Method and apparatus for pattern recognition employing the hidden Markov model

• US 5,608,841 A
• Filed: 06/03/1993
• Issued: 03/04/1997
• 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:
  US08/071,656
• Time in Patent Office: 1,370 Days
• Field of Search: 395/2.65, 395/2, 395/2.45, 395/2.54, 395/2.49, 395/2.48, 381/43
• 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. It is easily applicable to speech and other time series signals. In particular, the pattern recognition degree of a time series observation vector signal y, 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)εU (U=[a,b], a,bεR¹ and 0≦a≦b) and yεRⁿ (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 signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y, (C_m,y) into a signal u(y,m)ε

U=[a,b], where m=1, . . . , M, a,bε

R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M } is a set of signals against which the observation vector signal y is compared, yε

Rⁿ is an element of n-dimensional Euclidean space,

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

weighted sum calculating means for

a) generating a second calculation signal representing the weighted sum of logarithmic values of occurrence probabilities of signals in the set C or

b) generating a third calculation signal representing the weighted arithmetic mean of said logarithmic values of occurrence probabilities of signals in the set C under the definition that the weighting coefficient for the m-th signal C_m is u (y,m) obtained by said function value calculating means,

wherein the calculation signal representing the weighted sum or weighted arithmetic mean is a degree of pattern recognition of the observation vector signal y.

2. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 1 wherein said sound is human speech.

3. A computer implemented signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y, (C_m,y) into a signal u(y,m)ε

U=[a,b], where m=1, . . . , M, a,b ε

R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M } is a set of signals against which the observation vector signal y is compared, yε

Rⁿ is an element of n-dimensional Euclidean space,

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

power product calculating means for

a) generating a second calculation signal representing the product of powers of occurrence probabilities of signals in the set C or

b) generating a third calculation signal representing the weighted geometric mean of said occurrence probabilities of signals in the set C under the definition that the power or weighting coefficient for the m-th signal C_m is u(y,m) obtained by said function value calculating means,

wherein the product of power or the weighted geometric mean is a degree of pattern recognition of the observation vector signal y.

4. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 3 wherein said sound source is human speech.

5. A computer implemented signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

a signal cluster F_m (m=1, . . . , M), where each signal cluster is obtained by clustering a training vector signal set,

membership value calculating means for generating a first calculation signal representing the membership value of the observation vector signal y to each signal cluster F_m (m=1, . . . , M),

signal cluster occurrence probability memory means for storing the occurrence probability of each signal cluster, where said occurrence probabilities of signal clusters are received and stored, and

weighted sum calculating means for

a) generating a second calculation signal representing the weighted sum of logarithmic values of said occurrence probabilities of signal clusters or

b) generating a third calculation signal representing the weighted arithmetic mean of said logarithmic values of occurrence probabilities of signal clusters under the definition that the weighting coefficient for the m-th signal cluster F_m is said membership value of y for F_m obtained by said membership value calculating means,

wherein the weighted sum or weighted arithmetic mean is a degree of pattern recognition of the observation vector signal y.

6. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 5, wherein the calculation signal representing the weighted sum is generated on the closest K signal clusters of the observation vector signal y, to a predetermined K.

7. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 5, wherein the calculation signal representing the weighted sum is generated on the signal clusters of which membership value of the observation vector signal y is more than a predetermined threshold.

8. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 5 wherein said sound is human speech.

9. A computer implemented signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

a signal cluster F_m (m=1, . . . , M), where each signal cluster is obtained by clustering the training vector signal set,

membership value calculating means for generating a first calculation signal representing the membership value of the observation vector signal y to each signal cluster F_m (m=1, . . . , M),

signal cluster occurrence probability memory means for storing the occurrence probability of each signal cluster, where a priori occurrence probabilities of signal clusters are received and stored, and

power product calculating means for

a) generating a second calculation signal representing the product of powers of said a priori occurrence probabilities of signal clusters or

b) generating a third calculation signal representing the weighted geometric mean of said a priori occurrence probabilities of signal clusters under the definition that the power or weighting coefficient for the m-th signal cluster F_m is said membership value of y for F_m obtained by said membership value calculating means,

wherein the product of powers or the weighted geometric mean is a degree of pattern recognition of the observation vector signal y.

10. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 9, wherein the calculation signal representing the product of powers is generated on the closest K signal clusters of the observation vector signal y, to a predetermined K.

11. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 9, wherein the calculation signal representing the product of powers is generated on the signal clusters of which membership value of the observation vector signal y is more than a predetermined threshold.

12. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 9 wherein said sound is human speech.

13. A computer implemented signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

source signal occurrence probability memory means for storing a priori occurrence probability of each source signal, where said a priori occurrence probabilities of source signals are received and stored,

a posteriori probability calculating means for generating a first calculation signal representing the a posteriori probability of each of said source signals D₁, . . . , D_m for an observation vector signal y, where each source signal generates said observation vector y according to a probability density defined for the source signal, and

weighted sum calculating means for

a) generating a second calculation signal representing the weighted sum of said a priori occurrence probabilities of source signals or

b) generating a third calculation signal representing the weighted arithmetic mean of logarithmic value of said a priori occurrence probabilities of source signals under the definition that the weighting coefficient for the m-th source signal is defined by the a posteriori probability of the source signal C_m calculated by said a posteriori probability calculating means,

wherein the calculation signal representing the weighted sum or the weighted arithmetic mean is a degree of pattern recognition of the observation vector signal y.

14. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 13, wherein the source signal generates vector signals contained in each signal cluster obtained by clustering the training vector signal set, and the a posteriori probability to the observation vector signal of the source signal is the membership value of the observation vector signal y to the signal cluster.

15. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 13 wherein said sound is human speech.

16. A computer implemented signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

source signal occurrence probability memory means for storing a priori occurrence probability of each source signal, where said a priori occurrence probabilities of source signals are received and stored,

a posteriori probability calculating means for generating a first calculation signal representing the a posteriori probability of each of said source signals D₁, . . . , D_m for the observation vector signal y, where each source signal generates said observation vector signal y according to a probability density defined for the source signal, and

power product calculating means for

a) generating a second calculation signal representing the product of powers of said a priori occurrence probabilities of source signals or

b) generating a third calculation signal representing the weighted geometric mean of said a priori occurrence probabilities of source signals under the definition that the power or weighting coefficient for the m-th source signal is defined by the a posteriori probability of the source signal C_m calculated by said a posteriori probability calculating means,

wherein the calculation signal representing the product of power or the weighted geometric mean is a degree of pattern recognition of the observation vector signal y.

17. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 16, wherein the source signal generates vector signals contained in each signal cluster obtained by clustering a training vector signal set, and the a posteriori probability to the observation vector signal y of the source signal is the membership value of the observation vector signal y to the signal cluster.

18. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 16 wherein said sound is human speech.

19. A computer implemented signal processing apparatus used for a sound recognition apparatus comprising:

means for receiving a signal representing a sound and converting the signal representing the sound to a time series observation vector signal y,

signal occurrence degree memory means for storing the degree of occurrence of C_m in the conditions of predetermined S_n, for a first set S={S₁, S₂, . . . ,S_N } composed of N signals, and a second set C={C₁, C₂, . . . , C_M } composed of M signals,

first function calculating means for generating a first calculation signal representing a first function value expressing the strength of relation between a pattern set Z generated from information source A generating time series signals and the signal S_n,

second function calculating means for generating a second calculation signal representing a second function value expressing the strength of relation between the observation pattern signal y and signal C_m, and

observation pattern occurrence degree calculating means for generating a third calculation signal representing the possibility of occurrence of the observation pattern signal y from the information source A, from the occurrence degree of each signal of C stored in the signal occurrence degree memory means and the first function value and second function value.

20. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 19, wherein the observation pattern occurrence degree calculating means

a) generates a first calculation signal representing the weighted sum of logarithmic values of occurrence probabilities of signal in the set C or

b) generates a second calculation signal representing the weighted arithmetic mean of said occurrence probabilities in the set C,

where said occurrence probabilities of signal in the set C are received and stored in said signal occurrence degree memory means,

where weighting coefficients used to obtain said weighted sum or said weighted arithmetic mean are the product of first function values and second function values,

wherein said calculation signal representing the weighted sum or said weighted arithmetic mean is a degree of pattern recognition of the observation pattern signal y.

21. A computer implemented signal processing apparatus used for a sound recognition apparatus of claim 19, wherein the observation pattern occurrence degree calculating means

a) generates a first calculation signal representing the product of powers of occurrence probabilities of signals in the set C or

b) generates a second 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 memorized in said signal occurrence degree memory means,

where said power or weighting coefficients used to obtain said product of powers or said weighted geometric mean are products of the first function values and the second function values,

wherein said calculation signal representing the product of powers or said weighted geometric mean is the occurrence degree of the observation pattern signal y.

22. A computer implemented signal processing apparatus used for a sound recognition apparatus according to claim 19 wherein said sound is human speech.

23. A computer implemented signal processing apparatus comprising:

feature extraction means for receiving a sound and converting the sound to a time series observation vector signal y,

function value calculating means for generating a first calculation signal corresponding to a mapping of each pair C_m and y, (C_m,y) into a signal u(y,m)ε

U=[a,b], where m=1, . . . , M, a,b ε

R¹ for 0≦

a≦

b, C={C₁, C₂, . . . , C_M } is a set of signals against which the observation vector signal y is compared, yε

Rⁿ is an element of n-dimensional Euclidean space,

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

weighted sum calculating means for

a) generating a second calculation signal representing the weighted sum of logarithmic values of occurrence probabilities of signals in the set C or

b) generating a third calculation signal representing the weighted arithmetic mean of said logarithmic value of occurrence probabilities of signals in the set C under the definition that the weighting coefficient for the m-th signal C_m is u(y,m) obtained by said function value calculating means, and

decision means for generating a recognized word corresponding to the sound using the calculation signal representing the weighted sum or weighted arithmetic mean.