Adaptive training method for pattern recognition

Claim 3

The method of claim 1, wherein in said step (c) said first group of adapted models obtained by said maximum a posteriori estimation method in said step (a) are modified through a smoothing procedure by said vector field smoothing method to obtain a third group of adapted models, and said third group of adapted models is combined, as said first group of adapted models corresponding to said first group of reference models, with said second group of adapted models into said set of adapted models for pattern recognition.

Claim 4

The method of claim 1 or 2, wherein said reference models are continuous mixture hidden Markov models.

Claim 5

The method of claim 3, wherein said step (a) includes a step wherein: a statistical parameter of each of a plurality of distribution functions forming an output density distribution function of a model is to be estimated; said statistical parameter to be estimated is obtained by said maximum a posteriori estimation method using, as a control parameter common to all of said plurality of distribution functions of said output density distribution function, a statistical parameter of each prior distribution representing the reliability of said statistical parameter to be estimated; and a model to be trained in said training data is generated using distribution functions each including said statistical parameter obtained by said maximum a posteriori estimation method.

Claim 6

The method of claim 3, wherein said step (a) includes a step wherein: a statistical parameter of each of a plurality of distribution functions forming an output density distribution function of a model is to be estimated; said statistical parameter to be estimated is obtained by said maximum a posteriori estimation method using, as a control parameter corresponding to a statistical deviation of said training data for a prior distribution of each distribution function, a statistical parameter of said each prior distribution representing the reliability of said statistical parameter to be estimated; and a model to be trained in said training data is produced using distribution functions each including said statistical parameter obtained by said maximum a posteriori estimation method.

Claim 7

The method of claim 4 or 5, wherein said distribution functions are each a Gaussian distribution and said statistical parameter to be estimated is a mean vector of said model.

Claim 8

The method of claim 4 or 5, wherein said distribution functions are each a Laplacian distribution and said statistical parameter to be estimated is a mean vector of said model.

Claim 9

The method of claim 1 or 2, wherein said vector field smoothing method in said step (b) includes a process for obtaining transfer vectors on the basis of variations of statistical parameters of distribution functions of said first group of adapted models and distribution functions of said first reference models and for determining weighting coefficients of said transfer vectors in accordance with the distributions of said statistical parameters of said distribution functions of said first group of reference models.

Claim 10

The method of claim 8, wherein said statistical parameters of said distribution functions are mean vectors of models and said variations of said statistical parameters are each obtained as the difference between said mean vectors of two corresponding models.

Claim 11

The method of claim 9, which further comprises a step of calculating said weighting coefficient as a fuzzy membership function.

Claim 12

The method of claim 9, which further comprises a step of calculating said weighting coefficient as a Gaussian window function.

Claim 13

The method of claim 1, which repeats the following steps (d) through (f) upon each input of new training data after said step (c): (d) wherein those adapted models to be trained in said adapted model set which correspond to training models in said new training data are re-adapted by said maximum a posteriori estimation method to said new training data and said re-adapted models are added to said first group of adapted models; (e) wherein, letting reference models in said set corresponding to said first group of adapted models be identified as a first group of reference models and the other reference models in said set be identified as a second group of reference models, a second group of adapted models corresponding to said second group of reference models is obtained through interpolation by said vector field smoothing method using the relationship between said first group of adapted models and said first group of reference models corresponding thereto; and (f) said first group of adapted models obtained by said maximum a posteriori estimation method in said step (d) and said second group of adapted models obtained by said interpolation in said step (e) are combined into a set of updated adapted models for pattern recognition.

Claim 14

The method of claim 1, which repeats the following steps (d) through (f) upon each input of new training data after said step (c): (d) wherein said set of adapted models is regarded as a set of updated reference models, those updated reference models to be trained in said updated reference model set which correspond to said new training data are re-adapted by said maximum a posteriori estimation method using said new training data to obtain re-adapted models, and said first group of adapted models is updated with said re-adapted models; (e) wherein, letting said updated reference models in said set corresponding to said first group of adapted models be identified as a first group of updated reference models and the other updated reference models in said set of updated reference models be identified as a second group of updated reference models, a second group of adapted models corresponding to said second group of updated reference models is obtained through interpolation by said vector field smoothing method using the relationship between said first group of adapted models and said first group of updated reference models corresponding thereto; and (f) wherein said first group of adapted models obtained by said maximum a posteriori estimation method in said step (d) and said second group of adapted models obtained by said interpolation in said step (e) are combined into a set of updated adapted models for pattern recognition.

Claim 15

The method of claim 1, wherein said step (c) includes a step wherein those of reference models of an initial set which correspond to said first adapted model group are updated therewith to form a set of partially updated reference models, and which repeats the following steps (d) through (f) upon each input of new training data after said step (c): (d) wherein said set of adapted models is regarded as a set of updated reference models, those updated reference models to be trained in said updated reference model set which correspond to said new training data are adapted by said maximum a posteriori estimation method using said new training data to obtain adapted models, and said first group of adapted models are updated with said adapted models; (e) wherein, letting those partially updated reference models in said partially updated reference model set which correspond to said first group of adapted models be identified as a first group of partially updated reference models and the other partially updated reference models in said partially updated reference model set be identified as a second group of partially updated reference models, a second group of adapted models corresponding to said second group of partially updated reference models is obtained through interpolation by said vector field smoothing method using the relationship between said first group of adapted models and said first group of partially updated reference models corresponding thereto; and (f) wherein those models in said partially updated reference model set which correspond to said first group of adapted models updated in said step (d) are updated therewith, and said first group of adapted models obtained by said maximum a posteriori estimation method and said second group of adapted models by said interpolation in said step (e) are combined into a set of updated adapted models for pattern recognition.

Claim 18

The method of claim 16, wherein said maximum a posteriori estimation method includes a process wherein: a statistical parameter of each of a plurality of distribution functions forming an output density distribution function of a model is to be estimated; said statistical parameter to be estimated is obtained by said maximum a posteriori estimation method using, as a control parameter common to all of said plurality of distribution functions of said output density distribution function, a statistical parameter of each prior distribution representing the reliability of said statistical parameter to be estimated; and a model to be trained in said training data is produced using distribution functions each including said statistical parameter obtained by said maximum a posteriori estimation method.

Claim 19

The method of claim 16, wherein said maximum a posteriori estimation method include a process wherein: a statistical parameter of each of a plurality of distribution functions forming an output density distribution function of a model is to be estimated; said statistical parameter to be estimated is obtained by said maximum a posteriori estimation method using, as a control parameter corresponding to a statistical deviation of said training data for a prior distribution of each distribution function, a statistical parameter of said prior distribution representing the reliability of said statistical parameter to be estimated; and a model to be trained in said training data is produced using distribution functions each including said statistical parameter obtained by said maximum a posteriori estimation method.

Claim 20

The method of claim 17 or 18, wherein said distribution functions are each a Gaussian distribution and said statistical parameter to be estimated is a mean vector of said model.

Claim 21

The method of claim 17 or 18, wherein said distribution functions are each a Laplacian distribution and said statistical parameter to be estimated is a mean vector of said model.

Claim 22

The method of claim 17 or 18, wherein said vector field smoothing method in said step (b) includes a process for obtaining transfer vectors on the basis of variations of statistical parameters of distribution functions of said first group of adapted models and distribution functions of said first reference models and for determining weighting coefficients of said transfer vectors in accordance with the distributions of said statistical parameters of said distribution functions of said first group of reference models.

Claim 23

The method of claim 21, wherein said statistical parameters of said distribution functions are mean vectors of models and said variations of said statistical parameters are each obtained as the difference between said mean vectors of two corresponding models.

Claim 24

The method of claim 22, which further comprises a step of calculating said weighting coefficient as a fuzzy membership function.

Claim 25

The method of claim 22, which further comprises a step of calculating said weighting coefficient as a Gaussian window function.

Claim 26

The method of claim 1, wherein said step (c) includes a step wherein those of reference models of an initial set which correspond to said first adapted model group are updated therewith to form a set of partially updated reference models, and which repeats the following steps (d) through (f) upon each input of new training data after said step (c): (d) wherein those reference models to be trained in said partially updated reference model set which correspond to said new training data are adapted by said maximum a posteriori estimation method using said new training data to obtain adapted models, said adapted models are added to said first group of adapted models, and said reference models to be trained in said partially updated reference model set are updated with said adapted models; (e) wherein, letting reference models of said initial set corresponding to said first group of adapted models be identified as a first group of reference models and the other reference models in said set be identified as a second group of reference models, a second group of adapted models corresponding to said second group of reference models are obtained through interpolation by said vector field smoothing method using the relationship between said first group of adapted models and said first group of reference models corresponding thereto; and (f) wherein said first group of adapted models obtained by said maximum a posteriori estimation method in step (d) and said second group of adapted models obtained by said interpolation in said step (e) are combined into a set of updated adapted models for pattern recognition.

Claim 27

The method of claim 1, wherein said step (c) includes a step wherein those of reference models of an initial set which correspond to said first adapted model group are updated therewith to form a set of partially updated reference models, and which repeats the following steps (d) through (f) upon each input of new training data after said step (c): (d) wherein said set of adapted models is regarded as a set of updated reference models, those updated reference models to be trained in said updated reference model set which correspond to said new training data are adapted by said maximum a posteriori estimation method using said new training data to obtain adapted models, and said first adapted model group and said reference models to be trained in said partially updated reference model set are updated with said adapted models; (e) wherein, letting those updated reference models in said updated reference model set which correspond to said first group of adapted models be identified as a first group of updated reference models and the other updated reference models in said updated reference model set be identified as a second group of updated reference models, a second group of adapted models corresponding to said second group of updated reference models is obtained through interpolation by said vector field smoothing method using the relationship between said first group of adapted models and said first group of updated reference models corresponding thereto; and (f) wherein said first group of adapted models obtained by said maximum a posteriori estimation method in said step (d) and said second group of adapted models obtained by said interpolation in said step (e) are combined into a set of updated adapted models for pattern recognition.

Claim 28

The method of claim 1, wherein said step (c) includes a step wherein those of reference models of an initial set which correspond to said first adapted model group are updated therewith to form a set of partially updated reference models, and which repeats the following steps (d) through (f) upon each input of new training data after said step (c): (d) wherein those reference models to be trained in said partially updated reference model set which correspond to said new training data are adapted by said maximum a posteriori estimation method using said new training data to obtain adapted models, and said first group of adapted models is updated with said adapted models; (e) wherein, letting reference models in said partially updated reference model set corresponding to said first group of adapted models be identified as a first group of partially update reference models and the other reference models in said partially updated reference model set be identified as a second group of partially updated reference models, a second group of adapted models corresponding to said second group of partially updated reference models is obtained through interpolation by said vector field smoothing method using the relationship between said first group of adapted models and said first group of partially updated reference models corresponding thereto; and (f) wherein said reference models to be trained in said partially updated reference model set corresponding to said first group of adapted models updated in said step (d) are updated therewith, and said first group of adapted models obtained by said maximum a posteriori estimation method in said step (d) and said second group of adapted models obtained by said interpolation in said step (e) are combined into a set of updated adapted models for pattern recognition.

Claim 30

The method of claim 25, 26, or 27, wherein said reference models are continuous mixture hidden Markov models.

Claim 31

The method of claim 29, wherein said maximum a posteriori estimation method includes a process wherein: a statistical parameter of each of a plurality of distribution functions forming an output density distribution function of a model is to be estimated; said statistical parameter to be estimated is obtained by said maximum a posteriori estimation method using, as a control parameter common to all of said plurality of distribution functions of said output density distribution function, a statistical parameter of each prior distribution representing the reliability of said statistical parameter to be estimated; and a model to be trained in said training data is produced using distribution functions each including said statistical parameter obtained by said maximum a posteriori estimation method.

Claim 32

The method of claim 29, wherein said maximum a posteriori estimation method includes a process wherein: a statistical parameter of each of a plurality of distribution functions forming an output density distribution function of a model is to be estimated; said statistical parameter to be estimated is obtained by said maximum a posteriori estimation method using, as a control parameter corresponding to a statistical deviation of said training data for a prior distribution of each distribution function, a statistical parameter of said prior distribution representing the reliability of said statistical parameter to be estimated; and a model to be trained in said training data is produced using distribution functions each including said statistical parameter obtained by said maximum a posteriori estimation method.

Claim 33

The method of claim 30 or 31, wherein said distribution functions are each a Gaussian distribution and said statistical parameter to be estimated is a mean vector of said model.

Claim 34

The method of claim 30 or 31, wherein said distribution functions are each a Laplacian distribution and said statistical parameter to be estimated is a mean vector of said model.

Claim 35

The method of claim 30 or 31, wherein said vector field smoothing method in said step (b) includes a process for obtaining transfer vectors on the basis of variations of statistical parameters of distribution functions of said first group of adapted models and distribution functions of said first reference models and for determining weighting coefficients of said transfer vectors in accordance with the distributions of said statistical parameters of said distribution functions of said first group of reference models.

Claim 36

The method of claim 34, wherein said statistical parameters of said distribution functions are mean vectors of models and said variations of said statistical parameters are each obtained as the difference between said mean vectors of two corresponding models.

Claim 37

The method of claim 35, which further comprises a step of calculating said weighting coefficient as a fuzzy membership function.

Claim 38

The method of claim 35, which further comprises a step of calculating said weighting coefficient as a Gaussian window function.