KR100577515B1 - Method for choosing gaussian mixture number of hmm - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

The present invention relates to a method for setting the number of Gaussian mixtures of hidden Markov model parameters in a speech recognition system and a computer-readable recording medium recording a program for realizing the method.

2. The technical problem to be solved by the invention

The present invention provides a computer-readable recording method of increasing the number of mixtures during training and a method of setting the Gaussian mixture number of hidden Markov model parameters for finding the optimal number of mixtures and a program for realizing the method. To provide a record carrier.

3. Summary of Solution to Invention

The present invention includes a first step of initializing a Hidden Markov Model (HMM) parameter and the number of mixtures; Obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third step of changing the number of mixtures until the predetermined number is repeated and repeating from the second step; And a fourth step of testing a speech recognition rate using the Hidden Markov Model (HMM) parameter according to the number of the mixtures and setting the number of the mixtures as a result.

4. Important uses of the invention

The present invention is used in a voice recognition system.

Number of mixtures, hidden Markov model parameters, speech recognition, recognition rate

Description

How to set the number of Gaussian mixtures for the Hidden Markov model parameters {METHOD FOR CHOOSING GAUSSIAN MIXTURE NUMBER OF HMM}             

1 is an exemplary configuration diagram of a general voice recognition system.

2 is an exemplary diagram of a voice parameter training process in a general voice recognition system.

3 is an exemplary view of a speech recognition process in a general speech recognition system.

Figure 4 is a flow diagram of a process of training the voice parameter through the conventional number of mixtures setting.

5 is a flowchart illustrating a method for setting a Gaussian mixture number of hidden Markov model parameters according to the present invention.

FIG. 6 is an exemplary view of CDP recognition rate trained by extending the number of mixtures of CIP used in the present invention. FIG.

* Explanation of symbols for the main parts of the drawings

11: end point detector 12: feature extraction unit

13: pattern comparison part 14: reference pattern part

The present invention relates to a method of setting the number of Gaussian mixtures of a Hidden Markov Model (HMM) parameter and a computer-readable recording medium recording a program for realizing the method. A method of setting the number of Gaussian mixtures of hidden Markov model parameters used in voice parameter training, such as a recognition system, and a computer-readable recording medium recording a program for realizing the method.

1 is an exemplary configuration diagram of a general voice recognition system.

Operation in the speech recognition system to which the present invention is applied is as follows.

First, when a user's voice is input, the endpoint detector 11 first searches for a voice section excluding the silent section before and after the voice. Next, the feature extractor 12 extracts the feature of the speech from the signal of the speech section found above. The pattern comparison unit 13 performing the actual speech recognition operation compares the extracted voice feature data with the reference pattern data previously included in the reference pattern unit 14 to find the most similar reference pattern and uses the recognition result as the result.

The method proposed by the present invention corresponds to the reference pattern unit 14 of FIG. 1, and the Hidden Markov Model (HMM) parameter is one of the reference patterns. The present invention may be referred to as an invention relating to the generation of a reference pattern for increasing the recognition rate.

Speech recognition in the speech recognition system can be considered in two stages. The first step is the training process for generating the parameters to be used for recognition, and the second step is the actual recognition process for the input voice. Hidden Markov Medel (HMM) parameters are generated during training and are used later in the recognition process.

HMM is a widely used training method in speech recognition systems, and individual phoneme unit parameters may be configured as discrete output probability or continuous output probability.

In addition, the HMM parameter consists of state transition probabilities and observation probabilities for each phoneme, and there are differences between the two methods in configuring the observation probabilities. First, in the case of the discrete-hidden Markov model (discrete-HMM) with discrete output probability, we use a codebook, which is a representative value of the number and type of codewords that we set. Will be expressed.

On the other hand, in the case of the continuous-hidden Markov model (continuous-HMM) to which the present invention is applied, the probability value is configured to have a continuous probability density and is expressed as an average and a variance using a Gaussian distribution function. In this case, how many Gaussian functions are used affects the recognition performance, which is called the number of mixtures.

2 and 3 are exemplary flow charts of a voice parameter training process and a voice recognition process when using the HMM parameter in the voice recognition system.

2 is an exemplary diagram of a voice parameter training process in a general voice recognition system.

In the voice parameter training process in the voice recognition system, first, a voice file for training is read from a training voice file database to extract a feature (201). At this time, the feature is extracted for each frame.

The phoneme division information is received from the phoneme division information database, and the HMM phoneme parameter is initialized (202).

A pronunciation dictionary is generated using the basic phoneme database (203) and stored in the pronunciation dictionary database.

HMM independent phoneme parameter training is performed with the initialized HMM independent phoneme parameters obtained from the data from the pronunciation dictionary database, the feature information extracted from the training speech file, and the phoneme unit segmentation information (204). Independent phoneme training generates context-independent (CI) parameters.

In addition, the HMM-dependent phoneme (CD: Context-Dependent) parameter is obtained by extending the feature information extracted from the pronunciation dictionary database and the feature information extracted from the training speech file and the context-independent phoneme parameter (205). Perform training (206).

The HMM parameter is generated through the voice parameter training process as described above.

3 is an exemplary view of a speech recognition process in a general speech recognition system.

3 illustrates a process of performing voice recognition with voice data for recognition in a voice recognition system capable of real voice recognition.

In the speech recognition system, the speech data for recognition is read from the speech database for recognition and the feature is extracted (301).

The speech recognition is performed by using the pronunciation dictionary, the HMM parameter, and grammar information extracted from the speech data (302).

At this time, the pronunciation dictionary is generated from the basic phoneme database as shown through the description of FIG. 2, and when the data "grandfather" is stored, "grandfather" is "ㅎ + ㅏ + ㄹ + ㅇ + ㅏ + ㅂ + 음 + ㅈ + ㅣ ".

The HMM parameter has the characteristics of each phoneme, and has, for example, a probability value indicating a characteristic of "h", a probability value indicating a characteristic of "ㅏ", and the like.

The grammar information is determined by the scenario flow of the field in which speech recognition is performed in the speech recognition system. Therefore, the grammar information can be used to enumerate candidate values of speech data that may be present in a corresponding state. Information. For example, in the case of “voice recognition system for train ticket reservation”, the probability value of words other than the station name is zero in the step of entering the departure station name, and the probability value of words other than hours and minutes is zero in the step of entering the departure time. to be. The grammar information has a probability that each candidate word in the speech recognition process can appear.

In the speech recognition system, there is a general speech parameter training process and a speech recognition process. At this time, the HMM parameter setting process is performed to increase the recognition rate in the speech parameter training process.

The present invention relates to the efficient setting of the number of Gaussian mixtures of HMM parameters when the HMM parameters are set. In the related art, the performance of the voice recognition system was performed after the voice recognition system operator designated the number of mixtures to confirm the performance. .

If this is shown through the drawings as shown in FIG.

4 is a flowchart illustrating a process of training a voice parameter through a conventional number of mixtures.

This is a flowchart illustrating a description of the voice parameter training process in FIG. 2.

 First, the HMM parameter is initialized and the number of mixtures is designated (401).

Phoneme is estimated by performing forward-backward segmentation (402). That is, when there is input data such as "grandfather", the prediction is performed by dividing the position of a frame occupied by each phoneme "ㅎ", "ㅏ" and the like.

The HMM parameter is estimated (403). That is, the characteristics of the phonemes found for each frame are estimated.

After estimating the HMM parameters, it is determined whether this estimation has been performed sufficiently to reach the desired goal (404). In this case, the determination criterion may be a parameter estimation number or may be referred to as a target value designated by an operator. Since this is a voice parameter training process, it is to judge how well the data obtained by training represents the characteristics of the input data.

If the desired target has not been reached, the process is repeated from step 402 of performing forward-backward segmentation.

When the desired goal is reached, the number of the mixtures presented at the beginning is set (405) and the process ends. In this case, if the desired target is the parameter estimation training number, the training is performed as many times as the number of times, and the operator reports the result and the operator gives a different number of mixtures and then performs the work again.

As described above, although the number of mixtures has a significant influence on the performance of the voice recognition system, it is difficult to make an accurate judgment because the designated number is set by the operator's judgment.

The present invention has been made to solve the above problems, the method and method of setting the Gaussian mixture number of hidden Markov model parameters to increase the number of mixtures in the training process and find the optimal number of mixtures An object of the present invention is to provide a computer-readable recording medium having recorded thereon a program for realizing this.

The method of the present invention for achieving the above object is, in the method of setting the number of Gaussian (Mixture) Mixture of the Hidden Markov Model (HMM) parameter applied to the speech recognition system, HMM) a first step of initializing the number of parameters and the mixture (mixture); Obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third step of changing the number of mixtures until the predetermined number is repeated and repeating from the second step; And a fourth step of testing a speech recognition rate using the Hidden Markov Model (HMM) parameter according to the number of the mixtures and setting the number of the mixtures as a result.

In addition, another method of the present invention is a method for setting the number of Gaussian mixtures of a hidden markov model (HMM) parameter applied to a speech recognition system, wherein the hidden markov model (HMM) parameter is used. And a first step of initializing the number of the mixtures. Performing a speech recognition rate test by obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third step of changing the number of mixtures until the predetermined number is repeated and repeating from the second step; And a fourth step of setting the number of mixtures by comparing the voice recognition rate test results according to the number of each mixture.

The present invention also provides a speech recognition system having a large-capacity processor, comprising: a first function of initializing a Hidden Markov Model (HMM) parameter and the number of mixtures; A second function of obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third function of changing the number of mixtures until a predetermined number and repeatedly performing the second step; And a program for realizing a fourth function of testing a speech recognition rate with the Hidden Markov Model (HMM) parameter according to the number of the mixtures, and as a result, setting the number of the mixtures. It provides a recording medium that can be read by.

The present invention also provides a speech recognition system having a large-capacity processor, comprising: a first function of initializing a Hidden Markov Model (HMM) parameter and the number of mixtures; A second function of performing a speech recognition rate test by obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third function of changing the number of mixtures until a predetermined number and repeatedly performing the second step; And a computer-readable recording medium having recorded thereon a program for realizing a fourth function of setting the number of mixtures by comparing the voice recognition rate test results according to the number of each mixture. .

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.

5 is a flowchart illustrating a method of setting the number of Gaussian mixtures of a hidden Markov model parameter according to the present invention.

The present invention relates to setting the number of Gaussian mixtures and to a method of finding the number of mixtures having the best recognition performance. The initial value of the HMM parameter was generated using phoneme segmentation information. This refers to information that indicates which phonemes are uttered in a voice file by section (frame unit), and is determined by a phonetician based on the sound and waveform.

Here, the phoneme is a context-independent-phoneme (CIP), which is the basic phoneme defined and used in the speech recognition system of KT. It is defined as a phoneme that is more subdivided than the phoneme in Korean. Doing. The phoneme split information is used to generate an initial value of the HMM parameter in phoneme units. In this case, the mixture may be generated by splitting one or N pieces. In the present invention, an initial value is generated by one mixture.

HMM parameter estimation process is performed after the forward-backward segmentation process that separates the voice file by phoneme segment using the initial parameter values, training voice files, and pronunciation dictionary of each voice file. This series of processes is called parametric training.

The initial parameters created as one mixture go through the HMM training process using the training voice file, and then again generate parameters consisting of one mixture until the maximum mixture desired is generated. Repeatedly increasing the mix by one (here, the maximum mixture may be referred to as the maximum mixture allowed by the system's memory). Each time the HMM parameter of each mix unit is generated, the value is stored.

When the maximum number of mixtures is repeated to generate parameters consisting of each number of mixtures, the recognition performance is tested as voice files for recognition test using individual HMM parameters to obtain the best recognition performance. Selects the mixture parameter to indicate.

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

First, the HMM parameter is initialized, and the number of mixtures is set to 1 (501). It is determined whether the number of current mixtures is greater than the maximum number of mixtures (502). At this time, the maximum number of mixtures is determined by the memory capacity of the system or the setting of the operator.

If the number of current mixtures is not greater than or equal to the maximum number of mixtures, the voice file is separated by phoneme sections by performing forward-backward segmentation (503).

In operation 504, HMM parameter estimation is performed. Verify that the estimated HMM parameter has converged to the desired goal (505). As a result of the check, if it has not converged, the process is repeated from the forward-backward segmentation process 503 for separating the phoneme sections.

As a result of confirming that the target has been converged to the desired target, when the target has converged to the desired target, the obtained HMM parameter is stored (506), and the number of the mixtures is increased by 1 to increase the number of the current mixtures. The process is repeated from step 502 to determine whether the number of mixtures is greater than that.

If the number of current mixtures is greater than the maximum number of mixtures, the recognition rate is tested with the HMM parameter at each mixture number (508), and the optimal number of mixtures is used. Set the number (509).

In the above embodiment, after obtaining the HMM parameter for each number of mixtures, the optimal number of mixtures is set through the recognition rate test as a whole.However, the number of each mixture is obtained. There is also a method of obtaining an optimum number of mixtures by performing a recognition rate test each time, performing a recognition rate test up to the maximum number of mixtures, and comparing the result values obtained.

According to the method described above, the number of mixtures and the HMM parameters are obtained. The HMM parameters are referred to as HMM context-independent parameters (CIPs).

Real speech recognition system uses context-dependent parameter (CD) consisting of phonemes extended from CI-HMM to improve recognition performance.

CIPs are subdivided into CDPs based on the types of phonemes in the front and back context. This process is determined by the words in the domain. Here, the domain refers to the field where voice recognition is used. For example, in the case of voice recognition of railway stations, all target railway stations become domains in the voice recognition system.

For the corresponding domain word in the CDP, for example, even though the same "a" phonemes, the "a" of the superstar and the "a" of the star are strictly different in sound, and again the "a" in front and behind The sound is different depending on which phonemes are present. In this way, it is extended to CDP in consideration of the phoneme environment before and after.

When the phoneme unit is expanded to CDP, the same training process as the HIP parameter training process of the CIP unit is repeated. By the way, if the extension to the CDP to increase the value of the mixture (Ctureture) in the CIP training process, the HMM parameter of the CDP unit can generate only the mixture (mixture) more than the number of the mixture (mixture).

For example, if you created three mixtures during the CIP phoneme training phase of the HMM, if you continue with the training by expanding to CDP here, the three mixtures will start and increase.

That is, since it does not reduce from a large mixture to a small number of mixtures, it is a question of how many mixtures should be extended to the CDP during the training phase of the CIP. According to the experiments, the HMM training process of the CIP stage showed the best performance by increasing the mix until the number of mixtures, that is, one, was expanded to CDP and the recognition rate was the best. If this is presented through the drawings as shown in FIG.

FIG. 6 is an exemplary view of CDP recognition rate trained by extending the number of mixtures of CIP used in the present invention.

As shown in the figure, it was confirmed that it is better to repeat the training in the expanded state to the appropriate CDP than to repeat the training too excessively in the CIP stage.

The parameters of each mixture are saved through HMM training in CDP units, and the recognition performance is tested using the parameters of each mixture when the desired maximum number of mixtures are generated. do. In this case, the parameter representing the best performance is applied to the speech recognition system.

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.

As described above, the present invention can easily find the number showing the excellent performance among the number of the mixtures of the continuous probability hidden Markov model (HMM), and thus, N mixtures designated by the existing operators. ) Is much easier and faster to process than getting parameters.

In addition, the present invention generates and stores HMM parameters in the number of each mixture, so that any mix is required when extending voice parameter training in context independent phonemes (CIP) to training in context dependent phonemes (CDP). It's easy to see if it's good to expand on the number of mixes.

Claims (9)

In the method of setting the number of Gaussian mixtures of the Hidden Markov Model (HMM) parameter applied to the speech recognition system, A first step of initializing the hidden Markov model (HMM) parameter and the number of mixtures; Obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third step of changing the number of mixtures until the predetermined number is repeated and repeating from the second step; And A fourth step of testing a speech recognition rate using the Hidden Markov Model (HMM) parameter according to the number of each mixture and setting the number of mixtures as a result How to set the number of Gaussian mixtures of Hidden Markov model parameters, including. The method of claim 1, The second step, Perform a forward-backward segmentation operation to separate the voice file by phoneme sections with the Hidden Markov Model (HMM) parameter initial value, the number of the mixtures, the training voice file, and the pronunciation dictionary. A fifth step; A sixth step of obtaining an estimated value for the Hidden Markov Model (HMM) parameter with the speech file separated for each phoneme section; A seventh step of determining whether an estimated value of the Hidden Markov Model (HMM) parameter converges to a predetermined value; An eighth step of repeating from the fifth step if it does not converge to the predetermined value as a result of the determination of the seventh step; And A ninth step of storing the estimated value of the Hidden Markov Model (HMM) parameter as the Hidden Markov Model (HMM) parameter in the number of mixtures when converging to the predetermined value as a result of the determination in the seventh step. How to set the number of Gaussian mixtures of Hidden Markov model parameters, including. In the method of setting the number of Gaussian mixtures of the Hidden Markov Model (HMM) parameter applied to the speech recognition system, A first step of initializing the hidden Markov model (HMM) parameter and the number of mixtures; Performing a speech recognition rate test by obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third step of changing the number of mixtures until the predetermined number is repeated and repeating from the second step; And A fourth step of setting the number of mixtures by comparing the voice recognition rate test results according to the number of each mixture How to set the number of Gaussian mixtures of Hidden Markov model parameters, including. The method of claim 3, wherein The second step, Performing a forward-backward segmentation operation for separating the voice file by phoneme sections with the Hidden Markov Model (HMM) parameter initial value, the number of the mixtures, the training voice file, and the pronunciation dictionary. A fifth step; A sixth step of obtaining an estimated value for the Hidden Markov Model (HMM) parameter with the speech file separated for each phoneme section; A seventh step of determining whether an estimated value of the Hidden Markov Model (HMM) parameter converges to a predetermined value; An eighth step of repeating from the fifth step if it does not converge to the predetermined value as a result of the determination of the seventh step; A ninth step of storing the estimated value of the Hidden Markov Model (HMM) parameter as the Hidden Markov Model (HMM) parameter in the number of mixtures when converging to the predetermined value as a result of the determination in the seventh step; And A tenth step of performing a voice recognition rate test with the Hidden Markov Model (HMM) parameter in the number of mixtures and storing the result; How to set the number of Gaussian mixtures of Hidden Markov model parameters, including. The method according to any one of claims 1 to 4, The predetermined number is Method for setting the Gaussian Mixture number of Hidden Markov model parameters, characterized in that the number to indicate the range that can be set the number of mixtures. The method of claim 5, wherein The Hidden Markov Model (HMM) parameter is A method for setting the number of Gaussian mixtures of a Hidden Markov Model parameter, which is substantially a Hidden Markov Model (HMM) parameter in Context-Independent-phoneme (CIP) units. The method of claim 5, wherein The Hidden Markov Model (HMM) parameter is A method for setting the Gaussian mix number of Hidden Markov Model Parameters, which is substantially a Hidden Markov Model (HMM) parameter in units of Context-Dependent-phoneme (CDP). In voice recognition system with a large processor, A first function of initializing the Hidden Markov Model (HMM) parameters and the number of mixtures; A second function of obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third function of changing the number of mixtures until a predetermined number and repeatedly performing the second step; And A fourth function of testing a speech recognition rate with the Hidden Markov Model (HMM) parameter according to the number of each mixture and setting the number of mixtures as a result A computer-readable recording medium having recorded thereon a program for realizing this. In voice recognition system with a large processor, A first function of initializing the Hidden Markov Model (HMM) parameters and the number of mixtures; A second function of performing a speech recognition rate test by obtaining the Hidden Markov Model (HMM) parameter according to the number of the mixtures; A third function of changing the number of mixtures until a predetermined number and repeatedly performing the second step; And A fourth function of setting the number of mixtures by comparing the voice recognition rate test results according to the number of each mixture A computer-readable recording medium having recorded thereon a program for realizing this.

Images