Noise feature vector classification model construction method, speech enhancement method and device

Abstract

The invention relates to a noise feature vector classification model construction method, a speech enhancement method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: coding penalty parameters and kernel function parameters of a support vector machine; taking the penalty parameter and the kernel function parameter as optimization variables of a genetic algorithm, setting an initial population, and constructing a fitness evaluation function based on classification accuracy; performing selection operation, crossover operation and mutation operation on the initial population, and updating an iterative population according to a fitness evaluation function to obtain an optimal penalty parameter and an optimal kernel function parameter; and training a support vector machine through the optimal penalty parameter and the optimal kernel function parameter, and constructing a noise feature vector classification model. In a noisy environment, the classification effect of the noise feature vector is improved under the condition that nonlinear and non-stationary noise exists, and the noise feature classification requirement in a complex sound signal processing scene is met.

Description

Technical Field

[0001] The present application relates to the technical field of speech signal processing, and in particular to a noise feature vector classification model construction method, a speech enhancement method, an apparatus, a computer device, a storage medium and a computer program product. Background Art

[0002] With the development and progress of artificial intelligence technology, speech recognition has been widely used in various fields. It can convert human voice signals into corresponding text or commands, thereby realizing the intelligentization of human-computer interaction. Speech recognition technology plays an important role in voice assistants, autonomous driving, smart homes and other fields.

[0003] In traditional technologies, in terms of speech signal processing, in order to improve the robustness and accuracy of speech recognition technology, speech signal processing technology is also in the process of continuous development. This includes speech enhancement, speech denoising, speech coding and other technologies, which can effectively improve the quality of speech signals and improve the effect of speech recognition.

[0004] At present, the speech recognition rate of voice signals in a relatively quiet outdoor environment has basically met the user's usage needs. However, in outdoor environments with a large number of people, the audio collected by the terminal is full of various nonlinear and non-stationary noises. The recognition accuracy needs to be further improved, especially in response to the influence of sudden noise, which has a significant impact on subsequent speech semantic recognition. Summary of the invention

[0005] Based on this, it is necessary to provide a noise feature vector classification model construction method, speech enhancement method, device, computer equipment, storage medium and computer program product that can improve the accuracy of speech signal recognition in response to the above technical problems.

[0006] In a first aspect, the present application provides a method for constructing a noise feature vector classification model. The method comprises:

[0007] Encode the penalty parameters and kernel function parameters of the support vector machine;

[0008] The penalty parameters and kernel function parameters are used as optimization variables of the genetic algorithm, and the initial population is set to construct a fitness evaluation function based on classification accuracy.

[0009] Perform selection, crossover and mutation operations on the initial population, update the iterative population according to the fitness evaluation function, and obtain the optimal penalty parameters and the optimal kernel function parameters;

[0010] The support vector machine is trained by using the optimal penalty parameters and the optimal kernel function parameters to build a noise feature vector classification model.

[0011] In one embodiment, the selection operation on the initial population includes:

[0012] According to the initial population, the fitness value of the individuals in the initial population is obtained; the individual is a combination parameter of a set of penalty parameters and kernel function parameters;

[0013] Arrange the fitness values ​​of individuals in descending order and divide them into three parts evenly;

[0014] The part of the individual with the highest fitness is copied twice, and the part of the individual with medium fitness is copied once, and the new population is obtained by merging them.

[0015] In one embodiment, performing a crossover operation on the initial population includes:

[0016] According to the gene similarity of the parent individuals, determine whether the gene similarity is greater than the crossover critical threshold, and output the judgment result of the gene similarity;

[0017] Based on the results of gene similarity judgment, gene recombination is performed on parent individuals with low gene similarity;

[0018] Dynamically adjust the crossover probability according to the change of individual fitness during the population iteration process; the crossover probability is used to adjust the frequency of gene recombination;

[0019] The offspring individuals are genetically recombined by crossover probability, and the population is updated by partial mapping crossover.

[0020] In one embodiment, performing a mutation operation on the initial population includes:

[0021] Based on the individual fitness value, dynamically adjust the mutation probability of the mutation operator;

[0022] Determine whether to perform mutation operation on individuals according to the preset fitness threshold and mutation probability;

[0023] When the fitness value of an individual is greater than the fitness threshold, no mutation operation is performed;

[0024] When the fitness value of an individual is not greater than the fitness threshold, the individual is randomly mutated based on the mutation probability to update the population.

[0025] In one embodiment, the method further comprises:

[0026] The classification accuracy of the K-fold cross validation of the support vector machine is used as the standard for evaluating the fitness value of the individual.

[0027] In a second aspect, the present application provides a speech enhancement method. The method comprises:

[0028] Acquire a target speech signal and preprocess the target speech signal;

[0029] Performing feature extraction on the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector;

[0030] Call a pre-trained noise feature vector classification model such as any embodiment of the first aspect to classify the target speech feature vector and the noise feature vector; suppress the noise feature vector, retain the target speech feature vector, and reconstruct the enhanced target speech signal to output the enhanced speech signal.

[0031] In a second aspect, the present application also provides a speech enhancement device. The device comprises:

[0032] An acquisition module, used for acquiring a target speech signal and preprocessing the target speech signal;

[0033] A feature extraction module is used to extract features from the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector;

[0034] The enhancement module is used to call a pre-trained noise feature vector classification model such as any embodiment of the first aspect to classify the target speech feature vector and the noise feature vector; suppress the noise feature vector, retain the target speech feature vector, and reconstruct the enhanced target speech signal to output the enhanced speech signal.

[0035] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0036] Encode the penalty parameters and kernel function parameters of the support vector machine;

[0037] The penalty parameters and kernel function parameters are used as optimization variables of the genetic algorithm, and the initial population is set to construct a fitness evaluation function based on classification accuracy.

[0038] Perform selection, crossover and mutation operations on the initial population, update the iterative population according to the fitness evaluation function, and obtain the optimal penalty parameters and the optimal kernel function parameters;

[0039] The support vector machine is trained by using the optimal penalty parameters and the optimal kernel function parameters to build a noise feature vector classification model.

[0040] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0041] Encode the penalty parameters and kernel function parameters of the support vector machine;

[0042] The penalty parameters and kernel function parameters are used as optimization variables of the genetic algorithm, and the initial population is set to construct a fitness evaluation function based on classification accuracy.

[0043] Perform selection, crossover and mutation operations on the initial population, update the iterative population according to the fitness evaluation function, and obtain the optimal penalty parameters and the optimal kernel function parameters;

[0044] The support vector machine is trained by using the optimal penalty parameters and the optimal kernel function parameters to build a noise feature vector classification model.

[0045] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0046] Encode the penalty parameters and kernel function parameters of the support vector machine;

[0047] The penalty parameters and kernel function parameters are used as optimization variables of the genetic algorithm, and the initial population is set to construct a fitness evaluation function based on classification accuracy.

[0048] Perform selection, crossover and mutation operations on the initial population, update the iterative population according to the fitness evaluation function, and obtain the optimal penalty parameters and the optimal kernel function parameters;

[0049] The support vector machine is trained by using the optimal penalty parameters and the optimal kernel function parameters to build a noise feature vector classification model.

[0050] The noise feature vector classification model construction method, speech enhancement method, device, computer equipment, storage medium and computer program product achieve high efficiency in model parameter optimization by combining support vector machine with genetic algorithm; the penalty parameters and kernel function parameters of the support vector machine are globally searched and optimized by using genetic algorithm, which overcomes the defect of traditional methods that are prone to fall into local optimality, thereby finding the optimal parameter combination and significantly improving the classification performance of noise feature vectors. On the other hand, the selection operation, crossover operation and mutation operation that can be adjusted dynamically further enhance the search ability and population diversity of the genetic algorithm, which not only accelerates the convergence process, but also effectively avoids the risk of falling into local optimal solution; in noisy environment and in the presence of nonlinear non-stationary noise, the classification effect of noise feature vectors is improved to meet the noise feature classification requirements in complex sound signal processing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] FIG1 is an application environment diagram of a noise feature vector classification model construction method according to an embodiment;

[0052] FIG2 is a schematic diagram of a process of constructing a noise feature vector classification model in one embodiment;

[0053] FIG3 is a schematic diagram of a selection operation in one embodiment;

[0054] FIG4 is a schematic diagram of a flow chart of a speech enhancement method in one embodiment;

[0055] FIG5 is a flow chart of a speech enhancement method in another embodiment;

[0056] FIG6 is a structural block diagram of a speech enhancement device in one embodiment;

[0057] FIG7 is a structural block diagram of a noise feature vector classification model building device in one embodiment;

[0058] FIG. 8 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0060] The noise feature vector classification model construction method provided in the embodiment of the present application can be applied to the application environment shown in Figure 1. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. Among them, the terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented with an independent server or a server cluster consisting of multiple servers.

[0061] In one embodiment, as shown in FIG2 , a method for constructing a noise feature vector classification model is provided, and the method is applied to the server in FIG1 as an example for explanation, including the following steps:

[0062] Step 202: Encode the penalty parameters and kernel function parameters of the support vector machine.

[0063] Among them, Support Vector Machine (SVM) is a machine learning algorithm for classification and regression tasks. Its core idea is to divide data into different categories by finding an optimal separation hyperplane (a straight line or curve in low dimensions); it is particularly suitable for dealing with complex, nonlinear classification problems. The penalty parameter is a parameter used in the support vector machine to balance the model complexity and classification error. A larger value focuses more on reducing classification errors, but may lead to overfitting; a smaller value focuses more on simplifying the model, but may increase classification errors. The kernel function parameter is used to adjust the support vector machine's mapping ability to the input data. When the value is small, the model has stronger global characteristics, and when the value is large, the model pays more attention to local characteristics.

[0064] Step 204: Use the penalty parameter and the kernel function parameter as optimization variables of the genetic algorithm, set the initial population, and construct a fitness evaluation function based on classification accuracy.

[0065] Among them, the fitness evaluation function is used to evaluate the quality of individuals in the genetic algorithm, reflecting the quality of individuals in the solution space.

[0066] Step 206, performing selection operation, crossover operation and mutation operation on the initial population, updating the iterative population according to the fitness evaluation function, and obtaining the optimal penalty parameter and the optimal kernel function parameter.

[0067] Among them, the selection operation is used to adjust the overall trend of the population and determine the development direction of the next generation of individuals. Using a suitable selection operator will increase the probability of selecting excellent offspring. The crossover operation is used to recombine the genes of the parent individuals with a certain probability to produce new offspring individuals, which can improve the convergence speed and global search ability of the genetic algorithm. The mutation operation is used to randomly change the gene fragments of individuals, increase the diversity of the population, and avoid falling into the local optimal solution.

[0068] Step 208: Train the support vector machine using the optimal penalty parameter and the optimal kernel function parameter to construct a noise feature vector classification model.

[0069] In the above noise feature vector classification model construction method, by combining support vector machine with genetic algorithm, high efficiency of model parameter optimization is achieved; genetic algorithm is used to perform global search and optimization on the penalty parameters and kernel function parameters of support vector machine, overcoming the defect of traditional methods that are prone to fall into local optimality, thereby finding the optimal parameter combination and significantly improving the classification performance of noise feature vector. On the other hand, the selection operation, crossover operation and mutation operation that can be adjusted dynamically further enhance the search ability and population diversity of genetic algorithm, which not only accelerates the convergence process, but also effectively avoids the risk of falling into local optimal solution; in noisy environment and in the presence of nonlinear non-stationary noise, the classification effect of noise feature vector is improved to meet the noise feature classification needs in complex sound signal processing scenarios.

[0070] In one embodiment, the selection operation for the initial population includes: according to the initial population, obtaining the fitness value of the individuals in the initial population; the individuals are a combination parameter of a set of penalty parameters and kernel function parameters; the fitness values ​​of the individuals are arranged in descending order, and the individuals are evenly divided into three parts; the part of the individual with the highest fitness is copied twice, and the part of the individual with medium fitness is copied once, and the new population is obtained by merging.

[0071] Exemplarily, in combination with Figure 3, the penalty parameter 𝐶 and kernel function parameter 𝛾 of the support vector machine are encoded according to the set range and precision. The value range of the penalty parameter 𝐶 is [0.1, 100], represented by 10 bits of binary; the value range of the kernel function parameter 𝛾 is [0.01, 10], represented by 10 bits of binary. Each set of encoding results is composed of two parts, 𝐶 and 𝛾, to form an individual gene. The elite retention method is used to optimize the selection process, and the fitness value of each individual is obtained in turn, and the fitness values ​​are arranged in descending order, and the individuals are evenly divided into three parts, each of which meets the following formula:

[0072]

[0073] Among them, g(a j ) is the part with the highest individual fitness, and g(a j ) copies two copies to the offspring; g(b j ) is the part with medium fitness, and g(b j ) copies a copy to the offspring; g(c j ) is the part with the lowest individual fitness, and g(c j ) part, and finally a new offspring population is obtained with the same population size.

[0074] In this embodiment, individuals with higher fitness are preferentially retained and replicated to the next generation to avoid the loss of good genes; by segmenting the population according to the fitness value and increasing the retention ratio of offspring of individuals with medium fitness, it can not only ensure the transmission of good genes, but also enrich the diversity of the population, avoiding falling into local optimality due to excessive concentration on high fitness individuals; it is suitable for different population sizes and target optimization scenarios, and shows strong robustness in complex nonlinear optimization problems.

[0075] In one embodiment, performing a crossover operation on an initial population includes: determining whether the gene similarity is greater than a critical crossover threshold based on the gene similarity of the parent individuals, and outputting a determination result of the gene similarity; performing gene recombination on the parent individuals with low gene similarity based on the determination result of the gene similarity; dynamically adjusting the crossover probability based on changes in individual fitness during population iteration; the crossover probability is used to adjust the frequency of gene recombination; and determining gene recombination on offspring individuals based on the crossover probability, and updating the population using partial mapping crossover.

[0076] Exemplarily, obtain the gene codes of the parent individuals in the initial population. For example, a population contains 10 individuals, and the genes of each individual are represented by binary codes. Calculate the genetic similarity between any two parent individuals. Set a critical crossover threshold to determine whether the genetic similarity between the parent individuals is greater than the threshold; if the similarity is greater than the threshold, skip the crossover operation and retain the parent genes; if the similarity is less than or equal to the threshold, perform genetic recombination on the group of parents to generate offspring individuals. Adjust the crossover probability according to the change in individual fitness during the population iteration process. Specifically, randomly generate two numbers m and n, and cross-exchange the genes within the digital interval. The similarity expression is:

[0077]

[0078] Where L is the common gene length between m and n, N is the total gene length, and t is the population size.

[0079] The crossover probability changes according to the change of individual fitness during the population iteration process. The crossover probability expression is:

[0080]

[0081] In the formula, p jmax and p jinx are the maximum and minimum crossover probabilities, respectively, and f max and f min are the maximum and minimum values ​​of the fitness function, respectively, c is the larger of the two crossover individual fitness values.

[0082] In this embodiment, a dynamic adjustment mechanism based on crossover probability realizes the adjustment of the frequency of gene recombination of offspring, avoids the homogenization of gene regions within the population, improves the robustness and optimization efficiency of the genetic algorithm, and makes the model more sensitive to the detection and classification of sudden noise, and can quickly identify and isolate suddenly appearing noise signals.

[0083] In one embodiment, performing a mutation operation on an initial population includes: dynamically adjusting the mutation probability of a mutation operator based on the fitness value of the individual; determining whether to perform a mutation operation on the individual based on a preset fitness threshold and mutation probability; not performing a mutation operation when the fitness value of the individual is greater than the fitness threshold; and performing a random mutation operation on the individual based on the mutation probability to update the population when the fitness value of the individual is not greater than the fitness threshold.

[0084] Specifically, for each individual in the initial population, the fitness value of each individual is obtained based on the preset fitness evaluation function. The fitness value reflects the degree of excellence of the individual in the solution space and is used to guide the decision of the mutation operation. According to the fitness value range of all individuals in the population, the mutation probability of the mutation operator is dynamically adjusted. The specific expression is as follows:

[0085]

[0086] In the formula, p nmax and p nmin are the maximum and minimum crossover probability values, respectively, and f max and f min are the maximum and minimum values ​​of the fitness function, respectively, n is the larger of the two individual mutation fitness values.

[0087] According to the preset fitness threshold and the dynamically adjusted mutation probability p n , determine whether the individual needs to be mutated; when the fitness value of the individual is greater than the preset fitness threshold, it is regarded as an excellent individual and no mutation operation is performed to retain its excellent genes; when the fitness value of the individual is not greater than the preset fitness threshold, decide whether to mutate the individual according to the mutation probability 𝑃. For individuals that need to be mutated, randomly select their gene sites, and mutate the gene value of the site through a certain random perturbation. For example, for individuals with real gene values, gene mutation can add a random number near their value for adjustment; for individuals with discrete gene values, a random number can be selected from the candidate values ​​to replace the current gene value, and the mutated individual is updated to the population. Repeat the above mutation operation until the population reaches the iteration stop condition, and finally obtain an optimized population containing high fitness individuals, providing an excellent solution space foundation for subsequent operations.

[0088] In this embodiment, the crossover probability changes according to the change of individual fitness value during the population iteration process. The dynamic adjustment of the mutation probability can adaptively change the genetic diversity of the population according to the fitness value, avoiding the limitations brought by the fixed mutation rate; retaining individuals with high fitness to prevent the overall quality of the population from declining, while improving the convergence speed; retaining excellent individual genes to prevent the deterioration of the genetic goodness of the offspring, as well as reducing the convergence speed and global search capability.

[0089] In one embodiment, the method further includes: using the classification accuracy of the K-fold cross validation of the support vector machine as a standard for evaluating the fitness value of the individual.

[0090] Among them, K-fold cross validation is used to evaluate the performance of machine learning models. The purpose is to reduce the randomness of the evaluation results by dividing and training the data set multiple times. The specific method is to divide the data set into 𝐾 parts, use 𝐾-1 of them as the training set each time, and the remaining one as the test set. After repeating 𝐾 times, calculate the average classification accuracy of all test sets as the comprehensive performance of the model.

[0091] In this embodiment, in the process of optimizing the support vector machine parameters by the genetic algorithm, the K-fold cross-validation classification accuracy of the support vector machine is used as a criterion for evaluating individual fitness. That is, each individual in the genetic algorithm (i.e., a combination of a set of penalty parameters and kernel function parameters) will be used to train the support vector machine model, and its classification accuracy will be evaluated by 𝐾-fold cross-validation. The higher the classification accuracy, the greater the improvement of the model performance by the parameter combination, and the higher its fitness value; conversely, the lower the fitness value. Based on this, the genetic algorithm prioritizes retaining or adjusting individuals with high fitness, and gradually optimizes the support vector machine parameters to obtain the optimal classification performance.

[0092] In a second aspect, as shown in FIG4 , the present application provides a speech enhancement method, the method comprising:

[0093] Step 402: Acquire a target speech signal and preprocess the target speech signal.

[0094] Step 404: extract features from the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector.

[0095] Step 406, calling a pre-trained noise feature vector classification model such as any embodiment of the first aspect, classifies the target speech feature vector and the noise feature vector; suppresses the noise feature vector, retains the target speech feature vector, and reconstructs the enhanced target speech signal, and outputs the enhanced speech signal.

[0096] Specifically, the target speech signal is collected through a microphone or other audio input device, and the collected signal may contain environmental noise and other interference; the collected target speech signal is preprocessed, including denoising, signal equalization, amplitude normalization, and pre-emphasis filtering to improve the signal quality and provide a basis for subsequent feature extraction. Feature extraction is performed on the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector. Exemplarily, the target speech signal is decomposed into the time-frequency domain through signal processing methods such as short-time Fourier transform, and key features such as Mel-frequency cepstral coefficients, spectral entropy, signal-to-noise ratio, etc. are extracted. The result of feature extraction includes the feature vector of the target speech and the mixed noise feature vector. Using the noise feature vector classification model, the extracted speech feature vector is input, and the target speech feature vector and the noise feature vector are classified through the support vector machine model. For the noise feature vector, filtering or masking strategies are used to suppress it; for the target speech feature vector, amplification or optimization processing is performed to ensure the clarity and fidelity of the enhanced speech signal; after classification, the processed feature vector is subjected to inverse feature extraction, such as through inverse short-time Fourier transform, to reconstruct the enhanced target speech signal and output the final clear speech result.

[0097] In this embodiment, through preprocessing and feature extraction, combined with a support vector machine classification model optimized based on a genetic algorithm, accurate classification of target speech features and noise features is effectively achieved; by dynamically suppressing noise feature vectors, the target speech signal is retained and enhanced, which not only improves the clarity and fidelity of the speech signal, but also significantly improves the applicability and robustness of the speech enhancement system in complex noise environments; it is suitable for a variety of speech processing scenarios, can reduce the impact of noise on speech recognition, voice calls and voice interaction systems, improve user experience, and at the same time has efficient processing capabilities to support real-time application scenarios.

[0098] In one embodiment, as shown in FIG5 , a speech enhancement method is provided, comprising the following steps:

[0099] Step 502: Encode the penalty parameters and kernel function parameters of the support vector machine.

[0100] Step 504: Use the penalty parameter and the kernel function parameter as optimization variables of the genetic algorithm, set the initial population, and construct a fitness evaluation function based on classification accuracy.

[0101] Step 506, according to the initial population, obtain the fitness value of the individuals in the initial population; the individual is a combination parameter of a set of penalty parameters and kernel function parameters.

[0102] Step 508, sort the fitness values ​​of the individuals in descending order, and divide the individuals evenly into three parts.

[0103] Step 510, duplicate the part of the individual with the highest fitness twice, duplicate the part of the individual with the medium fitness once, and merge them to obtain a new population.

[0104] Step 512, based on the gene similarity of the parent individuals, determine whether the gene similarity is greater than the crossover critical threshold, and output the determination result of the gene similarity.

[0105] Step 514, based on the determination result of gene similarity, gene recombination is performed on the parent individuals with low gene similarity.

[0106] Step 516, dynamically adjust the crossover probability according to the change of individual fitness during the population iteration process; the crossover probability is used to adjust the frequency of gene recombination.

[0107] Step 518, perform genetic recombination on offspring individuals by crossover probability determination, and use partial mapping crossover to update the population.

[0108] Step 520: dynamically adjust the mutation probability of the mutation operator based on the fitness value of the individual.

[0109] Step 522, judging whether to perform mutation operation on the individual according to the preset fitness threshold and mutation probability.

[0110] Step 524: when the fitness value of the individual is greater than the fitness threshold, no mutation operation is performed.

[0111] Step 526: When the fitness value of the individual is not greater than the fitness threshold, the individual is randomly mutated based on the mutation probability to update the population.

[0112] Step 528, using the classification accuracy of the K-fold cross validation of the support vector machine as a standard for evaluating the fitness value of the individual.

[0113] Step 530, updating the iterative population according to the fitness evaluation function to obtain the optimal penalty parameter and the optimal kernel function parameter.

[0114] Step 532: Train the support vector machine using the optimal penalty parameter and the optimal kernel function parameter to construct a noise feature vector classification model.

[0115] Step 534: Acquire a target speech signal and preprocess the target speech signal.

[0116] Step 536: extract features from the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector.

[0117] Step 538, calling the pre-trained noise feature vector classification model to classify the target speech feature vector and the noise feature vector; suppressing the noise feature vector, retaining the target speech feature vector, and reconstructing the enhanced target speech signal to output the enhanced speech signal.

[0118] In this embodiment, by combining the support vector machine with the genetic algorithm, the high efficiency of the model parameter optimization is achieved; the genetic algorithm is used to perform global search and optimization on the penalty parameters and kernel function parameters of the support vector machine, overcoming the defect of the traditional method that is easy to fall into the local optimum, thereby finding the optimal parameter combination and significantly improving the classification performance of the noise feature vector. On the other hand, the selection operation, crossover operation and mutation operation that can be adjusted dynamically further enhance the search ability and population diversity of the genetic algorithm, which not only accelerates the convergence process, but also effectively avoids the risk of falling into the local optimal solution; in noisy environments and in the presence of nonlinear non-stationary noise, the classification effect of the noise feature vector is improved to meet the noise feature classification requirements in complex sound signal processing scenarios.

[0119] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are shown in sequence, these steps are not necessarily executed in sequence in the order indicated. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.

[0120] Based on the same inventive concept, the embodiment of the present application also provides a speech enhancement device for implementing the speech enhancement method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more speech enhancement device embodiments provided below can refer to the limitations on the speech enhancement method above, and will not be repeated here.

[0121] In a third aspect, in one embodiment, as shown in FIG6 , a speech enhancement device is provided, including: an acquisition module 602, a feature extraction module 604 and an enhancement module 606, wherein:

[0122] The acquisition module 602 is used to acquire a target speech signal and pre-process the target speech signal.

[0123] The feature extraction module 604 is used to extract features from the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector.

[0124] The enhancement module 606 is used to call a pre-trained noise feature vector classification model such as any one of the first aspect to classify the target speech feature vector and the noise feature vector; suppress the noise feature vector, retain the target speech feature vector, and reconstruct the enhanced target speech signal to output the enhanced speech signal.

[0125] In a fourth aspect, in one embodiment, as shown in FIG. 7 , a noise feature vector classification model construction device is provided, comprising: an encoding module 702, an evaluation module 704, an iteration module 706 and a construction module 708, wherein:

[0126] The encoding module 702 is used to encode the penalty parameters and kernel function parameters of the support vector machine.

[0127] The evaluation module 704 is used to use the penalty parameter and the kernel function parameter as optimization variables of the genetic algorithm, set the initial population, and construct a fitness evaluation function based on classification accuracy.

[0128] The iteration module 706 is used to perform selection operation, crossover operation and mutation operation on the initial population, update the iteration population according to the fitness evaluation function, and obtain the optimal penalty parameter and the optimal kernel function parameter.

[0129] The construction module 708 is used to train the support vector machine through the optimal penalty parameter and the optimal kernel function parameter to construct a noise feature vector classification model.

[0130] In one embodiment, the iteration module 706 includes:

[0131] The fitness module is used to obtain the fitness value of individuals in the initial population based on the initial population; the individual is a combination parameter of a set of penalty parameters and kernel function parameters.

[0132] The arrangement module is used to arrange the fitness values ​​of individuals in descending order and divide the individuals evenly into three parts.

[0133] Create a new module to copy the part of the individuals with the highest fitness twice and the part of the individuals with medium fitness once, and merge them to get a new population.

[0134] In one embodiment, the iteration module 706 further includes:

[0135] The similarity module is used to determine whether the gene similarity is greater than the cross-critical threshold based on the gene similarity of the parent individuals, and output the judgment result of the gene similarity.

[0136] The recombination module is used to perform gene recombination on parent individuals with low gene similarity based on the judgment results of gene similarity.

[0137] The frequency adjustment module is used to dynamically adjust the crossover probability according to the changes in individual fitness during the population iteration process; the crossover probability is used to adjust the frequency of gene recombination.

[0138] The crossover update module is used to perform genetic recombination on offspring individuals by judging the crossover probability, and to update the population using partial mapping crossover.

[0139] In one embodiment, the iteration module 706 further includes:

[0140] The mutation probability module is used to dynamically adjust the mutation probability of the mutation operator based on the individual fitness value.

[0141] The mutation operation module is used to determine whether to perform a mutation operation on an individual based on a preset fitness threshold and mutation probability.

[0142] The first processing module is used for not performing a mutation operation when the fitness value of the individual is greater than the fitness threshold.

[0143] The second processing module is used to update the population by performing random mutation on the individual based on the mutation probability when the fitness value of the individual is not greater than the fitness threshold.

[0144] In one embodiment, the apparatus further comprises:

[0145] The accuracy evaluation module is used to use the classification accuracy of the K-fold cross validation of the support vector machine as a standard for evaluating the fitness value of the individual.

[0146] Each module in the above-mentioned speech enhancement device and noise feature vector classification model construction device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0147] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG8. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a noise feature vector classification model construction method or a speech enhancement method is implemented.

[0148] Those skilled in the art will understand that the structure shown in FIG. 8 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0149] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0150] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0151] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0152] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0153] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0154] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0155] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for constructing a noise feature vector classification model, characterized in that: The method comprises: Encode the penalty parameters and kernel function parameters of the support vector machine; The penalty parameter and the kernel function parameter are used as optimization variables of the genetic algorithm, and an initial population is set to construct a fitness evaluation function based on classification accuracy; Performing selection, crossover and mutation operations on the initial population, updating the iterative population according to the fitness evaluation function, and obtaining an optimal penalty parameter and an optimal kernel function parameter; The support vector machine is trained by using the optimal penalty parameter and the optimal kernel function parameter to construct a noise feature vector classification model.

2. The method according to claim 1, characterized in that: The selection operation on the initial population includes: According to the initial population, the fitness value of the individual in the initial population is obtained; the individual is a set of combined parameters of the penalty parameter and the kernel function parameter; Arrange the fitness values ​​of the individuals in descending order, and evenly divide the individuals into three parts; The part of the individual with the highest fitness is copied twice, and the part of the individual with medium fitness is copied once, and the new population is obtained by merging them.

3. The method according to claim 1, characterized in that Performing a crossover operation on the initial population includes: According to the gene similarity of the parent individuals, determine whether the gene similarity is greater than the crossover critical threshold, and output the determination result of the gene similarity; Based on the determination result of the gene similarity, performing gene recombination on the parent individuals with low gene similarity; Dynamically adjust the crossover probability according to the change of individual fitness during the population iteration process; the crossover probability is used to adjust the frequency of gene recombination; The offspring individuals are genetically recombined by judging the crossover probability, and the population is updated by adopting partial mapping crossover.

4. The method according to claim 2, characterized in that: The mutation operation on the initial population includes: Dynamically adjusting the mutation probability of the mutation operator based on the fitness value of the individual; Determining whether to perform a mutation operation on the individual according to a preset fitness threshold and the mutation probability; When the fitness value of the individual is greater than the fitness threshold, no mutation operation is performed; In the case that the fitness value of the individual is not greater than the fitness threshold, the individual is subjected to random mutation based on the mutation probability to update the population.

5. The method according to claim 1, characterized in that: The method further comprises: The classification accuracy of the K-fold cross validation of the support vector machine is used as the standard for evaluating the fitness value of the individual.

6. A speech enhancement method, characterized in that: The method comprises: Acquire a target speech signal, and preprocess the target speech signal; Extracting features from the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector; Calling the pre-trained noise feature vector classification model as described in claims 1-5 to classify the target speech feature vector and the noise feature vector; suppressing the noise feature vector, retaining the target speech feature vector, and reconstructing the enhanced target speech signal to output the enhanced speech signal.

7. A speech enhancement device, characterized in that: The device comprises: An acquisition module, used for acquiring a target speech signal and preprocessing the target speech signal; A feature extraction module is used to extract features from the preprocessed target speech signal to obtain a target speech feature vector and a noise feature vector; The enhancement module is used to call the pre-trained noise feature vector classification model as described in claims 1-5 to classify the target speech feature vector and the noise feature vector; suppress the noise feature vector, retain the target speech feature vector, and reconstruct the enhanced target speech signal to output the enhanced speech signal.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images