KR100426844B1 - A method on the improvement of speaker enrolling speed for the MLP-based speaker verification system through improving learning speed and reducing learning data - Google Patents

Abstract

The present invention improves the learning speed by reducing the time required for speaker registration in the MLP-based speaker authentication system by using the method of improving the learning speed of the EBP and the background speaker reduction method in which the speaker grouping method is introduced in the existing speaker authentication method. The present invention relates to a method for improving registration speed of an MLP-based speaker identification system by reducing learning data. MLP (multilayer perceptron) has several advantages over other pattern recognition methods, and it is expected to be used as a speaker learning and recognition method of speaker identification system. However, learning of MLP takes considerable time because of the low speed of the error backpropagation (EBP) algorithm used for learning. This, together with the fact that many background speakers are required to achieve a high speaker recognition rate in speaker identification systems, creates a problem that it takes a long time to register a speaker in the system. This problem must be solved because the speaker authentication system must provide a proof service immediately after the speaker is registered. In order to solve this problem, the present invention uses a method of improving the learning speed of the EBP and a background speaker reduction method in which the speaker grouping method is introduced in the existing speaker identification method. Shorten.

Description

A method on the improvement of speaker enrolling speed for the MLP-based speaker verification system through improving learning speed and reducing learning data}

The present invention relates to a method for improving the registration speed of a speaker identification system. More specifically, the present invention relates to a method of improving the learning speed of an EBP and to using a background speaker reduction method in which a speaker grouping method is introduced in the existing speaker identification method. The present invention relates to a method of improving the registration speed of an MLP-based speaker identification system by improving the learning speed by reducing the time required for speaker registration in a speaker identification system and reducing the learning data.

MLP (multilayer perceptron) has several advantages over other pattern recognition methods, and it is expected to be used as a speaker learning and recognition method of speaker identification system. However, learning of MLP takes considerable time because of the low speed of the error backpropagation (EBP) algorithm used for learning. This, together with the fact that many background speakers are required to achieve a high speaker recognition rate in speaker identification systems, creates a problem that it takes a long time to register a speaker in the system. This problem must be solved because the speaker authentication system must provide a proof service immediately after the speaker is registered.

Therefore, the present invention is to solve the above-mentioned conventional problems, MLP-based speaker authentication using a method of improving the learning speed of the EBP, and a background speaker reduction method in which the speaker grouping method is introduced in the existing speaker identification method. The purpose of the present invention is to provide a method for improving the registration speed of the MLP-based speaker authentication system by reducing the learning speed and reducing the learning data.

1A to 1D are diagrams illustrating an EBP learning step according to the present invention.

Figure 1a is a two-model data distribution,

1b is a center position learning,

Figure 1c is distributed learning,

1D is a diagram illustrating contour learning.

Figure 2a to 2d is a view showing the repulsive force for each pattern of EBP learning according to the present invention.

Fig. 3a shows the registration speaker learning of the MLP in the case of the low density background speaker and Fig. 3b.

4 shows background speaker group selection of MLP-I and registration speaker learning of MLP-II.

5 is a flowchart of processing of a sustained sound MLP based speaker identification system.

Figure 6 is a graph of the experimental results of the online EBP algorithm according to an embodiment of the present invention.

7 is a graph of the results of COIL application according to an embodiment of the present invention.

Figure 8 is a graph of the experimental results using the DCS according to an embodiment of the present invention.

Figure 9 is a graph of the experimental results when the simultaneous application of COIL and DCS in accordance with an embodiment of the present invention.

10 is a graph of the improvement rate of each experimental result according to an embodiment of the present invention.

The method for improving the registration speed of the MLP-based speaker authentication system by improving the learning speed and reducing the learning data according to the present invention for achieving the above object is a frame of 30ms length that overlaps the input voice of a registered speaker sampled at 16bit 16kHz by 20ms. The 16th-order Mel interval filter bank is extracted for each frame and used to detect isolated words and continuous sounds.The filter bank coefficient is divided into coefficients up to 1kHz to remove the effect of the quantity on the entire spectral envelope. This value is subtracted from all coefficients and adjusted so that the average of all coefficients is zero again for effective learning of MLP, and the 50th-order 0-3kHz band equally spaced Mel filter bank is extracted for each frame. Filterbank coefficients are obtained by averaging the coefficients up to 1 kHz to remove the effect of the mass on the spectral envelope. Subtracted from all coefficients and speech analysis and feature extraction step of adjusting for effective learning of the MLP again, the average of all coefficients is zero; An isolated word detection and sustained sound detection step of detecting the isolated word and the sustained sound in the isolated word using the MLP trained to detect each continuous sound and the silence in a speaker-independent manner; In the case of basic online EBP or COIL application, the MLP is trained using registered speaker and background speaker data, while in the case of MLP-I and MLP-II application, each continuous sound of the whole isolated word is input to MLP-I by continuous sound. Then, the number of output neurons is averaged, and the average of these background values is selected in order of the highest n, and the background speakers are trained in the MLP-II using the n background speaker data selected for each continuous sound. The number of patterns used for learning is the continuous speech register speaker learning step of 10 per background speaker; In the case of basic online EBP or COIL application, each continuous sound of all isolated words is input to the MLP for each continuous sound, and the output neurons are averaged, while for the MLP-I and MLP-II applications, Each continuous sound is input to the MLP-I, and then the average of the output neurons is selected, and n background speakers are selected in descending order of the average, and one or more background speakers in the continuous speech register learning stage are selected among the selected background speakers. A sustained speech score evaluation step of averaging the output values of the MLP-II for all the sustained sounds and rejecting the requesting speaker when there is no continuous sound including one or more background speakers of the sustained-listed speaker learning step; Comparing the average value and the predetermined threshold value of the speaker score evaluation step for each continuous sound is characterized in that consisting of the registration word and the speaker score threshold comparison step to determine the final rejection / acceptance.

Hereinafter, exemplary embodiments of the present invention will be described with reference to the accompanying drawings.

In the age when various information devices such as computers, PDAs and mobile phones are utilized in everyday life, a means of protecting personal information is increasingly recognized. Among them, biometrics is a means of controlling access to such information devices by using unique features of each person, and it is excellent in convenience of security information management and excellent security capability itself. Major bio-characteristics are fingerprint, iris, face shape, and voice, but among these, voice is the one that we can easily access, and the cost of voice processing is relatively low compared to other features. .

Biometrics using speech is called speaker recognition. Speaker recognition identifies a speaker by registering several speakers in the system and selecting the speaker that matches the current voice, and determining whether the current voice matches the speaker after selecting the identity of a specific speaker registered in advance in the system. It is divided into speaker identification. Among them, the speaker identification is identified to an unspecified number of people, and the speaker identification system can be composed by combining a plurality of speaker identification modules.

Various recognition methods are used to determine the identity by learning the voice of a specific speaker and comparing the inputted voice with the learned voice. Among these, MLP (multilayer perceptron) has the following advantages and can be used in various recognition problems.

First, because it is nonparametric, there is no need for a lower probability distribution that must be assumed in the problem.

Second, the possibility of recognition error is minimized because it has the ability of rejection learning to distinguish the difference between each model to be learned.

Third, the feature space transformation ability is similar to that of linear discriminant analysis (LDA) when using a learning target of +1, 0 (or -1) for each learning model.

In general, learning an MLP takes considerable time. This is the main reason for the error backpropagation (EBP) algorithm used for learning MLP. The EBP algorithm is based on the maximum slope reduction method, and the final target is achieved by adjusting the weight while propagating the error between the current output and the target output of the MLP in the reverse direction from the output layer to the hidden layer. The slowness of EBP learning is due to the fact that the maximum slope reduction method uses only local information about the current weight.

On the other hand, the size of the learning data may also affect the learning speed of the MLP. In order to prove the speaker, a background speaker for comparison with the requesting speaker to measure the identity score is required, and the requesting speaker is divided into a real speaker and an impersonated speaker. For high speaker recognition rate, the requesting speaker should be compared with the background speaker with similar characteristics. However, because the characteristics of the calling speaker cannot be known in advance, as many background speakers as possible are prepared so that an accurate determination can be made even if a speaker requests identification. However, it is obvious that increasing the number of background speakers to achieve a high recognition rate also increases the registration time of the MLP speaker identification system with a long learning time.

The speaker identification system should not only be able to provide identification services at the same time as the speaker is registered, but also should be promptly verified in consideration of the convenience of frequently used information devices. The speaker registration time of the system using MLP is long due to the two factors described above, but the speaker identification time is satisfactory in view of the processing capacity of general information equipment because of the fast operation characteristic of MLP based on the distinction function. Accordingly, efforts should be focused on improving the registration speed to ensure real-time performance of the MLP-based speaker authentication system.

Attempts to improve the learning speed of MLP have been made in two directions. The first direction uses experience and experimental results. If the output is far from the target, the learning rate is increased and if it is close, the learning rate is decreased. This is divided into a method of changing the global learning rate that collectively affects the entire weight vector and a method of changing the local learning rate according to the change of the maximum slope for each weight. The second direction uses an optimization theory, which uses second derivatives of the weights. This class includes the use of momentum to reflect previous learning trends in the current update, or use Newton's optimization theory or a modified algorithm to calculate weighted vector updates that can converge to the target at the fastest. have.

Weight updating is done in two ways. One method is to apply the average of the change values after presenting all the training data, and the other is to apply the change value whenever the training data is presented one by one. The former is called on-line (or probabilistic) and the latter is called off-line (or batch). In both methods, one period in which all the learning data is presented is called an epoch, and the difference between the MLP target value and the output value is determined for each epoch to determine whether to continue the learning.

When MLP is used in pattern recognition including speaker identification, online learning is faster than offline learning. The following reasons can be found.

First, all the patterns in the model imply significant redundancy for each other, which contributes to the calculation of the maximum slope of the EBP. For this reason, the more patterns included in the model, the faster the learning speed in units of epochs.

Second, if the maximum slope of the patterns in the model is within 90 °, the learning proceeds to minimize the error.

Third, greatly reduce the likelihood of falling into the local minima. This characteristic is because when the weight vector is updated for every pattern in the model, random vibrations different from the overall traveling direction are generated by the pattern relatively far from the center position.

Several studies of pattern recognition have reported that online EBP learning is faster than the offline methods described above.

In the existing parametric speaker identification system, a method of selecting a limited number of background speakers similar to registered speakers when registering a speaker and then calculating the client's identity score by using the speaker information of this group when identifying the requesting speaker's identity is introduced. .

Higgins et al. Proposed a method of leveling the similarity score of the sponsor to the score of the background speaker most similar to the client in order to obtain a reliable identity score.

here, Is the requester's voice sequence, The identity asserted by the client, Indicates the background speaker closest to the requesting speaker.

However, this score is 1) because all the speakers are searched to find the closest background speaker, so the computation is proportional to the size of the background speaker, and 2) depending on the location of the sponsor in the group of background speakers (ie The value of the denominator term in Eq. (1) varies greatly, depending on the background speaker distribution density of.

In order to solve this problem, a speaker group method is proposed in which a certain number of adjacent background speakers are selected when registering the speaker in the system, and the sum of the similarity scores is used as the leveling term when calculating the client speaker score.

here, Means a speaker group close to the registered speaker. The larger the number of background speakers, the higher the reliability of the proof score.

In terms of processing speed, the speaker grouping method can be regarded as the distribution of the proximity background speaker searching process to the speaker registration stage in order to shorten the processing time required for identification of Equation (1). Among the registration and certification stages of the speaker authentication system, the registration stage is relatively less stringent for real time than the certification stage.

The MLP trained by the EBP algorithm in pattern recognition takes three stages of learning: 1) learn the central position of the model, 2) learn the model distribution, and 3) learn the contours of the model distribution. As we move toward the latter, not only the pattern area contributing to learning is limited to the outside, but the effective learning rate varies for each pattern, and the inefficiency in which patterns that do not contribute to learning continue to be included in the calculation process. If the existing online EBP algorithm can accommodate this change, it will be possible to learn more quickly, which can lead to faster registration speed of MLP-based speaker identification system.

In addition to increasing the learning speed of the MLP, reducing the amount of data required for learning the MLP can improve the registration speed of the speaker identification system. To this end, the existing speaker group method is introduced at the registration stage to reduce the amount of data generated during registration by selecting a few background speakers similar to the registered speakers.

In the present invention, a concrete method of realizing these two ideas is presented and the method is applied to a system in which the continuous sound is the speaker's recognition unit. Since the configuration of the invention for this purpose is as follows. First we cover the background of the two ideas, and then we explain the specific implementation of each idea. Next, the speaker authentication system implemented in the present invention will be described, and then the effects of the two methods will be verified through experiments using a voice database.

In the EBP algorithm, the displacement that gives the fastest approach to the target value for the current value of the weight is calculated as follows.

here, The pattern Error measurement function of the output neuron layer for Is With the fourth neuron Weight of the connection between the first neuron, Is Motion of the first neuron, Is The sum of the weighted inputs for the first neuron. Applying the amount of change calculated in this equation to the previous weight vector, we can derive a value closer to the target value and realize it by the following method.

here, Represents the specific state time of the weight vector, Visual Is the error value for the target value of the learning pattern (s) Is the learning rate that determines the percentage of change to apply.

In the EBP learning for pattern recognition, each model is trained in an iterative manner by Equation (4), so it goes through 1) central location learning, 2) domain dispersion learning, and 3) domain contour learning.

This process in the two-model classification problem is shown in Figures 1A-1D. 1A to 1D illustrate an EBP learning step according to the present invention, in which FIG. 1A shows two model data distribution, FIG. 1B shows center position learning, FIG. 1C shows distributed learning, and FIG. 1D shows outline learning.

In FIGS. 1B to 1D of the drawings, the black band represents the crystal boundary of the two models, and the darker the color, the clearer the boundary. Here, it can be seen that the boundary line crosses the center of the two models in FIG. 1B, the degree of area dispersion of the two models is distinguished in FIG. 1C, and the detailed outline is shown in FIG. 1D.

On the other hand, of formula (3) Is expressed as

here, Is the error value of the individual output neurons for the current pattern, Is the number of output neurons. Is again expressed as

here, Is your learning goal, Is the current output value.

In each step of Figs. 1A to 1B, the error value represented by each pattern by Equation (5) can be regarded as the repulsive force moving the decision boundary, which can be regarded as the degree to which the individual pattern contributes to learning at each step. Figures 2a to 2d shows a state depicted in Figure 1a to 1b by the concept of repulsive force for each pattern.

In Figs. 2A to 2D, the dotted line represents the crystal boundary line, and the arrow represents the repulsive force for each pattern. In the non-learning state of FIG. 2A, all patterns exhibit strong repulsive force so that the center position can be learned. Since FIG. 2B, patterns located in the other model region near the crystal boundary line and beyond the boundary line generate strong repulsive force. Change shape and position.

The reason why the online method achieves faster learning than the offline method in pattern recognition can be found in the application method of repulsive force by pattern. In the off-line method, the sum vector of all repulsion vectors for each pattern is applied, but in the on-line method, these repulsive force vectors are applied individually. Thus, the on-line approach has more opportunities to change the shape of the crystal boundary, and the effect is greater when the model area is complex.

Although the online method has an advantage in learning speed compared to the offline method, there is still room for achieving an optimal speed.

In the online EBP, more active use of the repulsive distribution of patterns according to the learning stages can further reduce the learning time.

In Figures 2a to 2d it can be seen that the crystal boundary of the model is deformed by the large repulsive force of the pattern that has not yet been learned. The effect of this repulsive force on the decision boundary is the learning rate in equation (4). Is adjusted by. That is, for less learned patterns The larger is, the more rapidly the neighboring crystal boundaries of the pattern move. But as learning progresses, when the decision boundary reaches around the learning objectives, Must decrease gradually to increase the likelihood that the decision boundary will converge on the learning objectives. In addition, overlapping regions between different models are not fully trained, so the pattern of these regions can be changed to avoid unnecessary vibration of the crystal boundary. Should be small.

For this purpose, as in the traditional online EBP algorithm, Instead of fixing it during the learning period, Should be adaptable to the situation. Variable like this Has already been tried in the offline EBP, but applying it in the online EBP can further enhance the benefits of using online data redundancy.

On the other hand, in the existing online EBP algorithm, even if the pattern has already been learned, the pattern continues to participate in the learning calculation. If a standard is set for a certain learning outcome and the pattern satisfies the criterion, the learning time can be shortened with little effect on the overall learning progress even if the pattern is excluded from the learning calculation.

If we assume that the number of background speakers in Eq. (2) is infinite, we can see Eq. (2) as post-probability as the speaker's pre-probability is fixed in the post-probability for the claimant's claimant according to Bayes' law.

Gish proved that if the MLP had sufficient learning capacity, it could learn the posterior probability of each model from pattern recognition that classifies multiple models. Accordingly, if two registered learners are provided as registered speakers and a sufficient number of background speakers, the learning of the MLP approximates the speaker group method represented by Equation (7).

However, if the number of background speakers constituting the speaker group is above a certain level, the computational amount increases linearly with little contribution to the recognition rate improvement. Therefore, it is advantageous to select the least background speakers at the satisfactory recognition rate through experiments.

In order to realize the idea of improving the learning speed of the on-line EBP algorithm, the present invention proposes two methods.

The first method is to vary the learning rate for each learning pattern.

The learning rate needs to change from large value to small value as the model progresses. In equation (5), the output neuron corresponding to the model of the current pattern Provides a numerical measurement means to know the learning status of the current pattern. That is, when the pattern is not sufficiently learned, a large value is displayed, and when the pattern is sufficiently learned, a small value is represented. Therefore, a large value is obtained at the beginning of learning, and a small value is obtained as the learning is completed.

The learning rate that is commonly selected in the online learning method is 1 to 0.0001. If the upper limit is exceeded in this range, the value of the internal variable tends to diverge, and if it exceeds the lower limit, the learning becomes unnecessarily long. However, the range of appropriate values varies depending on the problem, so this range should be determined by experiment. Of equation (5) If you use as a learning rate, you don't have to worry about the lower limit, so you only need to find the appropriate upper limit. Then from zero up to the upper bound Limit the scope of

here, Is the upper limit of the learning rate found through the experiment, Is Is the output of Is a limited value.

However, equation (8) uses only error information limited to the current pattern. If the learning model itself contains a lot of inevitable errors, that is, if the model distribution region overlaps largely among several models, Is largely irrelevant to the current learning situation and can result in disturbing the overall learning. To deal with this case, limit the value of Eq. (8) to the mean error of the epoch one step earlier.

here, Denotes the mean error square energy defined as Equation (10) below.

here, Is the number of patterns used during one epoch.

The second method is to omit the learning of the pattern.

The main calculations made in the online EBP are error calculation of patterns, error back propagation and weight update. Since the contribution of the current pattern in the current learning phase can be found through, if the value of is smaller than the final target error value of Eq. (10) set before learning, it is determined that the current pattern contributes less to the learning. The variable update calculation process can be omitted. If later learning of other patterns contributes to the higher contribution of the current pattern, this condition is found by, so that you can re-engage.

In the present invention, these two methods are referred to as a changing rate and omitting patterns in instant learning method.

As suggested above, to reduce the number of background speakers required for MLP learning, select a limited number of background speakers adjacent to the registered speaker as shown in Eq. (2), but select the minimum number at a level that does not reduce the recognition rate of the speaker identification system. do.

This method requires an MLP that evaluates how close the registered speaker is to each background speaker, and an MLP that learns the characteristics of the registered speaker using the background speaker selected through the MLP. In the present invention, the former is referred to as MLP-I and the latter as MLP-II.

The MLP-I has the same number of output neurons as the number of background speakers, and is learned to classify each using the background speaker's data in advance. In this case, if there are many hidden neurons, it may take a long time just to evaluate the proximity of the background speakers. Therefore, the number of hidden neurons is selected so that each background speaker data can be classified with more than 80% accuracy.

The MLP-II learns the registered speaker using the data of the background speaker selected through the MLP-I. Unlike the speaker grouping method, however, the learning model is set to two registered speakers and a background speaker, so the patterns of the background speakers selected by the MLP-I are integrated into one model.

The number of background speakers to be selected in MLP-I is determined experimentally. This number should be above the level that the selected background speakers can cover the distribution range of the registered speakers, and the background speakers should densely surround the registered speakers to prevent the recognition rate from falling.

The learning pattern of the registered speaker according to the density of the background speaker is illustrated in FIGS. 3A and 3B. In this figure, the dotted lines represent the boundaries that determine the area of the registered speaker and the background speaker. 4 illustrates the operation principle of MLP-I and MLP-II.

In the present invention, this method is called a DCS (discriminative cohort speakers) method.

The speaker identification system implemented in the present invention extracts the isolated word from the input voice, and the Korean continuous sound (/ a /, / e /, / o /, / u /, / i /, / l /, nasal sound) from the isolated word. ), Then learn the speaker using MLP-I and MLP-II for each continuous sound and calculate the proof score. The processing performed in this system is shown in FIG. 5, and description of each processing is as follows.

Step 1) Voice Analysis and Feature Extraction

The input speech of a 16bit 16kHz sampled speaker is divided into 30ms long frames that overlap 20ms.

A 16th-order Mel interval filter bank is extracted for each frame and used to detect isolated words and continuous sounds. The filterbank coefficients are then subtracted from all coefficients by averaging the coefficients up to 1 kHz to remove the effect of the mass on the entire spectral envelope. Then, for effective learning of MLP, again adjust the average of all coefficients to zero.

For each frame, the 50th-order 0-3kHz band equally spaced Mel filter bank is extracted and used for speaker identification. This voice feature is based on research showing that more speaker information is concentrated in the secondary formant. The filterbank coefficients are then subtracted from all coefficients by averaging the coefficients up to 1 kHz to remove the effect of the mass on the entire spectral envelope. Then, for effective learning of MLP, again adjust the average of all coefficients to zero.

Step 2) Isolated word detection and continuous sound detection

The MLP trained to detect each continuous sound and silence in speaker-independent manner is used to detect isolated words and sustained sounds within the isolated words.

Step 3) Learn continuous speaker registration

Step 3-1) In case of basic online EBP or COIL application, learn MLP using registered speaker and background speaker data.

Step 3-2) In case of applying MLP-I and MLP-II, each continuous sound of all isolated words is input to MLP-I for each continuous sound, and then the values of output neurons are averaged, and the background of n persons in the order of the average value is high. Select the speaker. The registered speaker is trained in MLP-II using the n background speaker data selected for each continuous sound. The number of patterns used for MLP-II learning per continuous sound is 10 per background speaker.

Step 4) Evaluate Speaker Score by Continuous Sound

Step 4-1) In the case of basic online EBP or COIL application, each continuous sound of all isolated words is input to the MLP for each continuous sound, and the output neurons are averaged.

Step 4-2) In case of applying MLP-I and MLP-II, each continuous sound of all isolated words is input to MLP-I for each continuous sound, and then the values of output neurons are averaged. Select the speaker. The output of the MLP-II is averaged over all sustained sounds containing at least one background speaker in three levels among the selected background speakers. If there is no continuous sound with one or more background speakers in level 3, the requesting speaker is rejected.

Step 5) Comparison of Registered Word and Speaker Score Threshold

The final rejection / acceptance is determined by comparing the average value in step 4 with a preset threshold.

In this speaker identification system, since the continuous sound is used as speaker recognition unit, the lower probability distribution takes the form of univariate. Therefore, MLP, MLP-I, and MLP-II related to speaker learning are sufficient as a two-layer structure including one hidden layer. Since MLP and MLP-II have only two models to learn, one output neuron and two hidden neurons can fully learn them.

Example

The data to be used for this experiment was a four-digit vocalization of 40 Korean men and women. Where the four-digit number is / goN /, / il /, / i /, / sam /, / sa /, / o /, / yug /, / cil /, / pal / corresponding to the Arabic numerals 0-9. , / gu / means four consecutive digits. Each speaker uttered a total of 35 different numeric arrays four times, and the utterances were recorded in 16-bit size at 16 kHz. Three out of four voices are used as the registered voice of each speaker and the remaining one is used as the proof test voice.

As the background speaker required for learning the registered speaker, 29 men and women besides the above 40 are used.

Various parameters of MLP (MLP-II in case of DCS method) used for speaker registration in the experiment are as follows.

The input data is leveled from -1.0 to +1.0, and the target value of each output neuron is +0.9 for the registered speaker and -0.9 for the background speaker to improve the learning speed. All internal variables are initialized to random values from -0.5 to +0.5 before learning. During training, the voice patterns of the two models are alternately presented in the MLP. In most cases, the number of learning patterns of the two models does not match, so the smaller one is repeatedly presented until all the patterns are presented. Fill it. The maximum learning epoch is limited to 1000 in consideration of falling into the local minima, and the learning goal is the average squared error energy rate in order to prevent premature learning interruption while the calculated value of equation (10) is less than 0.01. It shall be set to 0.01 or less.

40 experimental speakers are used one by one as registered speakers and actual speakers, and 39 others are used as impersonal speakers. As a result, 35 actual speaker attempts and 1,560 impersonal speaker attempts are evaluated per speaker, and as a whole experiment, 1,400 actual speaker attempts and 54,600 impersonal speaker attempts are evaluated.

The experiment was conducted on AMD 1.4GHz class computer. In the test results, the error rate means equal error rate, the number of learning patterns represents the total number of patterns learned to register one speaker, and the learning time is Indicates the actual time taken to learn the patterns. The error rate, the number of learning patterns, and the learning time are averages for all proof attempts.

First, the experimental results of finding the optimal learning rate by changing the learning rate using only the basic online EBP algorithm are shown in FIG. 6.

In this result, the highest learning rate was achieved at the learning rate of 0.5, but it is reasonable to select a learning rate of 1.0 considering the lowest error rate.

Next, the experimental results of applying COIL instead of online EBP at this learning rate are shown in FIG. 7. Here, the experiment using online EBP is denoted as OnEBP, the experiment using only variable learning rate as CIL, and the experiment using both variable learning rate and pattern omission as COIL.

In this result, the upper limit of the learning rate of CIL and COIL is 1.0, and this figure is the optimal value considering the learning time and error rate together with the online EBP. In each case, learning time improved by 95.1% and 229.2%.

8 shows experimental results when DCS is applied while using the online EBP. In this result, it is necessary to confirm that the number of patterns actually put into learning decreases, so instead, record the number of patterns used for learning during the total learning period instead of the epoch.

As a result, the minimum number of speakers in the cluster was 26 and the number of speakers was 14 when the number increased by 0.19%. In each case, the learning time improved 4.6% and 54.9% compared to the online EBP. In addition, when the number of speakers in the cluster is equal to the number of background speakers (that is, when there is no learning data reduction effect), a learning time delay of about 5% was observed, compared to the case where the speaker group method was not applied. This is because of the increase in the amount of calculation.

Finally, the experimental results using both COIL and DCS are shown in FIG. 9.

As a result, the error rate increased by 0.05%, the minimum number of speakers in the group was 20, and when the number increased by 0.19%, the number of speakers was 17. In each case, the learning time was improved by 305.1% and 351.4% compared to the online EBP.

The improvement rate of each experimental result when the error rate is 1.6% is summarized in FIG.

As can be seen from this result, when the COIL method and the DCS method are applied at the same time, they exhibit synergistic effects and record higher learning time improvement rates than when applied individually.

Although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the above-described embodiments, and the present invention is not limited to the above-described embodiments without departing from the spirit of the present invention as claimed in the claims. Various modifications can be made by those skilled in the art, and such modifications are intended to fall within the scope of the appended claims.

As described above, according to the method of improving the registration speed of the MLP-based speaker authentication system by improving the learning speed and reducing the learning data according to the present invention, the MLP, which provides some advantages over other pattern recognition methods, is a speaker of the speaker identification system. It is also promising as a learning and recognition method. However, the learning of MLP takes considerable time because of the low speed of the EBP algorithm used for learning. This, together with the fact that many background speakers are required to achieve a high speaker recognition rate in speaker identification systems, creates a problem that it takes a lot of time to register a speaker in the system. This does not meet the real-time requirements of the speaker identification system, which must provide the proof service immediately after the speaker is registered.

In order to solve this problem, the present invention shortens the time required for speaker registration in the MLP-based speaker authentication system by using the COIL method to improve the learning speed of the EBP and the DCS method which adopts the speaker grouping method in the existing speaker authentication. . As a result of applying these two methods in MLP-based speaker authentication system with continuous sound recognition unit, we confirmed the improvement of registration time, and when applied simultaneously, it is about 4 times faster than using the existing online EBP learning algorithm. Registration speed was recorded.

On the other hand, the experimental results showed that the DCS speedup was smaller than that of the COIL method. This suggests that the number of background speakers used in the speaker identification system was small, indicating that the DCS reduction effect was small. If more background speakers are applied in the future, the recognition rate will be improved and the greater effect of DCS will be expected.

Claims (3)

The input speech of 16bit 16kHz sampled is divided into 30ms long frames overlapping 20ms, and 16th Mel interval filter banks are extracted for each frame and used to detect isolated words and continuous sounds. Averages the coefficients up to 1 kHz to remove the effect of the mass on the full spectral envelope, subtracts this value from all coefficients, and again adjusts the average of all coefficients to zero for effective learning of the MLP. , Mel filter bank of 50th order 0 ~ 3kHz band equally spaced interval is extracted for each frame and used for speaker identification. Filter bank coefficient is obtained by averaging coefficients up to 1kHz to remove the effect of the quantity on the entire spectrum envelope. A speech analysis and feature extraction step of subtracting from all coefficients and adjusting the average of all coefficients to zero again for effective learning of MLP; An isolated word detection and sustained sound detection step of detecting an isolated word and a sustained sound in the isolated word using an MLP trained to detect each continuous sound and silence in a speaker-independent manner; In the case of basic online EBP or COIL application, the MLP is trained using registered speaker and background speaker data, In the case of MLP-I and MLP-II application, each continuous sound of all isolated words is input to MLP-I for each continuous sound, and then the average number of output neurons is averaged. Learning the registered speakers in the MLP-II using n background speaker data selected for each continuous sound, wherein the number of patterns used for the MLP-II learning for each continuous sound is 10 per background speaker; In the case of basic online EBP or COIL application, each continuous sound of all isolated words is input to the MLP for each continuous sound, and then the average value of output neurons is averaged. In the case of MLP-I and MLP-II application, each continuous sound of all isolated words is input to MLP-I for each continuous sound, and then the average number of output neurons is averaged, and n background speakers are selected in the order of higher average. Among the selected background speakers, the output of the MLP-II is averaged for all sustained sounds containing at least one background speaker in the continuous speech register learner stage, and there are no sustained sounds containing at least one background speaker in the continuous speech register learner stage. A case of sustained speech score evaluation that rejects the requesting speaker; Compared with the average value of the speaker score evaluation step for each continuous sound and the preset threshold value, the registration word and speaker score threshold value comparison step for determining final rejection / acceptance are made. How to improve registration speed of MLP based speaker identification system. The method of claim 1, MLP, MLP-I, and MLP-II related to speaker learning are all composed of a two-layer structure including one concealment layer. To improve registration speed. The method of claim 1, The MLP-II is a method of improving the registration speed of the MLP-based speaker authentication system by improving the learning speed, characterized in that consisting of one output neuron and two hidden neurons.