CN101740029B - Three-particle cooperative optimization method applied to vector quantization-based speaker recognition - Google Patents

Abstract

The present invention relates to a three-particle collaborative optimization method applied to speaker recognition, which is a vector quantization speaker codebook model optimization design method, including: the initial group is 6 particles divided into 2 subgroups, each particle represents a codebook, Each subgroup consists of three particles named three particles, and two three particles use different particle update parameters to realize global exploration and local search. In each iteration, the particle performs PSO velocity and position update and LBG algorithm operation with 3 iterations. Whenever the number of mixing updates is reached, the particles are mixed and divided into new three particles to achieve global information exchange and co-evolution. When the maximum number of iterations of the initial population is satisfied, two particles are selected from the two three particles to continue the search until the maximum number of iterations of the elite particles is reached, and the optimal one is used as the speaker codebook model. The invention better solves the problem that the initial codebook influences the optimization result, and obviously improves the speaker recognition performance of the short voice.

Description

A three-particle collaborative optimization method for speaker recognition based on vector quantization

technical field

The invention relates to the technical field of speech recognition, and more specifically, relates to a three-particle cooperative optimization method applied to speaker recognition based on vector quantization.

Background technique

Speaker recognition is a biometric technology that is urgently needed for the application of information technology in existing communication networks. Judicial field (such as monitoring and identification of criminals), security field (such as airport access control system control), information service field (such as automatic information retrieval or e-commerce), etc.

The speaker model is the core of high-performance speaker recognition. Since the late 1980s, speaker recognition has entered a new and thriving development due to the establishment of speaker models based on vector quantization, probability statistics models, and artificial neural networks. period. As a typical optimization problem, model training or optimization poses severe challenges to traditional optimization techniques. In the past, when researchers considered the optimization of the speaker model, they focused on the research and application of deterministic optimization methods, such as the LBG algorithm and the maximum expectation algorithm. It has serious defects and it is difficult to obtain a more accurate global optimal solution. In order to make a breakthrough, it must rely on the research and application of new algorithms. Training data has always been an important factor affecting the performance of the speaker recognition system. To build a model for each target speaker, this must require as much training data as possible for the target speaker. In the actual application system, since most applications of speaker recognition are to complete the verification of the target speaker when the target speaker is unwilling to actively cooperate, a large amount of speech of the target speaker is often unavailable, which is very important for training and In the practical application background of recognizing short voices whose voice time is less than 10 seconds, the recognition performance of the system will seriously deteriorate. How to establish an optimal short voice speaker model is a difficult problem that has not yet been solved.

1. Speaker recognition method based on vector quantization

The practical development of speaker recognition system makes short speech speaker recognition a key issue. For less training speech, the probabilistic statistical model cannot obtain more accurate model estimation parameters, and the vector quantization (Vector Quantization, VQ) method can achieve better recognition results. Vector quantization is a non-parametric (template) model, which is based on the assumption that the feature vector of each speaker's speech data has a certain distribution, and this distribution is the information that distinguishes the speaker from others. The vector quantization method tries to describe this distribution, and builds a codebook model (Code-Book Model, CBM) based on the principle of distortion minimization based on the training data of each speaker. During recognition, a set of feature vectors is extracted from the speech to be recognized, vector quantization is performed on each codebook model, and the average quantization error is calculated. The speaker of the codebook corresponding to the minimum average quantization error is the result of recognition. This approach has small storage requirements and computational overhead. The block diagram of the speaker recognition system based on vector quantization is shown in Figure 1.

In the speaker recognition method based on vector quantization, codebook design is a key issue. The main goal of codebook design is to find an optimal classification of training vectors, that is, the best scheme to divide the speaker's T L-dimensional training vectors into M categories. A good codebook can maximize the effect of vector quantization and describe the distribution of the original training sample data more accurately. The codebook is generally a subspace of the limited space, which is obtained by clustering a large amount of training data according to prior knowledge and some distortion measure. Codebook design is also an iterative process, which can be regarded as a multi-peak and multi-extreme problem. Like function optimization, the search for the optimal codebook also requires an optimization algorithm strategy. Therefore, people continue to seek new codebook design optimization strategies from newly proposed optimization algorithms, such as the following particle swarm optimization algorithm, hybrid leapfrog algorithm, and LBG algorithm.

2. Particle swarm optimization algorithm and hybrid leapfrog algorithm

Particle Swarm Optimization (PSO) is a bionic algorithm derived from the social biological world in nature. It is based on the theory of artificial life and the flocking phenomenon of birds and fish. With its simple algorithm structure and excellent problem solving It has attracted many researchers and achieved remarkable results. The particle swarm optimization algorithm regards each individual in the group as a particle (point) without mass and volume in the multidimensional search space, and each particle is a feasible solution to the optimization problem, and an fitness value is determined for it by the objective function (Fitness Value). These particles fly at a certain speed in the solution space, and dynamically adjust their flight speed according to the flight experience of the particle itself and the flight experience of its companions, that is, each particle passes through its own optimal value p _i and the group The optimal value of _pg is used to continuously correct its own direction and speed, gradually move to a better area, and finally search and find the optimal solution to the problem.

In a D-dimensional target search space, P particles are randomly generated, the position of the i-th particle can be expressed as z _i =(z _i1 , z _i2 ,..., z _iD ), and the velocity is v _i =(v _i1 , v _i2 ,...,v _iD ), calculate the current fitness value of z _i according to the fitness function to measure the quality of the particle position. The optimal position searched by particle i so far is p _i =(p _i1 , p _i2 ,..., p _iD ), and the optimal position searched by the entire particle swarm so far is p _g =(p _g1 , p _g2 ,... , p _gD ). The particle of particle i in each iteration updates its velocity and position according to the following formula:

$v_{\mathrm{id}}^{k+1}=w{v}_{\mathrm{id}}^k+c_1{r}_1(p_{\mathrm{id}}-z_{\mathrm{id}}^k)+c_2{r}_2(p_{\mathrm{gd}}-z_{\mathrm{id}}^k)$ (Formula 1)

$z_{\mathrm{id}}^{k+1}=z_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$ (Formula 2)

Among them, d=1, 2,..., D; k is the number of iterations; r ₁ and r ₂ are random numbers uniformly distributed between [0, 1]; w is inertia weight; c ₁ , c ₂ are learning factors , also known as the acceleration factor, which enables particles to have the ability to self-summarize and learn from outstanding individuals in the group, so as to approach their own historical optimal point and the historical optimal point of the group. Particles keep tracking p _i and p _g in the target search space until the predetermined number of iterations is reached.

In the particle swarm optimization algorithm, all the particles in the group use the same parameter value in the whole search process. Many scholars have shown that the different values of inertia weight w, learning factor c ₁ and c ₂ have great influence on the performance of the algorithm big impact. In the search process, the balance between global search ability and local search ability plays a crucial role in the success of the algorithm. Shi and Eberhart studied the influence of the inertia weight w on the optimization performance, and found that a larger w makes the particle have a good global exploration ability, which is conducive to jumping out of the local minimum point, while a smaller w value is conducive to improving the local search ability, for This proposes a method that linearly decreases the inertia weight according to the number of iterations of the algorithm. The algorithm uses larger inertia weights in the early stages and smaller inertia weights in the later stages. They also proposed the method of using fuzzy control system to adaptively adjust the inertia weight. Although these improved algorithms have improved the convergence speed and achieved better performance on single-peak problems, they are not effective in solving multi-peak function problems. The learning factors c ₁ and c ₂ determine the impact of the particle's own experience information and other particles' experience information on the particle's trajectory, reflecting the information exchange between particles, and can effectively control the global exploration and local search. Setting a larger value of c ₁ and a smaller value of c ₂ is beneficial for particles to fly in the entire search space and avoid flying to the optimal particle of the group. On the contrary, smaller c ₁ and larger c ₂ values are beneficial for the population to converge to the optimal point, but it will prompt the particles to converge to the local optimal value prematurely. Ratnaweera suggested to linearly adjust c ₁ and c ₂ during the whole iterative process, when c ₁ =[2.5□0.5], c ₂ =[0.5□2.5] can improve the optimization effect of some multimodal functions.

Shuffled Frog-Leaping Algorithm (SFLA) simulates the foraging characteristics of frog populations. In a D-dimensional target search space, randomly generate P frogs (solutions to the problem) to form the initial population, and the i-th frog represents the solution to the problem as X _i =( _xi1 , _xi2 ,..., _xiD ). The individual frogs are arranged from good to bad according to their fitness value, and the whole group is divided into m subgroups. Among them, the frog ranked first is divided into the first subgroup, the frog ranked second is divided into the second subgroup, the frog ranked m is divided into the m subgroup, and the frog ranked m+1 is divided into the first subgroup. And so on, until all the frogs are divided. Each subgroup conducts a local deep search, that is, in each iteration, the current worst individual X _w in the subgroup follows the current best individual X _b in the subgroup or the best individual X _g in the entire group for update operations, If there is no improvement in the fitness value of X _w , a new X _w is randomly generated. When all subgroups have completed a certain number of iterations of local search, the mixed strategy mixes and sorts all individuals and re-divides subgroups, and then performs local deep search, and so on until the termination condition is met.

3. Performance Evaluation Criteria

令说话人训练语音的L维特征矢量集为X＝{x₁，x₂，…x_T}，x_i＝{x_i1，x_i2，…，x_iL}，T为训练语音样本集中特征矢量的数目，Y是由M个L维的码字组成的码本，表示该说话人的模型，即Y＝{y₁，y₂，…，y_M}。The speaker codebook model design quality is represented by a distortion measure. The distortion measure is usually represented by the mean square error (MSE) between the training vector and the corresponding nearest codeword, abbreviated as

Let the L-dimensional feature vector set of the speaker's training voice be X={x ₁ , x ₂ ,...x _T }, x _i ={ _xi1 , x _i2 ,..., x _iL }, T is the feature vector in the training voice sample set Y is a codebook composed of M L-dimensional codewords, representing the speaker's model, that is, Y={y ₁ , y ₂ , . . . , y _M }.

$\widetilde{\mathrm{D.}}=\frac{1}{T}\underset{i=1}{\overset{T}{\mathrm{\Σ}}}{\left[d_{\min }\left(x_i\right)\right]}^2$ (Formula 3)

d(x_i，y_i)为Euclidean距离。in,

d(x _i , y _i ) is the Euclidean distance.

$d(x_i,{\mathrm{the\; y}}_j)=\left|\right|x_i-{\mathrm{the\; y}}_j\left|\right|=\sqrt{\underset{p=1}{\overset{L}{\mathrm{\Σ}}}{(x_{\mathrm{ip}}-{\mathrm{the\; y}}_{\mathrm{jp}})}^2}$ (Formula 4)

4. LBG Algorithm

The LBG algorithm is the most general method for designing speaker codebook models for vector quantization. Based on the nearest neighbor condition and centroid condition of the optimal vector quantizer, the LBG algorithm uses the Euclidean distance measure (distortion measure) for the nearest neighbor division between vectors. Given the initial codebook, the Voronoi division ^R of the training vector is Definitely, according to the distortion measure nearest neighbor condition, the new cell calculated by dividing R ⁰ gets a new codebook Y ¹ , and then divides R ¹ obtained from the codebook Y ¹ ; so iteratively continues until the distortion error condition is satisfied, Or end the iteration after reaching the number of iterations, and generate the final codebook. Although this iterative process cannot guarantee that the optimal codebook can be obtained in the end, each iteration can always reduce (or keep unchanged) the average distortion, so that the performance of the codebook is gradually improved. The flowchart of the LBG algorithm is shown in Figure 2.

Although the effect of the codebook generated by LBG is relatively good, it has an obvious shortcoming, because each iteration can only change the local changes of the codebook, that is, after each iteration, compared with the old codebook, the new codebook cannot There are very large changes, so LBG has very high requirements for the initial codebook. If you choose a different initial codebook, you will get different clustering effects. If the initial codebook is not well selected, it will greatly affect the performance of the final generated codebook and the program operation time. The instability of the LBG algorithm seriously affects the design quality of the codebook, which is why people attach importance to the codebook. Design and focus on the reasons for its optimal design approach.

5. Particle pair collaborative optimization algorithm

Xue Liping et al. (Xue Liping, Yin Junxun, Zhou Jiarui, Ji Zhen. Speaker identification based on particle pair collaborative optimization. Acta Electronica Sinica, 2009, 37(1): 207-211) proposed an optimal design of a new speaker codebook model Method - Particle-Pair Cooperative Optimizer (PPCO). On the basis of the traditional particle swarm optimization (PSO) algorithm, the idea of co-evolution based on the inter-species competition mechanism is used for reference, and the information exchange is achieved through the migration of particles between the initial particle pairs to realize the co-evolution of particle pairs. The particle pair collaborative optimization algorithm uses two particles to form a particle pair with a small group size, forming a cooperative working relationship, as shown in Figure 3. In addition to performing the basic operations of the PSO algorithm (velocity update and position update) in each iteration, each particle also makes full use of the local optimization ability of LBG to execute the LBG algorithm with 3 iterations. The initial population size of the PPCO algorithm is 4 particles, which are divided into two particle pairs respectively {P ₁ , P ₂ } and {P ₃ , P ₄ }, which are used as two independent groups for speed update and Position update, a parallel search for the optimal VQ codebook in the training speech sample vector space. The two particles of the initial particle pair have the ability of self-summarization and learning from the other individual in the particle pair, so as to approach their own historical optimal point and the historical optimal point of the other individual. Every certain number of iterations, implement information exchange between particle pairs: randomly select one particle from one particle pair to exchange with another particle pair, and realize the co-evolution of particle pairs. Repeatedly search and evolve until the maximum number of iterations of the initial particle pair is met, and the better one will be selected as the elite particle. The two elite particles EP ₁ and EP ₂ selected from the two initial particle pairs are recombined into a new elite particle pair {EP ₁ , EP ₂ }, and the search and evolution are continued, and the better one EP ₃ will be selected as the final solution. The block diagram of the vector quantization speaker recognition system based on Particle Pair Cooperative Optimization Algorithm (PPCO) is shown in Figure 1.

6. Preprocessing and Feature Extraction of Speaker Speech Signal

After the speaker's voice signal is sampled, it first goes through a pre-emphasis filter H(z)=1-0.95z ^-1 , and then performs frame division and window addition. The frame length is 20ms, and the frame shift is 10ms. The window function uses a Hamming window. The 15-dimensional Mel Frequency Cepstrum Coefficient (MFCC) is extracted as the speaker's speech feature vector.

7. Establishment of the speaker's codebook model

Let the L-dimensional feature vector set of the speaker's training voice be X={x ₁ , x ₂ ,...x _T }, x _i ={ _xi1 , x _i2 ,..., x _iL }, T is the feature vector in the training voice sample set Y is a codebook composed of M L-dimensional codewords, representing the speaker's model, that is, Y={y ₁ , y ₂ , . . . , y _M }. PPCO distributes the T feature vectors of the vector set X into M clusters according to the distortion measurement criterion, and each cluster is represented by a codeword. Finally, the training vectors in each cluster are replaced by their corresponding codewords. Compared with the number T of feature vectors in the training speech sample set, the size of the codebook (the number of codewords) is much smaller. Obviously, the statistical distribution of codewords is the same as the distribution of feature vectors of original training speech samples. Therefore, on the basis of maintaining the basic information of the original distribution, the codebook greatly reduces the amount of data that needs to be processed.

PPCO trains the speaker codebook model, and optimizes the codebook design by adopting various strategies such as small group size, multi-particle pairs, collaborative optimization, PSO and LBG mixed operation, and elite particle pairing. Its main steps are as follows:

(1) Generation of the speaker's initial codebook

The population is initialized randomly, and M vectors are randomly selected from the training feature vector set as the initial codeword of each particle. For each particle, calculate its fitness value. Set the maximum number of iterations for initial particle pairs, the number of interval iterations for particle exchange between initial particles, and the maximum number of iterations for elite particle pairs.

(2) Operation of two initial particle pairs

①Update and determine p _i and p _g according to the fitness value;

② Particles update their velocity and position according to the formulas (Formula 1) and (Formula 2);

③Particles perform LBG operations with an iteration number of 3 and process empty codewords;

④ Judgment, if the number of interval iterations for exchanging particles is reached, randomly select a particle from a particle pair to exchange particles with another particle pair, and implement information exchange between particle pairs.

Repeat steps ①, ②, ③, and ④ until the maximum number of iterations for the initial particle pair is met.

(3) Update of elite particle pairs and LBG operation

After the iteration of the two initial particle pairs reaches the termination condition, two optimal particles are selected from them to form elite particle pairs. According to the method of updating the initial particle pair, use the formulas (Formula 1) and (Formula 2) to update the velocity and position of the elite particle pair and perform LBG operations until the maximum number of iterations of the elite particle pair is reached. The best particle of the elite particle pair is used as the training result to obtain the speaker codebook model.

8. Speaker Recognition

In the speaker recognition stage, the recognition process is as follows:

(1) Extract feature vector set X={x ₁ , x ₂ ,...x _T } from the test speech;

(2) Use the N codebooks established in the system to perform vector quantization on the test speech feature vector set X={x ₁ , x ₂ ,...x _T } in turn, and calculate the respective quantization errors, namely

${\mathrm{\ξ}}_{\mathrm{no}}=\frac{1}{T}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}\underset{1\≤l\≤m}{\min }{\left[d(x_t,{\mathrm{the\; y}}_l^{\mathrm{no}})\right]}^2,\mathrm{no}=\mathrm{1,2},\mathrm{LN}$ (Formula 5)

Among them, y _l ⁿ is the lth codeword vector in the nth codebook (corresponding to the nth speaker).

(3) Select the speaker corresponding to the codebook with the smallest average quantization error as the recognition result of the system.

Nine, the main shortcoming of prior art scheme

The LBG algorithm has very high requirements on the initial codebook, and different clustering effects will be obtained if different initial codebooks are selected. If the initial codebook is not well selected, it will affect the codebook performance and program operation time, thereby affecting the speaker recognition performance.

Although the codebook design quality of the particle pair collaborative optimization algorithm has been greatly improved compared with the traditional LBG algorithm, there is still a certain sensitivity to the selection of the initial codebook, which may fall into the local optimum, and cannot guarantee that the global optimum can be found. Ucodebook.

Contents of the invention

The technical problem to be solved in the present invention is to propose a three-particle collaborative optimization method applied to speaker recognition based on vector quantization, which is an improved vector quantization speaker codebook model optimization design method-three-particle collaborative optimization method (Triple- Particle Cooperative Optimizer, TPCO).

The technical solution adopted by the present invention to solve the technical problem is: to construct a three-particle collaborative optimization method applied to speaker recognition based on vector quantization, including preprocessing and feature extraction of the speaker's voice signal, and establishing a speaker codebook Model and identifying the speaker, the steps of establishing the speaker codebook model include:

A. Set the maximum number of iterations of the initial group, the number of mixed updates of the three particles, and the maximum number of iterations of the elite particles. The initial group size is 6 particles, each particle represents a codebook, and is divided into 2 subgroups according to the division rules of the hybrid leapfrog algorithm. Each subgroup is composed of three particles and named three particles. The two three particles use different particle update parameters to realize that one three particles has a good global exploration ability, and the other three particles can find the best results in each part. easy to understand;

B, performing the speed update and position update operations of the PSO algorithm for each particle of the three particles in each iteration, and performing the mixed update operation of the LBG algorithm with a number of iterations of 3;

C, the two three-particles of the initial group simultaneously perform a large-scale global detection and a small-scale local fine search in the solution space. Whenever the number of three-particle mixed updates is reached, the sub-group division rules of the hybrid leapfrog algorithm and A mixed strategy, and the two three-particles are mixed and ranked into two new three-particles, so as to realize global information exchange and co-evolution between the two three-particles;

D. When the maximum number of iterations of the initial group is satisfied, the two elite particles selected from the two three particles continue to perform speed update and position update operations until the maximum number of iterations of the elite particles is reached, and the optimal one is the speaker codebook model.

In the present invention, said step A specifically also includes:

A1, establish particle structure and fitness function;

A2, the principle of three-particle division.

In the present invention, the step B specifically also includes:

B1, particle velocity and position update strategy.

In the present invention, the optimal position of the particle is determined by the fitness function value, and the step A1 specifically includes:

(j＝1，2，…，M)代表一簇；A11, the establishment of the particle structure is based on codewords, each particle represents a codebook Y, and Y is composed of M codewords of L dimension, representing the speaker codebook model, that is, Y={y ₁ , y ₂ ,...,y _M }, cluster the speaker training feature vector set into M clusters, each codeword

(j=1, 2, ..., M) represents a cluster;

A12, the dimension of the particle is D dimension, where D=M×L, the position of the i-th particle (i=1, 2, 3) is z _i =(z _i1 , z _i2 ,..., z _iD ), the particle The value range of position z _i is z _min ~ z _max . In the process of vector quantization of speech signals, z _min and z _max generally take the minimum value and maximum value of each dimension in the speech feature vector set, and the maximum value of velocity v _i Value v _max =z _max .

In the present invention, the step A2 specifically includes:

According to the subgroup division rules of the hybrid leapfrog algorithm, the 6 particles in the group are arranged from good to bad according to their fitness values, and are divided into two subgroups SWARM _S and SWARM _L. Among them, the particle ranked first is divided into SWARM _S , ranking The second particle is classified into SWARM _L , the third particle is classified into SWARM _S , the fourth particle is classified into SWARM _L , the fifth particle is classified into SWARM _S , and the sixth particle is classified into SWARM _L.

In the present invention, in the step B1, the formulas for updating the particle velocity and position of the two three particles are (Formula 6) to (Formula 9):

$v_{\mathrm{id}}^L=w^L{v}_{\mathrm{id}}^L+c_1^L{r}_1(p_{\mathrm{id}}^L-x_{\mathrm{id}}^L)+c_2^L{r}_2(p_{\mathrm{gd}}^L-x_{\mathrm{id}}^L)$ (Formula 6)

$x_{\mathrm{id}}^L=x_{\mathrm{id}}^L+v_{\mathrm{id}}^L$ (Formula 7)

$v_{\mathrm{jd}}^S=w^S{v}_{\mathrm{jd}}^S+c_1^S{r}_1(p_{\mathrm{jd}}^S-x_{\mathrm{jd}}^S)+c_2^S{r}_2(p_{\mathrm{gd}}^S-x_{\mathrm{jd}}^S)$ (Formula 8)

$x_{\mathrm{jd}}^S=x_{\mathrm{jd}}^S+v_{\mathrm{jd}}^S$ (Formula 9)

Among them, d=1, 2,..., D, r ₁ and r ₂ are random numbers evenly distributed between [0, 1], w ^L , c ₁ ^L , c ₂ ^L are the parameters of SWARM _L , w ^S , c ₁ ^S , c ₂ ^S are parameters of SWARM _S , w is inertia weight, c ₁ , c ₂ are learning factors;

Among them, the elite particle speed update and position update formulas use (Formula 8) and (Formula 9).

In the present invention, choose the mean square error MSE between feature vector and its corresponding nearest code word as fitness function, calculate the formula of mean square error MSE as:

$\widetilde{\mathrm{D.}}=\frac{1}{T}\underset{i=1}{\overset{T}{\mathrm{\Σ}}}{\left[d_{\min }\left(x_i\right)\right]}^2$ (Formula 3)

$d(x_i,{\mathrm{the\; y}}_j)=\left|\right|x_i-{\mathrm{the\; y}}_j\left|\right|=\sqrt{\underset{p=1}{\overset{L}{\mathrm{\Σ}}}{(x_{\mathrm{ip}}-{\mathrm{the\; y}}_{\mathrm{jp}})}^2}$ (Formula 4)

表示均方误差MSE，说话人训练语音的L维特征矢量集为X＝{x₁，x₂，…x_T}，x_i＝{x_i1，x_i2，…，x_iL}，T为训练语音样本集中特征矢量的数目，Y是由M个L维的码字组成的码本，表示该说话人的模型，即Y＝{y₁，y₂，…，y_M}，

d(x_i，y_j)为Euclidean距离。in,

Indicates the mean square error MSE, the L-dimensional feature vector set of the speaker's training speech is X={x ₁ , x ₂ ,...x _T }, x _i ={x _i1 , x _i2 ,..., x _iL }, T is the training The number of feature vectors in the speech sample set, Y is a codebook composed of M codewords of L dimension, representing the model of the speaker, that is, Y={y ₁ , y ₂ ,...,y _M },

d(x _i , y _j ) is the Euclidean distance.

In the present invention, the three-particle SWARM _L selects larger w, c ₁ and relatively small c ₂ , and the three-particle SWARM _S selects relatively small w, c ₁ and relatively large c ₂ .

In the present invention, after updating the particle velocity and position, and executing the LBG algorithm with an iteration number of 3, the empty codeword is processed, and the codeword with a problem of crossing the boundary is replaced by a training vector with a large error.

The beneficial effect of the present invention is that the present invention integrates the advantages of the particle pair collaborative optimization algorithm and the hybrid leapfrog algorithm, starts from controlling information dissemination in the group, coordinates global exploration and local deep search capabilities, improves the diversity of the group, overcomes premature maturity, and improves convergence Speed and solution accuracy, the generation time of the speaker model is short, can stably achieve better speech recognition results, and can more effectively prevent particles from falling into the local optimal codebook, making the overall codebook closer to the global optimal solution, and at the same time The impact of the initial codebook on the optimization result in vector quantization is better suppressed, and the performance of speaker recognition is improved.

Description of drawings

The present invention will be further described below in conjunction with accompanying drawing and embodiment, in the accompanying drawing:

Fig. 1 is a block diagram of a speaker recognition system based on vector quantization;

Fig. 2 is prior art LBG algorithm flowchart;

Fig. 3 is a schematic diagram of particle pair collaborative optimization algorithm;

Fig. 4 is a basic flowchart of the three-particle collaborative optimization method of the present invention.

Detailed ways

In order to have a further understanding and understanding of the structural features of the present invention and the achieved effects, the preferred embodiments and accompanying drawings are used for a detailed description, as follows:

The technical problem to be solved by the present invention is to propose an improved vector quantization speaker codebook model optimization design method - triple-particle cooperative optimization method (Triple-Particle Cooperative Optimizer, TPCO). The present invention integrates the advantages of the particle pair collaborative optimization algorithm and the hybrid leapfrog algorithm, starts from the information dissemination in the control group, coordinates global exploration and local deep search capabilities, improves the diversity of the group, overcomes prematurity, and improves the convergence speed and solution accuracy. It can more effectively prevent particles from falling into the local optimal codebook, and make the overall codebook closer to the global optimal solution. At the same time, it can better suppress the influence of the initial codebook on the optimization result in vector quantization, and improve the speaker recognition performance.

1. Basic principles

The three-particle collaborative optimization method described in the present invention follows the basic idea of the initial particle pair and the elite particle pair in the PPCO algorithm, and still adopts small group size, multi-particle swarm, collaborative optimization, PSO and LBG mixed operation, and elite particle pairs. A strategy to optimize the codebook design. The initial population size of TPCO is 6 particles, which are divided into two subgroups, SWARM _L and SWARM _S , according to the division rules of SFLA. Each subgroup consists of three particles and is named as three particles. SWARM _L selects larger w, c ₁ and smaller c ₂ to realize that the particles in the three particles can disperse as much as possible to reach various positions in the solution space, so that the three particles have good global exploration ability and local The neighborhood of extreme points can escape well, and the corresponding SWARM _S chooses smaller w, c ₁ and larger c ₂ to ensure that the particles of the three particles can find the best solution in each locality. In each iteration, the particle performs the velocity update and position update of PSO and the mixed update operation of LBG algorithm with the number of iterations being 3. Two three-particles conduct large-scale global detection and small-scale local fine search in the solution space at the same time. After completing a certain number of iterations of search, all particles are mixed and ranked according to the SFLA mixing strategy to divide the new two three-particles , to realize the global information exchange and co-evolution between two three particles. Repeatedly search and evolve until the maximum number of iterations of the initial population is met. Two elite particles selected from the two three particles respectively continue to conduct fine search and evolution until the maximum number of iterations of the elite particles is reached. The optimal one will be selected selected as the final codebook model.

2. Particle structure and fitness function

(j＝1，2，…，M)代表一簇，粒子结构如表1所示。The three-particle collaborative optimization method uses a codebook-based optimization scheme, and the design of the particle structure is based on codewords. Each particle represents a codebook Y, and Y is composed of M L-dimensional codewords, representing the speaker codebook model, that is, Y={y ₁ , y ₂ , . . . , y _M }. Cluster the speaker training feature vector set into M clusters, each codeword

(j=1, 2, ..., M) represents a cluster, and the particle structure is shown in Table 1.

Table 1 Particle structure

y ₁₁ , y ₁₂ , ..., y _1L y ₂₁ , y ₂₂ , ..., y _2L ---- y _j1 , y _j2 ,..., y _jL ---- y _M1 , y _M2 , ..., y _{M L}

The dimension of the particle is D dimension, where D=M×L. The position _zi = ( _zi1 , _zi2 , . . . , _ziD ) of the i-th particle (i=1, 2, 3) is shown in Table 1. The value range of the particle position z _i is z _min ~ z _max . In the process of vector quantization of speech signals, z _min and z _max generally take the minimum and maximum values of each dimension in the speech feature vector set, respectively. The maximum value v _max =z _max of the speed v _i .

The optimal position of a particle is determined by the value of the fitness function, and the selection of the fitness function should reflect the quality of the codebook design, so the three-particle collaborative optimization method selects the mean square error (MSE) between the training vector and its corresponding nearest codeword as The fitness function is calculated using the formulas (Formula 3) and (Formula 4).

3. Principle of three-particle division

According to the subgroup division rules of the SFLA algorithm, the three-particle collaborative optimization method arranges the six particles in the group according to their fitness values from good to bad, and divides them into two subgroups. Among them, the 1st particle is classified into SWARM _S , the 2nd particle is classified into SWARM _L , the 3rd particle is classified into SWARM _S , the 4th particle is classified into SWARM _L , and the 5th particle is classified into SWARM _S. The 6th particle is classified into SWARM _L. This division rule is widely used in the hybrid leapfrog algorithm and its improved algorithm. It has more advantages than random division, and can quickly complete long-distance information transmission and global information exchange between two subgroups.

4. Particle update strategy

① Particle speed and position update

The particle velocity and position update formulas of two three-particle swarms are as (Formula 6) to (Formula 9).

$v_{\mathrm{id}}^L=w^L{v}_{\mathrm{id}}^L+c_1^L{r}_1(p_{\mathrm{id}}^L-x_{\mathrm{id}}^L)+c_2^L{r}_2(p_{\mathrm{gd}}^L-x_{\mathrm{id}}^L)$ (Formula 6)

$x_{\mathrm{id}}^L=x_{\mathrm{id}}^L+v_{\mathrm{id}}^L$ (Formula 7)

$v_{\mathrm{jd}}^S=w^S{v}_{\mathrm{jd}}^S+c_1^S{r}_1(p_{\mathrm{jd}}^S-x_{\mathrm{jd}}^S)+c_2^S{r}_2(p_{\mathrm{gd}}^S-x_{\mathrm{jd}}^S)$ (Formula 8)

$x_{\mathrm{jd}}^S=x_{\mathrm{jd}}^S+v_{\mathrm{jd}}^S$ (Formula 9)

Among them, d=1, 2,..., D; r ₁ and r ₂ are random numbers uniformly distributed between [0, 1]; w ^L , c ₁ ^L , c ₂ ^L are the parameters of SWARM _L ; w ^S , c ₁ ^S , c ₂ ^S are parameters of SWARM _S.

In order to ensure the fine search ability of elite particles, the update formula of elite particles adopts (Formula 8) and (Formula 9).

② LBG operation of particles

After the particle velocity and position are updated, the LBG algorithm with an iteration number of 3 is executed to improve the particle local optimization ability. In the process of codebook design, the value of the particle’s codeword position value exceeding the maximum value will lead to the appearance of an empty codeword. After the particle performs the LBG operation, the empty codeword will be processed, and the training vector with a large error will be used to replace the one with the out-of-bounds problem. Codeword.

5. Basic process

Table 2 gives the pseudo code of the three-particle collaborative optimization method, and its flow chart is shown in Figure 4.

Table 2 Pseudocode of the three-particle collaborative optimization method

6. Differences from PPCO

Codebook optimization is a typical multi-peak and multi-extreme optimization problem. Compared with PPCO algorithm, the three-particle collaborative optimization method further improves the design performance of codebook optimization from two aspects.

(1) Introduce the hybrid leapfrog algorithm subgroup division rules and mixed strategies to increase the number of subgroup particles and improve the search ability of subgroups.

The TPCO subgroup changes from particle pairs to three particles, further enhancing the search capability. By observing the codebook optimization process of PPCO, it can be found that the information exchange between particle pairs is beneficial to improve the optimization effect. Since a particle is randomly exchanged between two particle pairs in PPCO, there are three possible exchange methods: ① optimal particle exchange; ② worst particle exchange; ③ optimal particle and worst particle exchange. The first two methods basically maintain the distribution of good and bad particles in each particle pair, and the third method forms a pair of better particles and another pair of worse particles, which destroys the original distribution of particle pairs. The existence of three particle exchange methods is beneficial to increase the diversity of particle pairs and increase the probability of converging to the global optimum. However, since it is possible that the particles are already close to the global optimal solution, a fine excavation search is required, and the random particle exchange will destroy the distribution of the original particle pairs, affecting the possible fine search of the particle pairs, thereby reducing the convergence speed and accuracy. The subgroup division rules and mixed strategy of the hybrid leapfrog algorithm have their distinct characteristics. Individuals are assigned to each subgroup according to their fitness values from good to bad, so that each subgroup has both better individuals and poorer individuals. While improving the diversity of subgroups, the influence of better individuals in subgroups is maintained, and the distribution of good and bad individuals in subgroups is not destroyed, which is conducive to improving the convergence speed and search accuracy. Therefore, in the mixed particle pair optimization algorithm, the subgroup division rules and mixed strategy of SFLA are used to divide and recombine the two three particles.

(2) The two three-particle SWARM _L and SWARM _S select different particle update parameters to further enhance the global exploration and local fine search capabilities.

The particle update parameters of the two initial particle pairs of PPCO are the same. In order to realize that the population has global exploration and local search capabilities throughout the iterative process, it can jump from the neighborhood of the local extremum to the neighborhood of the global optimal solution. The two three-particle SWARM _S and SWARM _L of TPCO choose different particle update parameters, so that SWARM _L has a good global exploration ability. The domain can escape well, and correspondingly, SWARM _S can find the best solution in each part. The method is like a microscope with a thick lens and a thin lens, the coarse lens is used to find the object of interest in a wide range, and the fine lens is used to observe the found object carefully. Then, the information exchange between the two three particles is realized through the mixing and reorganization of the particles.

The three-particle collaborative optimization method described in the present invention is based on the PPCO algorithm, aiming at codebook optimization is a typical multi-peak multi-extreme optimization problem, using two three-particles with different update parameters to realize local deep search and global exploration , through the hybrid strategy of SFLA to realize the rapid exchange of information between the three particles, so as to better balance the global optimization and local search, so that the algorithm can jump out of the local optimum. Applying the three-particle collaborative optimization method to the speaker recognition system, the convergence speed is fast, the generation time of the speaker model is short, and it does not depend on the selection of the speaker's initial codebook, and can stably achieve better recognition results. It better solves the problem that the short speech speaker's initial codebook affects the optimization result, and can be widely used in short speech speaker recognition products.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements without departing from the spirit and scope of the technical solutions of the present invention shall be covered by the claims of the present invention.

Claims (7)

1. A three-particle collaborative optimization method applied to speaker recognition based on vector quantization, including carrying out preprocessing and feature extraction to speaker voice signals, setting up speaker codebook models and identifying speakers, characterized in that, the The steps to build a speaker codebook model include: A. Set the maximum number of iterations of the initial group, the number of mixed updates of the three particles, and the maximum number of iterations of the elite particles. The initial group size is 6 particles, each particle represents a codebook, and is divided into 2 subgroups according to the division rules of the hybrid leapfrog algorithm. Each subgroup is composed of three particles and named three particles. The two three particles use different particle update parameters to realize that one three particles has a good global exploration ability, and the other three particles can find the best results in each part. easy to understand; B, performing the speed update and position update operations of the PSO algorithm for each particle of the three particles in each iteration, and performing the mixed update operation of the LBG algorithm with a number of iterations of 3; C, the two three-particles of the initial group simultaneously perform a large-scale global detection and a small-scale local fine search in the solution space. Whenever the number of three-particle mixed updates is reached, the sub-group division rules of the hybrid leapfrog algorithm and A mixed strategy, and the two three-particles are mixed and ranked into two new three-particles, so as to realize global information exchange and co-evolution between the two three-particles; D. When the maximum number of iterations of the initial group is satisfied, the two elite particles selected from the two three particles continue to perform speed update and position update operations until the maximum number of iterations of the elite particles is reached, and the optimal one is the speaker codebook model. 2. The three-particle collaborative optimization method applied to speaker recognition based on vector quantization according to claim 1, wherein said step A specifically further comprises: A1, establish particle structure and fitness function; A2, establish the principle of three-particle division. 3. The three-particle collaborative optimization method applied to speaker recognition based on vector quantization according to claim 1, wherein said step B specifically further comprises: B1, establish particle velocity and position update strategy. 4. the three-particle collaborative optimization method that is applied to the speaker recognition based on vector quantization according to claim 2, is characterized in that, the optimum position of particle is determined by fitness function value, and described step A1 specifically comprises: A11, the establishment of the particle structure is based on codewords, each particle represents a codebook Y, and Y is composed of M codewords of L dimension, representing the speaker codebook model, that is, Y={y ₁ , y ₂ ,...,y _M }, cluster the speaker training feature vector set into M clusters, each codeword

Represents a cluster, where j=1, 2, ..., M; A12, the dimension of the particle is D dimension, where D=M×L, the position of the i-th particle is z _i =(z _i1 , z _i2 ,..., z _iD ), where i=1, 2, 3, the particle The value range of position zi is z _min ～ z _max , in the process of vector quantization of speech signal, z _min and z _max respectively take the minimum value and maximum value of each dimension in the speech feature vector set, and the maximum value v of velocity v _{i max} =z _max . 5. The three-particle collaborative optimization method applied to speaker recognition based on vector quantization according to claim 2, wherein said step A2 specifically comprises: According to the subgroup division rules of the hybrid leapfrog algorithm, the 6 particles in the group are arranged according to the fitness value obtained by the fitness function from good to bad, and divided into two subgroups SWARM _S and SWARM _L , among which, the particles ranked first The second particle is classified into SWARM _L , the third _particle is classified into SWARM _S , the fourth particle is classified into SWARM _L , the fifth particle is classified into SWARM _S , and the sixth particle is classified into SWARM _L . 6. the three-particle collaborative optimization method that is applied to the speaker recognition based on vector quantization according to claim 3, is characterized in that, in described step B1, the formula of the particle velocity update and the position update of 2 three particles is formula 6 to formula 9: $v_{\mathrm{id}}^L=w^L{v}_{\mathrm{id}}^L+c_1^L{r}_1(p_{\mathrm{id}}^L-x_{\mathrm{id}}^L)+c_2^L{r}_2(p_{\mathrm{gd}}^L-x_{\mathrm{id}}^L)$ (Formula 6) $x_{\mathrm{id}}^L=x_{\mathrm{id}}^L+v_{\mathrm{id}}^L$ (Formula 7) $v_{\mathrm{jd}}^S=w^S{v}_{\mathrm{jd}}^S+c_1^S{r}_1(p_{\mathrm{jd}}^S-x_{\mathrm{jd}}^S)+c_2^S{r}_2(p_{\mathrm{gd}}^S-x_{\mathrm{jd}}^S)$ (Formula 8) $x_{\mathrm{jd}}^S=x_{\mathrm{jd}}^S+v_{\mathrm{jd}}^S$ (Formula 9) Among them, d=1, 2, ..., D, r ₁ and r ₂ are random numbers uniformly distributed between [0, 1], w ^L ,

is the parameter of SWARM _L , w ^S ,

is the parameter of SWARM _S , w is the inertia weight, c ₁ and c ₂ are the learning factors; Among them, the elite particle speed update and position update formulas use formula 8 and formula 9. 7. the three-particle collaborative optimization method that is applied to the speaker recognition based on vector quantization according to claim 4, is characterized in that, chooses the mean square error MSE between feature vector and its corresponding nearest codeword as fitness function , the formula for calculating the mean square error MSE is: $\widetilde{\mathrm{D.}}=\frac{1}{T}\underset{i=1}{\overset{T}{\mathrm{\Σ}}}{\left[d_{\min }\left(x_i\right)\right]}^2$ (Formula 3) $d(x_i,{\mathrm{the\; y}}_j)=\left|\right|x_i-{\mathrm{the\; y}}_j\left|\right|=\sqrt{\underset{p=1}{\overset{L}{\mathrm{\Σ}}}{(x_{\mathrm{ip}}-{\mathrm{the\; y}}_{\mathrm{jp}})}^2}$ (Formula 4) in,

Indicates the mean square error MSE, the L-dimensional feature vector set of the speaker's training speech is X={x ₁ , x ₂ ,...x _T }, x _i ={x _i1 , x _i2 ,..., x _iL }, T is the training The number of feature vectors in the speech sample set, Y is a codebook composed of M codewords of L dimension, representing the model of the speaker, that is, Y={y ₁ , y ₂ ,...,y _M },

d(x _i , y _j ) is the Euclidean distance.

Images