KR100633228B1 - Method for presenting gaussian probability density and method of speech recognition training for obtaining the same - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

The present invention relates to a Gaussian probability density representation method, a speech recognition training method for obtaining the same, and a computer readable recording medium having recorded thereon a program for implementing the methods.

2. The technical problem to be solved by the invention

The present invention divides a probability density space into a plurality of PGAM spaces, and then expresses this space as a GAM, and then represents a Gaussian probability density representation method representing a GAM having the largest value as a representative probability density, a speech recognition training method for obtaining the same, and the method. To provide a computer-readable recording medium that records a program for realizing them.

3. Summary of Solution to Invention

The present invention includes a first step of obtaining a trained speech recognition model through the recognition training receives the speech recognition model initialized the number of mix groups; A speech recognition test is performed on the trained speech recognition model to perform a recognition training on a misrecognized word to obtain a new speech recognition model and combine with the speech recognition model obtained in the first step to increase the number of mixture groups. A second step of obtaining a recognition model; A third step of performing recognition training on the speech recognition model obtained in the second step and confirming that the number of mixture groups has reached a desired number; A fourth step of repeating from the second step if the number of mixture groups does not reach a desired number as a result of the checking of the third step; And a fifth step of completing voice recognition training when the number of mixture groups reaches a desired number as a result of the checking of the third step.

4. Important uses of the invention

The present invention is used in a speech recognition system and the like.

Speech Recognition, Probability Density, Speech Recognition Training, Gaussian

Description

METHOD FOR PRESENTING GAUSSIAN PROBABILITY DENSITY AND METHOD OF SPEECH RECOGNITION TRAINING FOR OBTAINING THE SAME}             

1 is an exemplary configuration diagram of a general computer.

2 is an exemplary view of a probability model for expressing a conventional probability density.

3 is an exemplary explanatory diagram of a probability density model according to the present invention;

4 is a flowchart illustrating a Gaussian probability density representation method according to the present invention.

5 is a flowchart of an embodiment of a speech recognition training method for obtaining an HPGAM according to the present invention;

* Explanation of symbols for the main parts of the drawings

1: input device 2: central processing unit

3: main memory device 4: auxiliary storage device

5: display device

The present invention relates to a Gaussian probability density representation method, a speech recognition training method for obtaining the same, and a computer readable recording medium recording a program for realizing the methods, in particular, a Gaussian Autoregressive mixture (GAM). ) And a partitioned Gaussian Autoregressive mixture (PGAM) that combine Gaussian probability density to express Gaussian probability density, to obtain speech recognition training methods for obtaining them, and to a computer readable program that records the programs to realize the methods. It relates to a recording medium.

1 is an exemplary configuration diagram of a general computer.

In general, a computer includes an input device 1, a central processing unit 2, a main memory device 3, an auxiliary memory device 4, and a display device 5. A method of expressing probability density in such a computer and a method of performing speech recognition training using the same are implemented.

2 is an example of a probability model for expressing a conventional probability density.

Represents a widely used probability model. In other words, two Gaussian Mixture probability distributions (PDF) are overlapped to form one probability distribution (PDF) (dotted line).

When two Gaussian distributions with different probability densities are combined, the Mixture model weights each Gaussian distribution and simply adds them to create a new probability distribution (PDF). This is called the Gaussian Autoregressive mixture (GAM) model, and it considers the largest probability distribution of two Gaussian distributions and ignores the other probability distributions for the partitioned Gaussian Autoregressive mixture. It is called a PGAM (Partitioned Gaussian Autoregressive mixture) model.

When probability density is expressed as a sum of Gaussian density, two methods, called GAM and PGAM, are used. GAM expresses the probability density as a weighted sum of Gaussian density. In addition, PGAM expresses the probability space as Gaussian and represents the largest value as probability density and ignores the other probability density values.

However, there is a problem that the speech recognition rate is not high enough to be applied to speech recognition using the probability density expression method described above. Therefore, there is a demand for a better probability density representation method for improving speech recognition rate.

The present invention has been devised to solve the above problems, which is divided into several PGAM spaces, and then expressed as a GAM, and then a Gaussian probability density representing a GAM having the largest value as a representative probability density. It is an object of the present invention to provide a computer-readable recording medium recording a representation method and a program for realizing the method.                         

In addition, the present invention provides a voice recognition training method for combining and recognizing data that has been misidentified with an existing training method to obtain the Gaussian probability density expression, and a computer-readable recording medium having recorded thereon a program for realizing the method. That has another purpose.

In order to achieve the above object, the present invention relates to a Gaussian probability density expression method applied to a speech recognition system, comprising: dividing a probability density space into a plurality of partitioned Gaussian Autoregressive Mixture (PGAM) spaces. Stage 1; A second step of expressing the divided partitioned Gaussian auto-regressive mixture space as Gaussian Autoregressive Mixture (GAM); And a third step of expressing the Gaussian probability density by representing the GAM having the largest value as a hybrid partitioned Gaussian Autoregressive Mixture (HPGAM) which is a representative probability density.

According to another aspect of the present invention, there is provided a speech recognition system having a processor, comprising: a first function of dividing a probability density space into a plurality of partitioned Gaussian Autoregressive Mixture (PGAM) spaces; A second function of expressing the divided space in Gaussian Autoregressive Mixture (GAM); And a computer that records a program for realizing a third function of expressing Gaussian probability density by expressing the GAM having the largest value as a hybrid partitioned Gaussian Autoregressive Mixture (HPGAM), which is a representative probability density. Provide a readable recording medium.

According to another aspect of the present invention, there is provided a speech recognition training method for obtaining a hybrid partitioned Gaussian auto-regressive mixture applied to a speech recognition system, wherein the first training data is set using the number of mixture groups as an initial value. A first step of performing training on the first training data; A second step of performing a recognition test on the first training data on which the training is performed to perform a training on a misunderstanding word; A third step of combining the training data for the misrecognized word and the first training data to obtain new second training data of which the number of mixture groups is increased by one, and performing training on the second training data; A fourth step of checking whether the number of mixture groups of the training data of the third step reaches a desired number; A fifth step of repeating from the second step if the number of mixture groups does not reach a desired number in the fourth step; And a sixth step of terminating the speech recognition training when the number of the mixture groups reaches the desired number in the fourth step.

The present invention also provides a speech recognition system including a processor, comprising: a first step of setting first training data using the number of mixture groups as an initial value and then performing training on the first training data; A second step of performing a recognition test on the first training data on which the training is performed to perform a training on a misunderstanding word; A third step of combining the training data for the misrecognized word and the first training data to obtain new second training data of which the number of mixture groups is increased by one, and performing training on the second training data; A fourth step of checking whether the number of mixture groups of the training data of the third step reaches a desired number; A fifth step of repeating from the second step if the number of mixture groups does not reach a desired number in the fourth step; And a computer-readable recording medium having recorded thereon a program for realizing a sixth step of terminating voice recognition training when the number of mixture groups reaches the desired number in the fourth step.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

3 is an exemplary explanatory diagram of a probability density model according to the present invention.

3 shows two probability density (PDF) models. The hybrid partitioned Gaussian auto proposed by the present invention has a conventional Gaussian auto-regressive mixture model with 12 mixtures, 4 mixture groups and 3 mixtures. Representative HPGAM model.

In the present embodiment, a method of expressing probability density will be described. When probability density is expressed as a sum of Gaussian density, two methods, called GAM and PGAM, have been used.

GAM expresses the probability density as a weighted sum of Gaussian density. In addition, PGAM expresses the school space as Gaussian, and represents the largest value as probability density and ignores the remaining probability density values.

The method proposed in the present invention combines the above two and divides the probability density space into several PGAM spaces, and then expresses the GAM having the largest value as the representative probability density. In other words, the proposed Gaussian probability density method combines GAM and PGAM, and we call this method Hybrid Partitioned Gaussian Autoregressive mixture (HPGAM).

4 is a flowchart illustrating a Gaussian probability density representation method according to the present invention.

First, a probability density space is divided into a plurality of PGAM spaces (401), and this space is represented by GAM (402). The GAM having the largest value is represented as the representative probability density to represent the Gaussian probability density (403).

5 is a flowchart illustrating an embodiment of a speech recognition training method for obtaining an HPGAM according to the present invention.

In addition, this embodiment describes a training method for HPGAM. This method obtains HPGAM by training according to the existing training method and reinitializing the newly recognized data. That is, as shown in FIG. 5, a voice recognition model (1 mixture group and N mixtures) of a desired number of mixtures is obtained in a conventional manner, and then a voice recognition test is performed using this model. From these speech recognition results, a new speech recognition model (1 mixture group and N mixtures) is obtained using the misrecognition words. This model is combined with the previous model to obtain a new speech recognition model (2 mixture group and N mixtures) and trained using this model. After that, the voice recognition test is repeated with this model until the desired mix groups are obtained. This training compensates for the problem with initialization that is a disadvantage of the existing training.

This will be described according to the flow of the drawings.

First, training data is set with the number of mixture groups as 1 (501), and segmental k-means training and forward-backward training are performed on the training data. (502).

A recognition test is performed on the training data on which the training is performed (503), and segmental k-means training and forward-backward training on misperceived words are performed (504, 505).

Combining the existing training data with the training data obtained by training the misperceived words to obtain new training data with an increase in the number of mix group groups (506), performing forward-backward training (507), and mixing the training data. It is determined whether the number of the group groups is equal to the desired number of mix groups in step 508. As a result of the determination, if it is not the same, the number does not reach the number, and the process is repeated from the process 503 of performing the recognition test on the training data. As a result of the determination, the voice recognition training process for the training data is terminated.

The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes are possible in the art without departing from the technical spirit of the present invention. It will be clear to those of ordinary knowledge.

According to the present invention as described above, in order to improve the speech recognition rate, it is easier to obtain HPGAM with better recognition rate than the existing GAM or PGAM, and according to the speech recognition training to compensate for the problems caused by the initialization, which is a disadvantage of the existing training. There is an effect that can improve the speech recognition rate.

Claims (4)

In the Gaussian probability density expression method applied to the speech recognition system, A first step of dividing the probability density space into a plurality of partitioned Gaussian Autoregressive Mixture (PGAM) spaces; A second step of expressing the divided partitioned Gaussian auto-regressive mixture space as a Gaussian Autoregressive Mixture (GAM); And The third step of expressing the Gaussian probability density by representing the GAM having the largest value as a hybrid partitioned Gaussian Autoregressive Mixture (HPGAM), which is the representative probability density. Gaussian probability density expression method comprising a. In the speech recognition training method for obtaining a hybrid partitioned Gaussian auto-regressive mixture applied to a speech recognition system, A first step of setting first training data using the number of mixture groups as an initial value and then performing training on the first training data; A second step of performing a recognition test on the first training data on which the training is performed to perform a training on a misunderstanding word; A third step of combining the training data for the misrecognized word and the first training data to obtain new second training data of which the number of mixture groups is increased by one, and performing training on the second training data; A fourth step of checking whether the number of mixture groups of the training data of the third step reaches a desired number; A fifth step of repeating from the second step if the number of mixture groups does not reach a desired number in the fourth step; And A sixth step of terminating voice recognition training when the number of mixture groups reaches a desired number in the fourth step; Speech recognition training method comprising a. In the speech recognition system having a processor, A first function of dividing a probability density space into a plurality of partitioned Gaussian Autoregressive Mixture (PGAM) spaces; A second function of expressing the divided partitioned Gaussian Autoregressive Mixture space as Gaussian Autoregressive Mixture (GAM); And A third function that expresses Gaussian probability density by expressing GAM having the largest value as Hybrid Partitioned Gaussian Autoregressive Mixture (HPGAM), which is the representative probability density. A computer-readable recording medium having recorded thereon a program for realizing this. In a speech recognition system having a processor, A first step of setting first training data using the number of mixture groups as an initial value and then performing training on the first training data; A second step of performing a recognition test on the first training data on which the training is performed to perform a training on a misunderstanding word; A third step of combining the training data for the misrecognized word and the first training data to obtain new second training data of which the number of mixture groups is increased by one, and performing training on the second training data; A fourth step of checking whether the number of mixture groups of the training data of the third step reaches a desired number; A fifth step of repeating from the second step if the number of mixture groups does not reach a desired number in the fourth step; And The computer-readable recording medium having recorded thereon a program for realizing the sixth step of terminating the speech recognition training when the number of mixture groups reaches the desired number in the fourth step.

Images