the weighted geometric mean is the occurrence degree of the vector signal y(t) at time t, and parameter estimating means is provided for estimating the parameter of the model so that the occurrence degree of the pattern to be modeled, composed of observation vector series y(1) . . . ,y(t), . . . y(T), may be maximum on the basis of said occurrence degree of the vector signal y(t).

3. A computer implemented hidden Markov model memory apparatus comprising parameter memory means for storing the parameters of a model and characterized by storing the parameters estimated by a computer implemented hidden Markov model creating apparatus for processing an observation vector signal y(t) representing a sound input signal by the parameter memory means, employing the hidden Markov model creating apparatus comprising:

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y(t), (C_m,y(t)) into a signal u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . ,M, a,b.di-elect cons.R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M }, y(t).di-elect cons.Rⁿ, Rⁿ is an n-dimensional Euclidean space,

signal occurrence probability memory means for storing the occurrence probability of each signal of set C, where said occurrence probabilities of signals in the set C are received and stored, and

weighted sum calculating means for generating a second calculation signal representing the weighted sum of logarithmic values of occurrence probabilities of signals in the set C or generating a third calculation signal representing the weighted arithmetic mean of said logarithmic values of occurrence probabilities of signals in the set C, where said occurrence probabilities of signals in the set C are stored in said signal occurrence probability memory means and said weighting coefficient for the m-th signal is defined by u(y(t),m),

wherein the signal representing the weighted sum or the weighted arithmetic mean is the occurrence degree of the vector signal y(t) at time t, and parameter estimating means is provided for estimating the parameter of the model so that the occurrence degree of the pattern to be modeled, composed of observation vector series y(1) . . . ,y(t), . . . y(T), may be maximum on the basis of said occurrence degree of the vector signal y(t).

4. A computer implemented hidden Markov model memory apparatus comprising parameter memory means for storing the parameters of a model and characterized by storing the parameters estimated by a computer implemented hidden Markov model creating apparatus for processing an observation vector signal y(t) representing a sound input signal by the parameter memory means, employing the hidden Markov model creating apparatus comprising:

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y(t), (C_m,y(t)) into a signal u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . ,M, a,b.di-elect cons.R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M }, y(t).di-elect cons.Rⁿ, Rⁿ is an n-dimensional Euclidean space,

signal occurrence probability memory means for storing the occurrence probability of each signal of set C, where said occurrence probabilities of signals in the set C are received and stored, and

power product calculating means for generating a second calculation signal representing the product of powers of occurrence probabilities of signals in the set C or generating a third calculation signal representing the weighted geometric mean of occurrence probabilities of signals in the set C, where said occurrence probabilities of signals in the set C are stored in said signal occurrence probability memory means and said power or weighting coefficient for the m-th signal is defined by u(y(t),m),

wherein the signal representing the product of power or the weighted geometric mean is the occurrence degree of the vector signal y(t) at time t, and parameter estimating means is provided for estimating the parameter of the model so that the occurrence degree of the pattern to be modeled, composed of observation vector series y(1) . . . ,y(t), . . . y(T), may be maximum on the basis of said occurrence degree of the vector signal y(t).

5. A computer implemented likelihood calculating apparatus for processing an observation vector signal y(t) representing a sound input signal, comprising:

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y(t), (C_m,y(t)) into a signal u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . ,M, a,b.di-elect cons.R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M }, y(t).di-elect cons.Rⁿ, Rⁿ is an n-dimensional Euclidean space,

signal occurrence probability memory means for storing the occurrence probability of each signal of set C, where said occurrence probabilities of signals in the set C are received and stored, and

weighted sum calculating means for generating a second calculation signal representing the weighted sum of logarithmic values of occurrence probabilities of signals in the set C or generating a third calculation signal representing the weighted arithmetic mean of said logarithmic values of occurrence probabilities of signals in the set C, where said occurrence probabilities of signals in the set C are stored in said signal occurrence probability memory means and said weighting coefficient for the m-th signal is defined by u(y(t),m),

wherein the signal representing the weighted sum or the weighted arithmetic mean is the occurrence degree of the vector signal y(t) at time t, and parameter estimating means is provided for estimating the parameter of the model so that the occurrence degree of the pattern to be modeled, composed of observation vector series y(1) . . . ,y(t), . . . y(T), may be maximum on the basis of said occurrence degree of the vector signal y(t), and

a likelihood of a hidden Markov model possessing the parameters stored in the parameter memory means to the observation vector is calculated.

6. A computer implemented likelihood calculating apparatus for processing an observation vector signal y(t) representing a sound input signal, comprising:

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y(t), (C_m,y(t)) into a signal u(y(t),m).di-elect cons.U=[a,b], where m=1, . . . ,M, a,b.di-elect cons.R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M }, y(t).di-elect cons.Rⁿ, Rⁿ is an n-dimensional Euclidean space,

signal occurrence probability memory means for storing the occurrence probability of each signal of set C, where said occurrence probabilities of signals in the set C are received and stored, and

power product calculating means for generating a second calculation signal representing the product of powers of occurrence probabilities of signals in the set C or generating a third calculation signal representing the weighted geometric mean of occurrence probabilities of signals in the set C, where said occurrence probabilities of signals in the set C are stored in said signal occurrence probability memory means and said power or weighting coefficient for the m-th signal is defined by u(y(t),m),

wherein the signal representing the product of powers or the weighted geometric mean is the occurrence degree of the vector signal y(t) at time t, and parameter estimating means is provided for estimating the parameter of the model so that the occurrence degree of the pattern to be modeled, composed of observation vector series y(1) . . . ,y(t), . . . y(T), may be maximum on the basis of said occurrence degree of the vector signal y(t), and

a likelihood of a hidden Markov model possessing the parameters stored in the parameter memory means to the observation vector is calculated.

7. A computer implemented hidden Markov model creating apparatus for processing an observation vector signal y(t) representing a sound input signal, comprising:

a feature extraction part for convening a plurality of training speech signals of r=1 to R^w for model creation corresponding to recognition unit w(=1, . . . , W) into a feature vector series

space="preserve" listing-type="equation">Y.sup.w(r) =(y.sup.w(r) (1), y.sup.w(r) (2), . . . , y.sup.w(r) (T.sup.w(r))),

a recognition unit pattern memory part for storing a plurality of training patterns for model creation corresponding to the recognition unit w(=1, . . . ,W) by R^w pieces for w in a form of the feature vector series, a buffer memory temporarily storing patterns Y^w (1), . . . ,Y^w (R^w) for the recognition unit pattern memory part,

a clustering part for clustering the feature vector set from the recognition trait pattern memory part,

a centroid memory part for storing centroids of clusters F₁, . . . ,F_M obtained by clustering by the clustering part,

a vector membership value calculating memory part for calculating and storing the membership values u(y^w(r) (t),1), . . . , u(y^w(r) (t),M), each of which is derived for the vector y^w(r) (t) composing the training pattern stored in said buffer memory to said cluster F_m using centroids stored in said centroid memory part,

a parameter estimating part for estimating parameters of model λ

^w corresponding to recognition unit w, by executing a step of creating the model λ

^w by giving the occurrence degree ω

^w₁ (y^w(r) (t)) of feature vector y^w(r) (t) in state i of model λ

^w corresponding to the recognition unit in a form of ##EQU62## in the relation between the membership value u(y^w(r) (t),m) and occurence probability b^w_im in state i of cluster F_m,

a first parameter memory part for temporarily storing the reestimated value of the parameter estimating part, and a second parameter memory part for storing the parameters as the final result of the re-estimation corresponding to the recognition unit w(=1, . . . , W),

wherein the parameter estimating part re-estimates parameters of the model λ

^w by using the value of the first parameter memory part where the parameters are renewed iteratively by the re-estimation and the renewed parameters are used for next re-estimation step.