Pattern recognition scheme using probabilistic models based on mixtures distribution of discrete distribution

• US 6,064,958 A
• Filed: 09/19/1997
• Issued: 05/16/2000
• Est. Priority Date: 09/20/1996
• Status: Expired due to Fees

• Current Assignee: Nippon Telegraph and Telephone Corporation
• Original Assignee: Nippon Telegraph and Telephone Corporation
• Inventors: Sagayama, Shigeki, Takahashi, Satoshi
• Primary Examiner(s): Voeltz, Emanuel Todd
• Assistant Examiner(s): SOFOCLEOUS, MICHAEL D
• Application Number:
  US08/934,376
• Time in Patent Office: 970 Days
• Field of Search: 704/230, 704/243, 704/256, 704/245
• US Class Current: 704/243
• 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 15/144 Training of HMMs

Abstract

A pattern recognition scheme using probabilistic models that are capable of reducing a calculation cost for the output probability while improving a recognition performance even when a number of mixture component distributions of respective states is small, by arranging distributions with low calculation cost and high expressive power as the mixture component distribution. In this pattern recognition scheme, a probability of each probabilistic model expressing features of each recognition category with respect to each input feature vector derived from each input signal is calculated, where the probabilistic model represents a feature parameter subspace in which feature vectors of each recognition category exist and the feature parameter subspace is expressed by using mixture distributions of one-dimensional discrete distributions with arbitrary distribution shapes which are arranged in respective dimensions. Then, a recognition category expressed by a probabilistic model with a highest probability among a plurality of probabilistic models is outputted as a recognition result.

Claims

1. A pattern recognition method, comprising the steps of:

calculating a probability of each probabilistic model expressing features of each recognition category with respect to each input feature vector derived from each input signal, wherein the probabilistic model represents a feature parameter subspace in which feature vectors of each recognition category exist and the feature parameter subspace is expressed by using mixture distributions of one-dimensional discrete distributions with arbitrary distribution shapes which are arranged in respective dimensions; and

outputting a recognition category expressed by a probabilistic model with a highest probability among a plurality of probabilistic models as a recognition result.

2. The method of claim 1, wherein the calculating step includes the steps of:

obtaining a scalar quantization code by scalar quantizing a value of each dimension component of the input feature vector;

determining a probability of each dimension of each mixture component distribution by referring to a scalar quantization code book according to the scalar quantization code;

calculating a product of probabilities of all dimensions of each mixture component distribution as a probability of each mixture component distribution; and

setting a sum or a maximum value of probabilities of all mixture component distributions within each state as a probability of each probabilistic model in each state with respect to the input feature vector.

3. The method of claim 1, wherein a scalar quantization code book for the discrete distributions is shared among distributions existing in all or a part of the probabilistic models.

4. The method of claim 1, wherein a scalar quantization code book for the discrete distributions is shared among distributions existing in an identical dimension of all or a part of the probabilistic models.

5. The method of claim 1, wherein the discrete distributions are distributions based on quantization points defined by a linear scalar quantization.

6. The method of claim 1, wherein the discrete distributions are distributions based on quantization points defined by a non-linear scalar quantization, and the quantization points are given in a tree structure.

7. The method of claim 1, further comprising the step of:

estimating parameters of the discrete distributions from training data, by first training a continuous distribution type model with a number of mixture component distributions equal to a desired number of distributions to be mixed in the mixture distribution, and then training a probabilistic model obtained by discretizing each continuous distribution of the continuous distribution type model as an initial model.

8. The method of claim 1, further comprising the step of:

estimating parameters of the discrete distributions from training data, by first training a continuous distribution type model with a number of mixture component distributions greater than a desired number of distributions to be mixed in the mixture distribution, then merging continuous distributions of the continuous distribution type model until a number of mixture component distributions becomes equal to the desired number of distributions, and then training a probabilistic model obtained by discretizing each continuous distribution of the continuous distribution type model after merging as an initial model.

9. The method of claim 1, wherein the probabilistic models are hidden Markov models.

10. The method of claim 1, wherein the input signal is a speech signal, the input feature vector expresses features of an input speech, and the recognition result indicates the recognition category for the input speech.

11. A pattern recognition apparatus, comprising:

a calculation unit for calculating a probability of each probabilistic model expressing features of each recognition category with respect to each input feature vector derived from each input signal, wherein the probabilistic model represents a feature parameter subspace in which feature vectors of each recognition category exist and the feature parameter subspace is expressed by using mixture distributions of one-dimensional discrete distributions with arbitrary distribution shapes which are arranged in respective dimensions; and

a recognition unit for outputting a recognition category expressed by a probabilistic model with a highest probability among a plurality of probabilistic models as a recognition result.

12. The apparatus of claim 11, wherein the calculation unit includes:

a first unit for obtaining a scalar quantization code by scalar quantizing a value of each dimension component of the input feature vector;

a second unit for determining a probability of each dimension of each mixture component distribution by referring to a scalar quantization code book according to the scalar quantization code;

a third unit for calculating a product of probabilities of all dimensions of each mixture component distribution as a probability of each mixture component distribution; and

a fourth unit for setting a sum or a maximum value of probabilities of all mixture component distributions within each state as a probability of each probabilistic model in each state with respect to the input feature vector.

13. The apparatus of claim 11, wherein the calculation unit uses a scalar quantization code book for the discrete distributions which is shared among distributions existing in all or a part of the probabilistic models.

14. The apparatus of claim 11, wherein the calculation unit uses a scalar quantization code book for the discrete distributions which is shared among distributions existing in an identical dimension of all or a part of the probabilistic models.

15. The apparatus of claim 11, wherein the calculation unit uses the discrete distributions which are distributions based on quantization points defined by a linear scalar quantization.

16. The apparatus of claim 11, wherein the calculation unit uses the discrete distributions are distributions based on quantization points defined by a non-linear scalar quantization, and the quantization points are given in a tree structure.

17. The apparatus of claim 11, further comprising:

a training unit for estimating parameters of the discrete distributions from training data, by first training a continuous distribution type model with a number of mixture component distributions equal to a desired number of distributions to be mixed in the mixture distribution, and then training a probabilistic model obtained by discretizing each continuous distribution of the continuous distribution type model as an initial model.

18. The apparatus of claim 11, further comprising:

a training unit for estimating parameters of the discrete distributions from training data, by first training a continuous distribution type model with a number of mixture component distributions greater than a desired number of distributions to be mixed in the mixture distribution, then merging continuous distributions of the continuous distribution type model until a number of mixture component distributions becomes equal to the desired number of distributions, and then training a probabilistic model obtained by discretizing each continuous distribution of the continuous distribution type model after merging as an initial model.

19. An article of manufacture, comprising:

a computer usable medium having computer readable program code means embodied therein for causing a computer to function as a pattern recognition system, the computer readable program code means including;

first computer readable program code means for causing said computer to calculate a probability of each probabilistic model expressing features of each recognition category with respect to each input feature vector derived from each input signal, wherein the probabilistic model represents a feature parameter subspace in which feature vectors of each recognition category exist and the feature parameter subspace is expressed by using mixture distributions of one-dimensional discrete distributions with arbitrary distribution shapes which are arranged in respective dimensions; and

second computer readable program code means for causing said computer to output a recognition category expressed by a probabilistic model with a highest probability among a plurality of probabilistic models as a recognition result.

20. The article of manufacture of claim 19, wherein the first computer readable program code means includes:

a first computer readable sub-program code means for obtaining a scalar quantization code by scalar quantizing a value of each dimension component of the input feature vector;

a second computer readable sub-program code means for determining a probability of each dimension of each mixture component distribution by referring to a scalar quantization code book according to the scalar quantization code;

a third computer readable sub-program code means for calculating a product of probabilities of all dimensions of each mixture component distribution as a probability of each mixture component distribution; and

a fourth computer readable sub-program code means for setting a sum or a maximum value of probabilities of all mixture component distributions within each state as a probability of each probabilistic model in each state with respect to the input feature vector.

21. The article of manufacture of claim 19, wherein the first computer readable program code means uses a scalar quantization code book for the discrete distributions which is shared among distributions existing in all or a part of the probabilistic models.

22. The article of manufacture of claim 19, wherein the first computer readable program code means uses a scalar quantization code book for the discrete distributions which is shared among distributions existing in an identical dimension of all or a part of the probabilistic models.

23. The article of manufacture of claim 19, wherein the first computer readable program code means uses the discrete distributions which are distributions based on quantization points defined by a linear scalar quantization.

24. The article of manufacture of claim 19, wherein the first computer readable program code means uses the discrete distributions are distributions based on quantization points defined by a non-linear scalar quantization, and the quantization points are given in a tree structure.

25. The article of manufacture of claim 19, further comprising:

a third computer readable program code means for causing said computer to estimate parameters of the discrete distributions from training data, by first training a continuous distribution type model with a number of mixture component distributions equal to a desired number of distributions to be mixed in the mixture distribution, and then training a probabilistic model obtained by discretizing each continuous distribution of the continuous distribution type model as an initial model.

26. The article of manufacture of claim 19, further comprising:

a third computer readable program code means for causing said computer to estimate parameters of the discrete distributions from training data, by first training a continuous distribution type model with a number of mixture component distributions greater than a desired number of distributions to be mixed in the mixture distribution, then merging continuous distributions of the continuous distribution type model until a number of mixture component distributions becomes equal to the desired number of distributions, and then training a probabilistic model obtained by discretizing each continuous distribution of the continuous distribution type model after merging as an initial model.

27. A pattern recognition apparatus, comprising:

an A/D conversion unit for converting input speech signals into digital signals;

an input speech feature vector extraction unit for extracting an input speech feature vector from the digital signals converted by the A/D conversion unit;

a model training speech data storage unit for storing model training speech data;

a model training speech feature vector extraction unit for extracting model training speech feature vectors from the model training speech data stored in the model training speech data storage unit;

an initial model generation unit for generating initial models of discrete mixture distribution type models by carrying out a training of continuous mixture distribution type models using the model training speech feature vectors extracted by the model training speech feature vector extraction unit;

a model parameter estimation unit for estimating model parameters of the discrete mixture distribution type models using the model training speech feature vectors extracted by the model training speech feature vector extraction unit, with discrete mixture distributions of the initial models generated by the initial model generation unit as initial distributions;

a model parameter memory unit for storing the model parameters obtained by the model parameter estimation unit;

a model probability calculation unit for calculating a probability of the input speech feature vector with respect to each discrete mixture distribution type model according to the model parameters stored in the model parameter memory unit; and

a recognition result output unit for outputting a recognition category expressed by a discrete mixture distribution type model with a highest probability calculated by the model probability calculation unit among the discrete mixture distribution type models as a recognition result.

28. A pattern recognition method, comprising the steps of:

(a) converting input speech signals into digital signals;

(b) extracting an input speech feature vector from the digital signals converted by the step (a);

(c) storing model training speech data;

(d) extracting model training speech feature vectors from the model training speech data stored by the step (c);

(e) generating initial models of discrete mixture distribution type models by carrying out a training of continuous mixture distribution type models using the model training speech feature vectors extracted by the step (d);

(f) estimating model parameters of the discrete mixture distribution type models using the model training speech feature vectors extracted by the step (d), with discrete mixture distributions of the initial models generated by the step (e) as initial distributions;

(g) storing the model parameters obtained by the step (f);

(h) calculating a probability of the input speech feature vector with respect to each discrete mixture distribution type model according to the model parameters stored by the step (g); and

(i) outputting a recognition category expressed by a discrete mixture distribution type model with a highest probability calculated by the step (h) among the discrete mixture distribution type models as a recognition result.