Pattern recognition apparatus, pattern recognition method, and storage medium

1. A pattern recognition apparatus based on a Neural Network (NN), comprising:

at least one memory storing a set of instructions; and

at least one processor configured to execute the set of instructions to:

train an NN model to generate NN parameters, based on first feature vectors and domain vectors respectively indicating subsets in a plurality of domains respectively indicating languages other than a specific language,

wherein the first feature vectors are respectively extracted from the subsets, and the domain vectors respectively indicate identifiers corresponding to the subsets;

verify a pair of second feature vectors in a specific domain to output whether the pair is extracted from first data and second data of a same individual or not, based on a target domain vector and the NN parameters, the specific domain indicating the specific language, the target domain vector indicating a subset in the specific domain,

wherein the at least one processor trains the NN model by using the domain vectors as training input to NN model, and verifies the pair by using the target domain vector as verification input to NN model;

read, from a storage, the subsets being speech recordings in the plurality of domains other than the specific domain, the subsets being respectively associated with domains and speakers labels;

extract sets of the first feature vectors respectively from the subsets;

calculate, as the domain vectors, mean vectors of the sets of feature vectors;

store the domain vectors in the storage;

store the NN parameters in the storage after training the NN model;

read the subset in the specific domain from the storage, the subset in the specific domain being speech recordings in the specific domain;

extract third feature vectors from the subset;

calculate, as the target domain vector, a mean vector of the third feature vectors;

store the target domain vectors in the storage;

read enrollment data in the specific domain from the storage;

extract, as one of the pair of second feature vectors, an enrollment feature vector from the enrollment data;

read test data in the specific domain from the storage;

extract, as another of the pair of second feature vectors, a test feature vector from the test data;

read the target domain vector from the storage;

read the NN parameters from the storage; and

verify the pair of second feature vectors.

2. The pattern recognition apparatus according to claim 1 , wherein

the at least one processor is further configured to utilize a specific subset in the specific domain as the target domain vector.

3. The pattern recognition apparatus according to claim 1 , wherein

the at least one processor is further configured to calculate means respectively corresponding to the subsets as the domain vectors thereof.

4. The pattern recognition apparatus according to claim 1 , wherein

the at least one processor is further configured to train a multi-layer perceptron (MLP) based on the first feature vectors to generate MLP parameters to extract the domain vectors corresponding to the subsets.

5. The pattern recognition apparatus according to claim 1 , wherein

the at least one processor is further configured to:

train a multi-layer perceptron (MLP) with the NN model training to generate MLP-NN parameters, based on the first feature vectors; and

verify the pair of second feature vectors based on the MLP-NN parameters.

6. A pattern recognition method using a Neural Network (NN), comprising:

training an NN model to generate NN parameters, based on first feature vectors and domain vectors respectively indicating subsets in a plurality of domains respectively indicating languages other than a specific language,

wherein the first feature vectors are respectively extracted from the subsets, and the domain vectors respectively indicate identifiers corresponding to the subsets;

verifying a pair of second feature vectors in a specific domain to output whether the pair is extracted from first data and second data of a same individual or not, based on a target domain vector and the NN parameters, the specific domain indicating the specific language, the target domain vector indicating a subset in the specific domain,

wherein the NN model is trained by using the domain vectors as training input to NN model, and the pair is verified by using the target domain vector as verification input to NN model;

reading, from a storage, the subsets being speech recordings in the plurality of domains other than the specific domain, the subsets being respectively associated with domains and speakers labels;

extracting sets of the first feature vectors respectively from the subsets;

calculating, as the domain vectors, mean vectors of the sets of feature vectors;

storing the domain vectors in the storage;

storing the NN parameters in the storage after training the NN model;

reading the subset in the specific domain from the storage, the subset in the specific domain being speech recordings in the specific domain;

extracting third feature vectors from the subset;

calculating, as the target domain vector, a mean vector of the third feature vectors;

storing the target domain vectors in the storage;

reading enrollment data in the specific domain from the storage;

extracting, as one of the pair of second feature vectors, an enrollment feature vector from the enrollment data;

reading test data in the specific domain from the storage;

extracting, as another of the pair of second feature vectors, a test feature vector from the test data;

reading the target domain vector from the storage;

reading the NN parameters from the storage; and

verifying the pair of second feature vectors.

7. The pattern recognition method according to claim 6 , wherein

the verifying includes utilizing a specific subset in the specific domain as the target domain vector.

8. The pattern recognition method according to claim 6 , further comprising:

calculating means respectively corresponding to the subsets as the domain vectors thereof.

9. The pattern recognition method according to claim 6 , further comprising:

training a multi-layer perceptron (MLP) based on the first feature vectors to generate MLP parameters to extract the domain vectors corresponding to the subsets.

10. The pattern recognition method according to claim 6 , wherein

the training includes training a multi-layer perceptron (MLP) with the NN model training to generate MLP-NN parameters, based on the first feature vectors, and

the verifying includes verifying the pair of second feature vectors based on the MLP-NN parameters.

11. A non-transitory computer readable storage medium storing a pattern recognition program using a Neural Network (NN) for causing a computer to recognize a pattern, the program causing a computer to execute:

NN training processing of training an NN model to generate NN parameters, based on first feature vectors and domain vectors respectively indicating subsets in a plurality of domains respectively indicating languages other than a specific language,

wherein the first feature vectors are respectively extracted from the subsets, and the domain vectors respectively indicate identifiers corresponding to the subsets;

NN verification processing of verifying a pair of second feature vectors in a specific domain to output whether the pair is extracted from first data and second data of a same individual or not, based on a target domain vector and the NN parameters, the specific domain indicating the specific language, the target domain vector indicating a subset in the specific domain,

wherein the NN model is trained by using the domain vectors as training input to NN model, and the pair is verified by using the target domain vector as verification input to NN model;

read, from a storage, the subsets being speech recordings in the plurality of domains other than the specific domain, the subsets being respectively associated with domains and speakers labels;

extract sets of the first feature vectors respectively from the subsets;

calculate, as the domain vectors, mean vectors of the sets of feature vectors;

store the domain vectors in the storage;

store the NN parameters in the storage after training the NN model;

read the subset in the specific domain from the storage, the subset in the specific domain being speech recordings in the specific domain;

extract third feature vectors from the subset;

calculate, as the target domain vector, a mean vector of the third feature vectors;

store the target domain vectors in the storage;

read enrollment data in the specific domain from the storage;

extract, as one of the pair of second feature vectors, an enrollment feature vector from the enrollment data;

read test data in the specific domain from the storage;

extract, as another of the pair of second feature vectors, a test feature vector from the test data;

read the target domain vector from the storage;

read the NN parameters from the storage; and

verify the pair of second feature vectors.

12. The storage medium according to claim 11 , wherein

the NN verification processing utilizes a specific subset in the specific domain as the target domain vector.

13. The storage medium according to claim 11 , the program further causing a computer to execute

mean extraction processing of calculating means respectively corresponding to the subsets as the domain vectors thereof.

14. The storage medium according to claim 11 , the program further causing a computer to execute

multi-layer perceptron (MLP) training processing of training a MLP based on the first feature vectors to generate MLP parameters to extract the domain vectors corresponding to the sub sets.

15. The storage medium according to claim 11 , wherein

the NN training processing further trains a multi-layer perceptron (MLP) with the NN model training to generate MLP-NN parameters, based on the first feature vectors, and

the NN verification processing verifies the pair of second feature vectors based on the MLP-NN parameters.