CN102543075A - Speaker VQ-SVM Parallel Recognition System Based on Virtual Instrument Technology - Google Patents

Abstract

The present invention relates to a speaker VQ-SVM parallel recognition system based on virtual instrument technology. This speaker VQ-SVM parallel recognition system based on virtual instrument technology includes a speech preprocessing unit, a feature extraction unit, a speaker model unit, a recognition Unit, LabVIEW virtual instrument platform, on the virtual instrument platform, a large program is divided into small modules through LabVIEW subVI, and the program parts that call MATLAB nodes involved in the program are written into subVIs, and by calling these Sub VI to realize the construction of the system. The present invention overcomes the need for serial recognition to waste time when the existing VQ-SVM two methods are mixed for speaker recognition, and proposes to concentrate the VQ and SVM two methods on the same platform to realize parallel recognition processing, thereby improving the overall system The recognition time is saved under the premise of the recognition effect.

Description

Speaker VQ-SVM Parallel Recognition System Based on Virtual Instrument Technology

the

1. Technical field:

The invention relates to the field of signal processing and pattern recognition, in particular to a speaker VQ-SVM parallel recognition system based on virtual instrument technology.

2. Background technology:

Speaker recognition is to identify the identity of the speaker by analyzing the speaker's speech characteristics. Speaker recognition methods mainly include vector quantization methods, probability statistics methods, and discriminant classifier methods. According to the composition principle of the speaker recognition system, speaker recognition mainly includes two stages of training and recognition, as shown in Figure 1. Firstly, the original speech signal is obtained, and then a clean speech signal is obtained through preprocessing, and then the speech feature parameters are extracted, and then speaker training and recognition are realized through a specific method. The speaker model usually uses a database to store a large number of speech feature samples processed by a specific algorithm, and the speech to be recognized is matched with the sample set in the database after preprocessing and feature extraction to achieve discrimination.

Any single method has both advantages and limitations. At present, most of the research is a hybrid recognition method that combines two or more methods. VQ technology is a data compression and coding technology; SVM is a machine learning method based on statistical theory. These two methods are complementary. The advantages of the vector quantization (VQ) method are that the classification characteristics of large samples are better, the number of models is small, the training time is short, and the recognition response is faster. The disadvantage is that it cannot solve nonlinear problems and has poor anti-noise performance. The advantage of the support vector machine (SVM) method is that small sample classification is better, and it shows unique advantages in solving nonlinear and high-dimensional pattern recognition problems. The disadvantage is that the training algorithm is complex and the training speed is slow, and it is difficult to handle large sample data. Although there have been two methods of VQ-SVM mixed for speaker recognition, various operations are usually implemented on the MATLAB platform. Therefore, if a variety of different methods are used, it can only be carried out in a serial manner. Similarly, the existing VQ-SVM two methods are mixed for speaker recognition, which is also the so-called serial recognition in which one method is used for initial recognition and then another method is used for secondary recognition. It is not difficult to find that the biggest weakness of this serial identification method is that it takes up machine resources and wastes identification time. the

3. Contents of the invention:

The object of the present invention is to provide a speaker VQ-SVM parallel recognition system based on virtual instrument technology, which is used to solve the problem that the existing VQ-SVM two methods are mixed for speaker recognition, which occupies machine resources and wastes recognition time.

The technical solution adopted by the present invention to solve its technical problems is: this speaker VQ-SVM parallel recognition system based on virtual instrument technology includes a speech preprocessing unit, a feature extraction unit, a speaker model unit, a recognition unit, and a LabVIEW virtual instrument platform On the virtual instrument platform, a large program is divided into small modules through LabVIEW subVIs, and the program parts that call MATLAB nodes involved in the program are written into subVIs, and the system is realized by calling these subVIs. Construct;

Using the VQ algorithm, the VQ model is established. The initial codebook adopts the split method, and the centroid of the feature vector is selected as the initial codebook. In LabVIEW, the establishment and storage of the speaker model is realized by calling the MATLAB node. The algorithm formula is as follows:

              Total Distortion:

                                 Compute the new codeword:

—集合中矢量的个数，

—中所有矢量的质心； In the formula:

— the number of vectors in the set,

— The centroid of all vectors in ;

Relative distortion improvement:

  ，

 ； The SVM algorithm is used to establish the SVM model, and the radial basis kernel function is used to establish the speaker model. The algorithm formula is as follows:

,

;

In the recognition unit, output the recognition result through the speaker recognition front panel in the result judgment part. When the results of the two recognition methods of VQ and SVM are inconsistent, as long as there is one method that can recognize, the recognition result of this method will be output as the correct result; When the results of the two methods are the same, the recognition result is output on the front panel of speaker recognition, and the correct recognition is indicated by a green light, and the non-recognition is indicated by a red light.

In the above scheme, the feature extraction unit uses the Mel frequency cepstral coefficient MFCC and its first-order difference as the feature parameters for recognition, and realizes the feature parameter extraction by programming in the MATLAB7.0 environment. The specific parameters are set as: frame length 512, frame shift 256, the number of filters is 12, the sampling frequency is 44100Hz, and the first and last two frames are removed, because the first-order difference between the two frames is zero, so a 24-dimensional speech feature vector is obtained. the

Beneficial effect:

1. The present invention overcomes the need for serial recognition to waste time when the existing VQ-SVM two methods are mixed for speaker recognition, and proposes to concentrate the VQ and SVM two methods on the same platform to realize parallel recognition processing, thereby increasing the The recognition time is saved under the premise of the recognition effect of the whole system.

2. The present invention combines two recognition methods on a virtual instrument technology platform to perform speaker parallel recognition. In the case of small samples, the SVM method is better than the VQ method; as the number of samples increases, the recognition performance of the SVM shows a downward trend, while the recognition performance of the VQ method has an upward trend, which makes full use of the two methods in the number of samples. The complementarity they have can improve the overall performance of the system. the

4. Description of the drawings:

Figure 1 is a schematic diagram of the speaker recognition system;

Fig. 2 is a structural representation of the present invention;

Fig. 3 is the schematic diagram of speaker identification front panel in the present invention;

Figure 4 is a flowchart of the LBG algorithm.

1 speaker recognition front panel 2 lights. the

5. Specific implementation methods:

Below in conjunction with accompanying drawing, the present invention will be further described:

As shown in Figure 2 and Figure 3, this speaker VQ-SVM parallel recognition system based on virtual instrument technology includes a speech preprocessing unit, a feature extraction unit, a speaker model unit, a recognition unit, and a LabVIEW virtual instrument platform. On the platform, a large program is divided into small modules through LabVIEW subVI, and the program parts that call MATLAB nodes involved in the program are written into subVIs, and the system construction is realized by calling these subVIs. In order to realize the parallel recognition of VQ and SVM, in view of the characteristics that LabVIEW can realize multi-task and multi-thread, based on virtual instrument technology and fusion of speaker recognition technology, LabVIEW manages and calls MATLAB to realize the system processing. When the result judging part of the present invention is selected, when the results of the two recognition methods are inconsistent, as long as there is one method that can recognize, the recognition result of this method will be output as the correct result. When the results of the two methods are the same, the speaker's front panel 1 outputs the recognition result, the correct recognition is indicated by the green light 2, and the red light 2 is indicated for no recognition. The performance comparison of the system is shown in Table 4.

The key issues in the speaker recognition system are the extraction of speech feature parameters and the establishment of speaker models. Mel frequency cepstral coefficients (MFCC) and their first-order differences are selected as the feature parameters for recognition. MFCC parameters only reflect the static characteristics of speech parameters, while the human ear is more sensitive to the dynamic characteristics of speech, and the parameters that reflect the dynamic changes of speech are differential cepstrum. The extraction of feature parameters is realized by programming in the MATLAB7.0 environment. The specific parameters are set as follows: frame length 512, frame shift 256, number of filters 12, sampling frequency 44100Hz, and remove the first and last two frames, because the two The first-order difference of the frame is zero, so that the 24-dimensional speech feature vector is obtained. the

=0.01，最大迭代次数

=log，得到了较好的识别效果。这些工作是在MATLAB7.0环境下通过编程实现的，然后在LabVIEW中通过调用MATLAB节点来实现模型的建立及存储。其中涉及到核心算法的公式如下：  One of the algorithms used in the present invention is VQ. There are many ways to establish the VQ model (codebook). The most basic and commonly used method is the LBG algorithm. The generation of the book, the algorithm process is shown in Figure 4. The initial codebook adopts the splitting method, and the centroid (centroid) of the feature vector is selected as the initial codebook. Select codebook capacity through experiments =16, distortion threshold

=0.01, the maximum number of iterations

= log , a better recognition effect was obtained. These tasks are realized through programming in the MATLAB7.0 environment, and then the establishment and storage of the model is realized by calling the MATLAB node in LabVIEW. The formulas involved in the core algorithm are as follows:

                        Total Distortion:

                                 Compute the new codeword:

—集合中矢量的个数，—

中所有矢量的质心) (in:

— the number of vectors in the set, —

centroid of all vectors in

                         Relative distortion improvement:

Another algorithm adopted by the present invention is SVM, the key of which is the selection of kernel function and the optimization of its parameters.

  ，

              Radial basis kernel function:

,

In order to select a suitable kernel function, 10 boys and 10 girls were randomly selected as samples in practice, and a comparative experiment was carried out with RBF and Poly kernel functions, see Table 1 and Table 2 for details. It can be seen that under the same conditions, the experimental effect of using RBF is better. In view of the basis of previous experiments and the parameters of the RBF kernel function are less than those of the Poly kernel function, the radial basis kernel function is comprehensively selected to establish the speaker model. Kernel function parameters are an important factor affecting the performance of support vector machine classifiers, so the optimization of kernel function parameters is very important. At present, the commonly used method is to let C and γ take values within a certain range. For the given parameters, the training set As the original data set, use the cross-validation method to obtain the verification classification accuracy rate, and finally take the set of parameters with the highest verification classification accuracy rate in the training set as the optimal parameter. This method is also the idea of the grid search method. The optimization of parameters C and γ can be achieved by using the grid.py program under the subdirectory of python under the libsvm toolbox. The screenshot of parameter optimization of 20 people and 5 frames of data is shown in Figure 4.

the

Table 1 Experimental results of 10 girls number of frames ( C , γ ) (Degree, Coeff) Recognition rate%(RBF) Recognition rate% (Poly) RBF identification time (s) Poly recognition time (s) 1 (0.25,0.25) (3 , 1) 100 100 0.01 0.02 3 (0.25,0.25) (3 , 1) 96.67 96.67 0.02 0.01 5 (0.25,0.25) (3 , 1) 100 98.00 0.02 0.02 7 (0.25,0.25) (3 , 1) 100 98.57 0.02 0.01 10 (0.25,0.25) (3 , 1) 97.00 95.00 0.02 0.02 15 (4.00 ,1.00) (3 , 1) 96.00 96.00 0.03 0.02 20 (4.00 ,1.00) (3 , 1) 96.00 97.00 0.03 0.03 30 (4.00 ,1.00) (3 , 1) 93.00 93.33 0.08 0.08

Table 2 Experimental results of 10 boys number of frames ( C , γ ) (Degree, Coeff) Recognition rate%(RBF) Recognition rate%(Poly) RBF identification time (s) Poly recognition time (s) 1 (1.32,0.76) (3 , 1) 90.00 90.00 0.01 0.01 3 (1.00 ,1.00) (3 , 1) 93.33 93.33 0.02 0.02 5 (4.00 ,1.00) (3 , 1) 96.00 94.00 0.02 0.01 7 (4.00 ,1.00) (3 , 1) 95.71 90.00 0.01 0.02 10 (4.00,1.00) (3 , 1) 92.00 84.00 0.02 0.02 15 (4.00 ,1.00) (3 , 1) 86.67 84.00 0.05 0.03 20 (4.00 ,1.00) (3 , 1) 87.00 85.00 0.05 0.05 30 (4.00 ,1.00) (3 , 1) 84.67 85.33 0.11 0.09

Through the above analysis and experiments, the present invention first constructs a speaker recognition system of a single method on a virtual instrument platform, and debugs it to achieve a better recognition effect, and then combines these two methods in parallel to realize the virtual instrument technology-based The construction of the speaker recognition system. In the process of building the system, the idea of modularization is fully utilized, and a large program is divided into small modules for realization, which not only simplifies the program but also increases the readability of the program. That is to say, it is realized by LabVIEW subVI on the virtual instrument platform, and the program parts that call MATLAB nodes involved in the program are all written into subVIs, and the system construction is realized by calling these subVIs. The front panel of the system is shown in Figure 3.

Experimental verification and result analysis

The speaker's voice is recorded in a common laboratory environment through the tape recorder that comes with the computer. The corpus is a 16-bit monophonic voice signal with a sampling frequency of 22050 Hz, and the file type is wav format. In this paper, 30 people are selected as speakers, and each person records 20 segments of speech (each segment is 2-4s) as a sample library, all of which have nothing to do with the samples. The first 10 segments are used to establish the speaker model, and the last 10 segments are used for testing. After testing 5 times and 10 times of voice randomly selected by 20 and 30 people, the experimental results are shown in Table 3. From the analysis, as the sample size increases, the VQ method is better than the SVM method, which means that the SVM method has classification advantages in small samples, and the VQ method has classification advantages in large samples. It can be inferred that as the sample size increases or even Thousands of them, the VQ method will outperform the SVM method. In the result judging part of the present invention, when the results of the two recognition methods are inconsistent, as long as there is one method that can recognize, the recognition result of this method will be output as the correct result. When the results of the two methods are the same, the speaker's front panel 1 outputs the recognition result, the correct recognition is indicated by the green light 2, and the red light 2 is indicated for no recognition. The performance comparison of the system is shown in Table 4. It can be seen that the performance of the system has been improved by using the two methods in parallel to realize the speaker recognition, although the recognition time has increased, but the increase is very little.

the

Table 3 Comparison of experimental results

Table 4 System performance comparison recognition methods Recognition rate(%) Misrecognition rate (%) Recognition time (s) Q 94.88 10.09 0.06 SVM 97.38 9.71 0.08 VQ-SVM 98.54 5.28 0.15

in conclusion

The present invention combines two recognition methods on a virtual instrument technology platform to perform parallel speaker recognition. In the case of small samples, the SVM method is better than the VQ method; as the number of samples increases, the recognition performance of the SVM shows a downward trend, while the recognition performance of the VQ method has an upward trend, which makes full use of the two methods in the number of samples. The complementarity they have can improve the overall performance of the system.

Claims (2)

1. one kind based on the parallel recognition system of the speaker VQ-SVM of virtual instrument technique; It is characterized in that: the parallel recognition system of this speaker VQ-SVM based on virtual instrument technique comprises voice pretreatment unit, feature extraction unit, speaker model unit, recognition unit, LabVIEW virtual instrument platform; On virtual instrument platform, realize a large program is divided into each little module through the sub-VI of LabVIEW; The program part that calls the MATLAB node that relates in the program all is written as each sub-VI, through calling the structure that this a little VI realizes system;

Adopt the VQ algorithm, set up the VQ model, initial codebook adopts disintegrating method, and the centre of form of selected characteristic vector realizes foundation and the storage of speaker model through calling the MATLAB node as inceptive code book in LabVIEW, and algorithmic formula is following:

Total distortion:

Calculate new code word:

In the formula: the number of vector in

-set, the barycenter of all vectors among -

;

Relative distortion improvement amount:

Adopt the SVM algorithm; Set up the SVM model; Select for use radially basic kernel function to set up speaker's model; Its algorithmic formula is following:

,

;

Export recognition result in decision section as a result on through the Speaker Identification front panel in the recognition unit, when the result of VQ, two kinds of recognition methodss of SVM is inconsistent, export the recognition result of this method as correct result as long as there is a kind of method just to discern; When the coming to the same thing of two kinds of methods, on the Speaker Identification front panel, export recognition result, correct identification is with the green light indication, and nonrecognition is indicated with red light.

2. the parallel recognition system of the speaker VQ-SVM based on virtual instrument technique according to claim 1; It is characterized in that: described feature extraction unit adopts Mei Er frequency cepstral coefficient MFCC and first order difference thereof as the characteristic parameter of identification, realizes the extraction of characteristic parameter through under the MATLAB7.0 environment, programming, and concrete parameter is set to: frame length 512; Frame moves 256; The number of wave filter is 12, SF 44100Hz, and removed each two frame of head and the tail; Because the first order difference of this two frame is zero, so just obtained the speech feature vector of 24 dimensions.

Images