Speech processing method and system based on LLM model, electronic equipment and storage medium

Abstract

The invention provides a speech processing method and system based on an LLM model, electronic equipment and a computer storage medium, and the method comprises the steps: obtaining a plurality of language recognition learning models on the same hardware platform, and guaranteeing that all the language recognition learning models are operated in a fair resource environment; obtaining vocal appeal text data, and respectively inputting the vocal appeal text data into each language recognition learning model; running all the language recognition learning models in parallel, comparing, analyzing and predicting the vocal appeal text data with appeal tags in a language recognition learning model appeal tag library, and outputting respective prediction results by each language recognition learning model; and performing statistical scoring on prediction results output by each language recognition learning model according to the weight information, and obtaining a most matched vocal appeal text data recognition result according to the score. According to the method, the influence of subjectivity is reduced, and the objectivity and stability of prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of text classification, and in particular to a speech processing method, system, electronic device and storage medium based on an LLM model. Background Art

[0002] The traditional BERT has limited ability to understand long texts and cannot effectively learn more contextual information. Secondly, it cannot handle diversity better and does not have a thorough understanding of more vocabulary and grammatical rules. In addition, it cannot handle ambiguity better and does not have a clear understanding of more contextual information.

[0003] In addition, due to differences in training samples, large models have inconsistent understanding capabilities for the same text. This leads to differences in the understanding capabilities of different models when processing the same text. In order to improve the effect of text classification, we used multiple large language models (LLMs) for experiments and output text classification results by voting in order to obtain the best effect. Summary of the invention

[0004] The embodiment of the present invention provides a speech processing method, system, electronic device and storage medium based on the LLM model, aiming to solve the problems of strong subjectivity and unstable classification results in the traditional customer voice appeal classification method.

[0005] The technical solution of the present invention is:

[0006] In a first aspect, the present application provides a speech processing method based on an LLM model, characterized by comprising:

[0007] Obtain multiple language recognition learning models on the same hardware platform to ensure that all language recognition learning models run in a fair resource environment;

[0008] Acquire voiced appeal text data, and input the voiced appeal text data into each language recognition learning model respectively;

[0009] All language recognition learning models are run in parallel, and the utterance demand text data is compared, analyzed and predicted with the demand labels in the demand label library of the language recognition learning model, and each language recognition learning model outputs its own prediction result, wherein the prediction result includes the weight information of the prediction result in the demand label library of the language recognition learning model;

[0010] The prediction results output by each language recognition learning model are statistically scored based on the weight information, and the most matching voice demand text data recognition result is obtained based on the score.

[0011] Furthermore, the specific method of comparing, analyzing and predicting the uttered appeal text data with the appeal labels in the appeal label library of the language recognition learning model, and each language recognition learning model outputting its own prediction result includes:

[0012] Perform data cleaning on the demand labels in the demand label library, and perform feature extraction after data cleaning;

[0013] Cluster the feature vectors and obtain preliminary preprocessing, determine the cluster center labels, adjust the cluster boundaries, and then iteratively update the cluster centers and boundaries;

[0014] The weight of each cluster is calculated based on the number of labels in the cluster. The more labels there are in the cluster, the greater the weight of the corresponding cluster center label, and vice versa.

[0015] Furthermore, clustering the feature vectors and obtaining preliminary clustering results, determining cluster center labels, adjusting cluster boundaries, and then iteratively updating cluster centers and boundaries include:

[0016] The preprocessed label feature vectors are clustered using the K-means algorithm to determine the optimal number of clusters K and obtain preliminary clustering results;

[0017] The centroid of each cluster is selected as the representative label of the cluster, that is, the cluster center label;

[0018] A greedy algorithm is used to gradually adjust the cluster boundaries to optimize the clustering results.

[0019] Furthermore, the method of using a greedy algorithm to gradually adjust cluster boundaries to optimize clustering results includes:

[0020] a1. Determine the initial cluster center through K-means algorithm;

[0021] a2. Determine the initialization boundary of each cluster based on the initial clustering results;

[0022] a3. Obtain the similarity between all points in each cluster and its cluster center;

[0023] a4. Obtain the similarity of cluster centers between different clusters;

[0024] a5. Adjust and iterate the labels;

[0025] a6. Adjust cluster boundaries by moving selected labels from one cluster to another;

[0026] a7. After the label is moved, recalculate the cluster center of each cluster;

[0027] a8. Repeat a5 to a7 until the preset number of iterations is reached and the impact of cluster boundary adjustment on the total similarity is less than a certain threshold.

[0028] Furthermore, the method of clustering the preprocessed label feature vectors by the K-means algorithm to determine the optimal number of clusters K includes: running the K-means algorithm for different K values, calculating the total square error corresponding to each K value, obtaining a relationship graph between the total square error and K, selecting the point where the total average error begins to decrease as the K value, or for different K values, obtaining the silhouette coefficient of each sample, selecting the value with the highest average silhouette coefficient as the K value, or for each K value, calculating the total average error of the K-means clustering and the total average error of a reference data set, the reference data set is generated by adding random noise to the original data set, then calculating the difference between the two, and selecting the one with the largest difference as the K value.

[0029] Furthermore, the format of the utterance appeal text data needs to be unified when input into each language recognition learning model.

[0030] Furthermore, the prediction results output by each language recognition learning model are statistically scored according to the weight information, and the specific method of obtaining the most matching voiced appeal text data recognition result according to the score includes:

[0031] When the content of the appeal labels output by each language recognition learning model is different, the appeal labels output by each model are weighted, the weight information of each label is counted, and the appeal label with the highest appeal label weight is selected as the recognition result label that best matches the current utterance appeal text data;

[0032] When the contents of the appeal labels output by each language recognition learning model are the same, the appeal labels with the same content output by each model are weighted, the weight information of the appeal labels with the same content is summed up, and then the summed weight information is compared. The appeal label corresponding to the weight information with the highest score is selected as the recognition result label that best matches the current voice appeal text data.

[0033] In a second aspect, the present application also provides a speech processing system based on the LLM model, characterized in that it includes:

[0034] A language recognition learning model acquisition unit is used to acquire multiple language recognition learning models on the same hardware platform to ensure that all language recognition learning models run in a fair resource environment;

[0035] A voiced appeal text data acquisition unit, used to acquire voiced appeal text data, and input the voiced appeal text data into each language recognition learning model;

[0036] A prediction result output unit is used to run all language recognition learning models in parallel, compare and analyze the voiced demand text data with the demand labels in the demand label library of each language recognition learning model, and each language recognition learning model outputs its own prediction result, wherein the prediction result includes the weight information of the prediction result in the demand label library of the language recognition learning model;

[0037] The unit for obtaining the most matching recognition result of the utterance demand text data performs statistical scoring on the prediction results output by each language recognition learning model according to the weight information, and obtains the most matching recognition result of the utterance demand text data according to the score.

[0038] Furthermore, the voiced appeal text data acquisition unit further includes a text data format unification unit, and the text data format unification unit is used to unify the format of the voiced appeal text data input into each language recognition learning model.

[0039] Furthermore, the best matching voiced appeal text data recognition result acquisition unit includes a text data weight calculation and comparison unit, which is used to

[0040] When the content of the appeal labels output by each language recognition learning model is different, the appeal labels output by each model are weighted, the weight information of each label is counted, and the appeal label with the highest appeal label weight is selected as the recognition result label that best matches the current utterance appeal text data;

[0041] When the contents of the appeal labels output by each language recognition learning model are the same, the appeal labels with the same content output by each model are weighted, the weight information of the appeal labels with the same content is summed up, and then the summed weight information is compared. The appeal label corresponding to the weight information with the highest score is selected as the recognition result label that best matches the current voice appeal text data.

[0042] In a third aspect, the present application also provides a computer storage medium, characterized in that: the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above-mentioned method steps.

[0043] In a fourth aspect, the present application further provides an electronic device, characterized in that it comprises a processor and a memory, wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the above-mentioned method steps.

[0044] The beneficial effects brought about by the technical solutions provided by some embodiments of the present application include at least the following: the present application abandons the problem that traditional small models have limited ability to understand long texts and cannot effectively learn more contextual information. Secondly, the language recognition learning model based on multiple large models can have a more thorough understanding of diverse vocabulary and grammatical rules. In addition, it can better handle ambiguity and have a clearer understanding of more contextual information. By introducing language recognition learning models of multiple large models, the accuracy and stability of classification can be improved; the scoring comparison method can fully consider the prediction results of each model, reduce the influence of subjectivity, and improve the objectivity and stability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0047] Figure 2 is a system block diagram of an embodiment of the present invention;

[0048] Figure 3 It is a schematic diagram of the structure of the electronic device of the present invention.

[0049] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings. DETAILED DESCRIPTION

[0050] The technical solution provided by the present invention is described in more detail below in conjunction with the accompanying drawings. The description of the exemplary implementation force is merely illustrative and is by no means intended to limit the present disclosure and its application or use. The present disclosure can be implemented in many different forms and is not limited to the embodiments described herein. These embodiments are provided to make the present disclosure thorough and complete and to fully express the scope of the present disclosure to those skilled in the art. It should be noted that unless otherwise specifically stated, the relative arrangement of the parts and steps, and the composition of the materials described in these embodiments. Numerical expressions and numerical values ​​should be interpreted as being merely exemplary and not as limitations.

[0051] The serial numbers assigned to the components in the present disclosure, such as "first", "second", etc., are only used to distinguish the objects described and do not have any order or technical meaning. The "connection" and "coupling" mentioned in the present invention, unless otherwise specified, include direct and indirect. In the description of the present invention, it should be understood that the directions or positional relationships indicated by the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc. are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the invention.

[0052] In the present disclosure, when a specific device is described as being located between a first device and a second device, there may or may not be an intermediate device between the specific device and the first device or the second device. When a specific device is described as being connected to other devices, the specific device may be directly connected to the other devices without an intermediate device, or may not be directly connected to the other devices but have an intermediate device.

[0053] All terms (including technical terms or scientific terms) used in the present disclosure have the same meanings as those understood by ordinary technicians in the field to which the present disclosure belongs, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant technology, and should not be interpreted in an idealized or extremely formal sense, unless explicitly defined herein.

[0054] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0055] The dimensions of the various parts shown in the drawings are not drawn according to the actual proportional relationship. In the drawings, the same figure marks are attached to common structural elements or structural elements of the same type, and their repeated descriptions are appropriately omitted.

[0056] First, if Figure 1 As shown, the present application provides a method for classifying customer voice demands, including the steps of:

[0057] S1. Obtain multiple language models with language recognition and classification learning functions on the same hardware platform, and ensure that all language models run in a fair resource environment;

[0058] S2, obtaining voiced appeal text data, and inputting the voiced appeal text data into each language model respectively;

[0059] S3, running all language models in parallel, comparing and analyzing the voiced appeal text data with the appeal labels in the language model appeal label library, and predicting each language model outputting its own prediction result, wherein the prediction result includes the weight information of the prediction result in the language model appeal label library;

[0060] S4. Perform statistical scoring on the prediction results output by each language model according to the weight information, and obtain the most matching voice demand text data recognition result according to the score.

[0061] In some embodiments, each of the multiple language models is a language model with deep language learning training function and self-iteration function.

[0062] By adopting multiple large language models (LLM models) with language recognition and classification learning functions, we can make full use of the advantages and characteristics of different models to obtain more comprehensive and accurate text understanding results. Each LLM model may give different classification results when processing the same text, but by obtaining the high and low scores, we can combine the opinions of each model to make the final classification decision. This method can improve the accuracy and stability of classification, reduce the impact of single model bias, and better meet the needs of actual application scenarios.

[0063] In some embodiments, the language model may adopt, but is not limited to, any combination of two or more of the ChatGLM2-6B, baichuan2-13b, and PingAnGPT data training models. These models are natural language processing models based on deep learning, used to understand and generate human language. ChatGLM2-6B is a large pre-trained language model developed by the Natural Language Processing Laboratory of Tsinghua University. It uses 6 billion parameters and can be used for various natural language processing tasks, such as text generation, question-answering systems, machine translation, etc. Baichuan2-13b is a large pre-trained language model launched by Baidu, which uses 13 billion parameters. This model can be used for various natural language processing tasks, such as text generation, question-answering systems, machine translation, etc. PingAnGPT is a large pre-trained language model launched by Ping An Technology, which uses billions of parameters. This model can be used for various natural language processing tasks, such as text generation, question-answering systems, machine translation, etc. These models can all be trained with large amounts of text data. In addition, ChatGLM2-6B is known for its unique encoder-decoder architecture and rich pre-training strategies, while Baichuan2-13B represents the latest progress in large-parameter models, especially its breakthrough in parallel computing optimization. PingAnGPT, with its professional data training in the financial field, has demonstrated a strong grasp of knowledge in specific fields. The training data sets behind each model have different focuses, ranging from general knowledge to in-depth knowledge of specific fields. These differences are directly related to the adaptability and accuracy of the model for specific tasks. Therefore, cross-model comparisons can help understand the impact of data diversity on model output, especially when dealing with complex, professional problems.

[0064] ChatGLM2-6B, baichuan2-13b and PingAnGPT are all large language models based on the Transformer architecture, which are used to handle natural language processing tasks. However, this application is not limited to the above models. Under certain conditions, alternative models of these models can be found. The alternative models can consider the following aspects:

[0065] Choose other large pre-trained language models: There are many other large pre-trained language models on the market, such as GPT-3, BERT, XLNet, etc. These models perform well on natural language processing tasks and can be used as alternatives.

[0066] Use open source models: Many large pre-trained language models are open source and can be found on platforms such as GitHub. For example, Hugging Face's Transformers library provides many pre-trained models, and you can choose the appropriate model for fine-tuning according to your needs.

[0067] Custom trained models: If the existing pre-trained models do not meet your needs, you can consider training a custom model from scratch. This requires a lot of computing resources and time, but can be optimized according to the needs of specific tasks.

[0068] Transfer learning: You can use an existing pre-trained model as a starting point and fine-tune it to suit a specific task through transfer learning. This method can achieve better performance on smaller datasets.

[0069] Ensemble multiple models: You can try to ensemble multiple different pre-trained models together to improve the overall performance. This can be achieved through simple weighted averaging, voting, or other ensemble methods.

[0070] In summary, when choosing an alternative approach, factors such as model performance, computing resources, and ease of use need to be considered to find the solution that best suits a specific task.

[0071] In some embodiments, all of the above language models need to be run on the same hardware platform, such as on the same computer, deployed on the same hardware platform, and initially tuned according to their respective official best practices to ensure that all language models run in a fair resource environment.

[0072] Specifically, to ensure that all language models can run on the same hardware platform, the following issues need to be noted:

[0073] Choose the right hardware platform: First, you need to identify a hardware platform that can meet all the language model computing needs. You can choose a high-performance CPU, GPU, memory, and storage space. Make sure the selected hardware platform has enough scalability to add more models or upgrade existing models in the future.

[0074] Operating system and drivers: An appropriate operating system (such as Linux, Windows, or macOS) installed on the hardware platform should ensure that all necessary drivers are installed so that the hardware devices can function properly.

[0075] Programming language and library requirements: Depending on the language model you want to run, you will need a programming language (such as Python, Java, or C++) and a library (such as TensorFlow, PyTorch, or Scikit-learn). Make sure these packages are compatible with your hardware platform.

[0076] Configuration environment: Each language model requires an independent running environment to ensure that the dependencies between them do not conflict. This can be achieved through a virtual environment (such as Python's virtualenv) or container technology (such as Docker).

[0077] Language model: You can obtain the required language model code from the official website or GitHub repository, and make sure to install and configure it according to the model's documentation.

[0078] When running a model, you can monitor and optimize the hardware, software, and data configuration to identify possible sources of problems. In addition, you can also pay attention to the usage of hardware resources, such as CPU, GPU, and memory usage. If you find a resource bottleneck, you can optimize the model parameters or upgrade the hardware to improve performance through optimization and upgrade methods.

[0079] Documentation and Logging: To ensure reproducibility and maintainability, all steps, configurations, and results can be documented to allow replication and extension of results.

[0080] In some embodiments, the specific method of comparing, analyzing and predicting the uttered appeal text data with the appeal labels in the appeal label library of the language recognition learning model, and each language recognition learning model outputting its own prediction result includes:

[0081] The demand tags in the demand tag library are cleaned, and feature extraction is performed after data cleaning; wherein data cleaning can remove redundant, repeated and irrelevant tags to ensure the quality of the data set. In some embodiments, feature extraction: TF-IDF, word embedding or other suitable methods are used to convert text tags into numerical feature vectors to facilitate subsequent clustering analysis.

[0082] Cluster the feature vectors and obtain preliminary preprocessing, determine the cluster center labels, adjust the cluster boundaries, and then iteratively update the cluster centers and boundaries;

[0083] The weight of each cluster is calculated based on the number of labels in the cluster. The more labels there are in the cluster, the greater the weight of the corresponding cluster center label. Conversely, the smaller the weight is. For example, if a cluster contains 10 samples, the weight of the center label of the cluster is set to 10.

[0084] This weight is the weight of which label to use in the subsequent calculation

[0085] In some embodiments, the preprocessed label feature vectors are clustered using the K-means algorithm to determine the optimal number of clusters K and obtain preliminary clustering results. In terms of cluster center determination, the centroid of each cluster is selected as the representative label of the cluster, that is, the cluster center label.

[0086] In some embodiments, cluster boundary adjustment: a greedy algorithm is used to gradually adjust the cluster boundaries to optimize the clustering results. This step aims to find an optimal solution that can maximize the similarity within a cluster while minimizing the similarity between clusters.

[0087] Iterative update: Continuously iteratively update the cluster centers and boundaries until convergence conditions are reached or the preset stopping criteria are met.

[0088] In some embodiments, clustering the feature vectors and obtaining preliminary clustering results, determining cluster center labels, adjusting cluster boundaries, and then iteratively updating cluster centers and boundaries includes:

[0089] The preprocessed label feature vectors are clustered using the K-means algorithm to determine the optimal number of clusters K and obtain preliminary clustering results;

[0090] The centroid of each cluster is selected as the representative label of the cluster, that is, the cluster center label;

[0091] A greedy algorithm is used to gradually adjust the cluster boundaries to optimize the clustering results.

[0092] In some embodiments, the method of using a greedy algorithm to gradually adjust cluster boundaries to optimize clustering results includes:

[0093] a1. Determine the initial cluster center through K-means algorithm;

[0094] a2. Determine the initialization boundary of each cluster based on the initial clustering results;

[0095] a3. Obtain the similarity between all points in each cluster and its cluster center;

[0096] a4. Obtain the similarity of cluster centers between different clusters;

[0097] a5. Adjust and iterate the labels;

[0098] a6. Adjust cluster boundaries by moving selected labels from one cluster to another;

[0099] a7. After the label is moved, recalculate the cluster center of each cluster;

[0100] a8. Repeat a5 to a7 until the preset number of iterations is reached and the impact of cluster boundary adjustment on the total similarity is less than a certain threshold.

[0101] Specifically, the above steps can be divided into several large blocks according to their functions:

[0102] 1. Initialization:

[0103] -Determine the initial cluster centers, which are usually obtained through the K-means algorithm.

[0104] - Initialize the boundaries of each cluster, possibly based on the initial clustering results.

[0105] 2. Calculate the initial similarity:

[0106] -Calculate the similarity between all points in each cluster and its cluster center. The inverse of the Euclidean distance can be used as the similarity indicator.

[0107] -Calculate the similarity of cluster centers between different clusters.

[0108] 3. Select the adjustment object:

[0109] - In each iteration, a label (or labels) is selected, which may be a point on the cluster boundary or a point that is adjusted to maximize the intra-cluster similarity or minimize the inter-cluster similarity.

[0110] 4. Adjust the border:

[0111] -Move the selected label from one cluster to another, which requires calculating the total similarity change before and after the move.

[0112] - If moving a label increases the total similarity (or decreases the total dissimilarity), move it, otherwise keep it as it is.

[0113] 5. Update cluster center:

[0114] - Recalculate the cluster center of each cluster after the label is moved, because changes in cluster membership may affect the location of the cluster center.

[0115] 6. Iteration:

[0116] - Repeat steps 3 to 5 until the stopping condition is met. The stopping condition can be that the preset number of iterations is reached, the effect of cluster boundary adjustment on the total similarity is less than a certain threshold, or the cluster center changes very little.

[0117] The specific steps are as follows:

[0118] Python

[0119] while (stop condition not reached):

[0120] for each label in all labels on the boundary:

[0121] for each target cluster in other clusters:

[0122] Calculate the similarity change of moving the label to the target cluster

[0123] If similarity increases:

[0124] Move the label to the target cluster

[0125] Update cluster centers

[0126] Calculate the current total similarity

[0127] If the total similarity change is less than the threshold or the maximum number of iterations is reached:

[0128] Stop Iteration

[0129] In this process, "greed" is reflected in the fact that each iteration selects the optimal move in the current situation without considering whether this move will lead to a local optimal solution in future iterations.

[0130] In some embodiments, the method of clustering the preprocessed label feature vectors by the K-means algorithm to determine the optimal number of clusters K includes but is not limited to: running the K-means algorithm for different K values, calculating the total square error corresponding to each K value, obtaining the relationship diagram of the total square error and K, and selecting the point where the total average error decreases as the K value, or obtaining the silhouette coefficient of each sample for different K values, selecting the value with the highest average value of the silhouette coefficient as the K value, or for each K value, calculating the total average error of the K-means clustering and the total average error of a reference data set, the reference data set is generated by adding random noise to the original data set, and then calculating the difference between the two, and selecting the largest difference as the K value. It is also possible to calculate the Davies-Bouldin index for different K values, which is the average of the ratio of the average distance of samples within the cluster to the distance of the nearest cluster between clusters, and select the K value with the smallest Davies-Bouldin index. It is also possible to use the Canopy algorithm to initialize the center point before K-means, which can provide a better estimate of the K value. The Canopy algorithm uses two thresholds to roughly determine the number of clusters. The K value can also be determined by calculating the density distribution of the data. Methods such as DBSCAN can be used to understand the structure of the data first, and then help determine the value of K in K-means. When choosing the value of K, in addition to using the above methods, the following factors should also be considered:

[0131] Business understanding: Sometimes business requirements or domain knowledge can guide the choice of K.

[0132] Data characteristics: The distribution and characteristics of the data itself will also affect the choice of K.

[0133] Computational resources: Higher values ​​of K require more computing resources.

[0134] Typically, determining the optimal K value is an iterative process that may require combining multiple approaches and considering the above factors to make a decision.

[0135] In some embodiments, the appeal labels and their weights in the language model library in each language model are the same. For example, the appeal labels reflect product / service usage feedback, functional experience, price disputes, service preferences, write-off / exchange process feedback, system timeliness, product focus-product compensation information, product focus-insurance liability, incomplete notification, can't find people, complaints, service preferences-need to add, product focus-underwriting conditions, product compensation information, information registration, communication attitude, product focus-others, cumbersome processes, other praises, underwriting conditions and sales misleading, etc. For example, the appeal labels in the language model library based on the communication between operators and customers in an insurance company are as follows:

[0136] “1. Product / service usage feedback (feedback on problems encountered during use after redeeming products / services, such as: car wash not clean, poor quality and damage of products); 2. Functional experience (dissatisfaction with the functions of online platforms such as APP / WeChat due to unusable functions, operation lag, material loss, etc.); 3. Price disputes (due to doubts about the price amount in the current scenario, which is inconsistent with expectations / misunderstanding); 4. Service preferences, personnel timeliness (ignoring, ignoring, replying time is too long or waiting time is too long, no one handles on weekends); 5. Verification / redemption process feedback (feedback on problems encountered in the redemption process of products / services, such as: the coupons sent cannot be found) , don’t know how to change it); 6. System timeliness (mainly refers to the system processing cycle, the process is too long, resulting in slowness and customers can’t wait); 7. Product focus-product compensation information (product claim process, claim method, deductible, claim cycle related consultation); 8. Product focus-insurance liability (consultation on product insurance liability, excluded liability, and whether compensation is given in various situations); 9. Incomplete notification (incomplete explanation of the scope of terms and policy-related standards, or related matters are not fully informed and there are omissions, and the conditions for using the service are not clearly informed, such as the medical card can only be used for 50 yuan per month and is valid for only 180 days); 10. Can’t find anyone (the customer doesn’t know who to contact and can’t find the person to contact ); 11. Complaints, service preferences - need to be added (obvious service preferences appear in the voice, such as: others give 5 gas tickets, you only give 2, touch-up paint is useless and I need to charge, is there any discount for manicure); 12. Product focus - underwriting conditions (consult product underwriting conditions, what conditions can be met for underwriting, age restrictions, medical history, current physical condition, etc.); 13. Product compensation information (product claim process, claim method, deductible, claim cycle related consultation); 14. Information registration (information errors in registration records); 15. Communication attitude (bad attitude, insults, or non-reception in various scenarios, etc.); 16. Product focus - others; 17 , complicated processes (related issues caused by complex processes and multiple operations in the corresponding business processes, such as customers need to download the Good Car Owner APP, Good Life APP, Ping An Pocket Bank, and add corporate WeChat when purchasing insurance); 18. Other praise (praise for insurance / comprehensive financial products, purchase process, activity experience, and platform experience); 19. Underwriting conditions (consult product underwriting conditions, what conditions must be met for underwriting, age restrictions, medical history, current physical condition, etc.); 20. Misleading sales (concealment during the sales process, confusion of product information, hidden bundling misleading, exaggeration and fictitious misleading); 21. Insurance liability (consult product insurance liability, excluded liability, and whether compensation will be given in various situations).

[0137] The above-mentioned appeal tags have corresponding weight information configured. For example, 1. The weight of "Feedback on the use of goods / services" is 1; 2. The weight of "Function experience tag" is 2; 3. The weight of "Price dispute" is 3; 4. The weight of "Service preference, personnel timeliness" is 4; 5. The weight of "Feedback on the verification / exchange process" is 5; 6. The weight of "System timeliness" is 6; 7. The weight of "Product concern - Product compensation information" is 7; 8. The weight of "Product concern - Insurance liability" is 8; 9. The weight of "Incomplete information" is 9; 10. The weight of "Unable to find someone" is 10; 11. The weight of "Complaint, Service preference - To be added" is 11; 12. The weight of "Product concern - Underwriting conditions" is 12; 13. The weight of "Product compensation information" is 13; 14. The weight of "Information registration" is 14; 15. The weight of "Communication attitude" is 15; 16. The weight of "Product concern - Others" is 16; 17. The weight of "Complicated process" is 17; 18. The weight of "Other praises" is 18; 19. The weight of "Underwriting conditions" is 19; 20. The weight of "Sales misrepresentation" is 20; 21. The weight of "Insurance liability" is 21. The weight information of each appeal tag corresponds one-to-one with the content of the appeal tag. Each tag represents a specific customer concern or problem type, and the weight indicates the importance or priority of the problem. The larger the weight value, the higher the priority that should be given to this problem when dealing with it.

[0138] In some embodiments, after receiving the voice appeal text data, the hardware platform needs to perform noise reduction processing on the text data, such as cleaning and annotation, with the aim of removing unnecessary noises, repeated sounds, etc., so as to make the output not affected by other factors and make the output more accurate.

[0139] Specifically, the noise reduction processing is mainly carried out on the following content for noise reduction:

[0140] Removing stop words: Stop words refer to words that frequently appear in the text but contribute little to the meaning of the text, such as "de (的), shi (是), zai (在), he (和)", etc.

[0141] Removing punctuation marks and special characters: These characters usually contribute little to the meaning of the text and can be removed to reduce noise. Removing punctuation marks and special characters is an important step in text preprocessing, especially when performing natural language processing (NLP) tasks. These characters usually include:

[0142] Punctuation marks: Such as full stop (。), comma (,), semicolon (;), colon (:), question mark (?), exclamation mark (!), quotation marks (“”‘’), parentheses (()【】{}), dash (——), ellipsis (……), etc.

[0143] Special characters: Such as tab (\t), line break (), carriage return (\r), space, and other non-printable characters.

[0144] Characters other than numbers and letters: such as currency symbols ($￡￥), percent signs (%), copyright symbols Trademark symbol ( TM ), emoticons, arrow symbols, etc.

[0145] Remove sparse words: Remove words that occur less frequently in the text and may contribute less to the meaning of the text.

[0146] Use topic models to identify the main topics in the text and remove noise information that is not related to the topic.

[0147] Use automatic summarization technology to extract key information from the text, remove redundant and unimportant content, etc.

[0148] In addition, in some embodiments, the format of the voiced appeal text data needs to be unified when input into each language recognition learning model to eliminate deviations caused by differences in data processing.

[0149] There are several important reasons for unifying the format of voice appeal text data:

[0150] Improve processing efficiency: A unified format can make data processing more efficient because the model does not need to write specific parsing code for each different format.

[0151] Reduce errors: When data is not in the same format, it is more likely to cause parsing errors or misunderstandings. A unified format can reduce this risk.

[0152] Improved compatibility: If the system needs to be extended to support new language identification learning models, a unified format will make this process simpler and more straightforward.

[0153] Facilitates data analysis: When conducting data analysis, a unified format can simplify the analysis process and make it easier to compare and integrate the results from different models.

[0154] Improve model accuracy: Some models may have specific requirements for the format of input data, and failure to meet these requirements may result in reduced recognition accuracy. By unifying the format, you can ensure that each model receives data suitable for its best performance.

[0155] In general, the unified unit of text data format plays a vital role in processing complex systems with multiple languages ​​and models, ensuring data consistency and system robustness.

[0156] When the data format is unified, the following methods can be used:

[0157] Determine the target format: First, determine what format you want to convert the data to. You can choose CSV, JSON, XML, Excel, etc. depending on the actual situation.

[0158] Data cleaning: Check the data for missing values, duplicate values, outliers, or inconsistent data to ensure data quality and consistency.

[0159] Data transformation: Use tools or programming languages ​​(such as Python, R, etc.) to convert data into the target format. You can use ready-made libraries or tools.

[0160] Data Validation: After conversion, you need to verify that the data has been correctly converted to the target format and that no data has been lost or converted incorrectly.

[0161] Documentation: Record the conversion process and any special processing steps so that the process can be replicated in the future if needed.

[0162] Automation: When converting, use an automated process, such as using a scheduled task or script to automatically execute.

[0163] The above are general steps. The specific operations may be set differently depending on the type and complexity of the data and the target format.

[0164] After that, all models are run in parallel under the same conditions to predict the same voice appeal text data. For example, the following voice appeal text data: ["Hello, ok, it feels very small, ah, you said hello, hello, what, what, I didn't hear it clearly, can you call the insured because I am just the driver, I am the driver, I am not the insurance, I am responsible for driving the car, but the insurance insured is Mr. Li, you call him and say ok, ok, ok, ok, ok, ok, um"]. Each model outputs different prediction results, for example:

[0165] Output 1:

[0166] The output of the ChatGLM2-6B model is: product focus - underwriting conditions (the corresponding weight is 12);

[0167] The output of the baichuan2-13b model is: product focus - other (the corresponding weight is 16);

[0168] The output of the PingAnGPT model is: product focus-product compensation information (weight is 7).

[0169] For example, output result 2:

[0170] The output of the ChatGLM2-6B model is: product focus - underwriting conditions (the corresponding weight is 12);

[0171] The output of the baichuan2-13b model is: product focus-product compensation information (weight is 7);

[0172] The output of the PingAnGPT model is: product focus-product compensation information (weight is 7).

[0173] After obtaining the appeal labels output by each model, the prediction results output by each language recognition learning model are statistically scored according to the weight information. The specific method for obtaining the most matching voice appeal text data recognition result according to the score includes:

[0174] When the content of the appeal labels output by each language recognition learning model is different, the appeal labels output by each model are weighted, the weight information of each label is counted, and the appeal label with the highest appeal label weight is selected as the recognition result label that best matches the current utterance appeal text data;

[0175] For example, in the above output result 1, the voice appeal labels output by each model are different, and the corresponding weights are also different. Then the output result with the highest weight score is selected as the result that best matches the text data of this voice appeal, that is, product focus-underwriting conditions (the corresponding weight is 12).

[0176] When the contents of the appeal labels output by each language recognition learning model are the same, the appeal labels with the same content output by each model are weighted, the weight information of the appeal labels with the same content is summed up, and then the summed weight information is compared. The appeal label corresponding to the weight information with the highest score is selected as the recognition result label that best matches the current voice appeal text data.

[0177] Refer to the above output result 2, where the appeal label output by the ChatGLM2-6B model is different from that of the baichuan2-13b model and the PingAnGPT model, while the prediction results output by the baichuan2-13b model and the PingAnGPT model are the same. The final output result should first calculate the sum of the weights of the output of the same prediction results. In this result, the sum of the weights of the results output by the baichuan2-13b model and the PingAnGPT model is 7+7=14, and the weight of the ChatGLM2-6B model is 12, which is less than 14. Therefore, the product focus-product compensation information should be selected as the output result that best matches the voice appeal label for this time.

[0178] Of course, in some embodiments, the language model selects more than the above three types. When more than three types are selected, the method is the same as above. First, compare the weights of the demand labels output by each language model. If they are different, select the one with the highest weight as the output result of this time. When there are the same output demand labels, first calculate the sum of the weights of the same labels, and then compare the score of the total with the weights of the labels output by other single models and the sum of the scores of other identical labels, and select the one with the highest score as the output result of this time.

[0179] This application abandons the problem that traditional small models have limited ability to understand long texts and cannot effectively learn more contextual information. Secondly, the language recognition learning model based on multiple large models can have a more thorough understanding of diverse vocabulary and grammatical rules. In addition, it can better handle ambiguity and have a clearer understanding of more contextual information. By introducing language recognition learning models of multiple large models, the accuracy and stability of classification can be improved; the scoring comparison method can fully consider the prediction results of each model, reduce the influence of subjectivity, and improve the objectivity and stability of the prediction.

[0180] like Figure 2 As shown, the present application also provides a customer voice demand classification system, including:

[0181] A language recognition learning model acquisition unit is used to acquire multiple language recognition learning models on the same hardware platform to ensure that all language recognition learning models run in a fair resource environment;

[0182] A voiced appeal text data acquisition unit, used to acquire voiced appeal text data, and input the voiced appeal text data into each language recognition learning model;

[0183] A prediction result output unit is used to run all language recognition learning models in parallel, compare and analyze the voiced demand text data with the demand labels in the demand label library of each language recognition learning model, and each language recognition learning model outputs its own prediction result, wherein the prediction result includes the weight information of the prediction result in the demand label library of the language recognition learning model;

[0184] The unit for obtaining the most matching recognition result of the utterance demand text data performs statistical scoring on the prediction results output by each language recognition learning model according to the weight information, and obtains the most matching recognition result of the utterance demand text data according to the score.

[0185] In some embodiments, each of the multiple language recognition classification learning models acquired by the language recognition learning model acquisition unit is a language model with deep language learning training function and self-iteration function.

[0186] In some embodiments, the language recognition learning model acquisition unit is used to acquire multiple language recognition classification learning models including ChatGLM2-6B, baichuan2-13b, and PingAnGPT data training models.

[0187] In some embodiments, the language recognition learning model acquisition unit may be a component for acquiring and managing multiple language recognition learning models. Its main function is to ensure that all language recognition learning models run on the same hardware platform and can be trained and inferred in a fair resource environment.

[0188] The language recognition learning model acquisition unit can acquire multiple language recognition learning models: acquire the required language recognition learning models from different data sources or model libraries for training and reasoning on the hardware platform.

[0189] In some embodiments, the language recognition learning model acquisition unit should ensure fair resource allocation: allocate fair resources to each language recognition learning model on the hardware platform, such as computing power, memory, and storage space. This can be achieved through a dynamic resource scheduling algorithm to ensure that each model can obtain sufficient resources for training and reasoning.

[0190] In some embodiments, the language recognition learning model acquisition unit can also monitor the resource usage of each language recognition learning model in real time, such as CPU occupancy, memory usage, etc. According to the monitoring results, the resource allocation is adjusted to ensure that all models can run in a fair resource environment.

[0191] In some embodiments, the language recognition learning model acquisition unit can also monitor the running status of each language recognition learning model, and promptly discover and handle possible faults, such as model crash, insufficient resources, etc. When a fault occurs, it can try to automatically restore the model or report the fault information to the administrator for processing.

[0192] Through the above functions, the language recognition learning model acquisition unit can ensure that multiple language recognition learning models on the same hardware platform can run in a fair resource environment, thereby improving the training and reasoning performance of the model.

[0193] In some embodiments, the voiced appeal text data acquisition unit includes a noise reduction unit, and the noise reduction unit is used to perform noise reduction processing on the acquired voiced appeal text data.

[0194] In some embodiments, the voiced appeal text data acquisition unit further includes a text data format unification unit, and the text data format unification unit is used to unify the format of the voiced appeal text data input into each language recognition learning model.

[0195] In some embodiments, the best matching vocal appeal text data recognition result acquisition unit includes a text data weight calculation and comparison unit, and the text data weight calculation and comparison unit is used to:

[0196] When the content of the appeal labels output by each language recognition learning model is different, the appeal labels output by each model are weighted, the weight information of each label is counted, and the appeal label with the highest appeal label weight is selected as the recognition result label that best matches the current utterance appeal text data;

[0197] When the contents of the appeal labels output by each language recognition learning model are the same, the appeal labels with the same content output by each model are weighted, the weight information of the appeal labels with the same content is summed up, and then the summed weight information is compared. The appeal label corresponding to the weight information with the highest score is selected as the recognition result label that best matches the current voice appeal text data.

[0198] The present application system is used to implement the present application method. The specific working principle can be found in the present application method, which will not be repeated here.

[0199] This application is designed to achieve more accurate and natural communication. By adopting an experimental approach with multiple LLM models, the advantages and characteristics of different models are fully utilized to obtain more comprehensive and accurate text understanding results. Each LLM model may give different classification results when processing the same text, but through voting, the opinions of each model can be combined to reach a final classification decision. This approach can improve the accuracy and stability of classification, reduce the impact of single model bias, and thus better meet the needs of actual application scenarios.

[0200] The present application also provides a computer storage medium, which stores multiple instructions, which are suitable for being loaded by a processor and executing the above method steps. Computer storage medium is a physical device for storing data, such as a hard disk, an optical disk, a USB flash drive, etc. These storage media can store various types of data, including text, images, audio, video, and program code. The storage medium stores some programs or scripts, which contain a series of instructions suitable for being loaded by a processor and executing the above method steps, and these instructions can be read, parsed, and executed by the computer's processor.

[0201] For example, an executable file (such as an .exe file) stored on a USB flash drive contains a series of instructions. When this file is copied to a computer and run, the computer's processor will operate according to the instructions in the file to achieve a specific function or task. The specific execution process can be found in the detailed description of the system and method, which will not be repeated here.

[0202] The present application also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the above method steps.

[0203] See also Figure 3 , is a schematic diagram of the main hardware structure of the electronic device provided in the embodiment of the present application. Figure 3 As shown,

[0204] The customer voice demand classification device 300 may include: at least one processor 301 , at least one network interface 304 , a user interface 303 , a memory 305 , and at least one communication bus 302 .

[0205] The communication bus 302 is used to realize the connection and communication between these components.

[0206] The user interface 303 may include a touch screen and a camera.

[0207] The network interface 304 may optionally include but is not limited to a standard wired interface, a wireless interface (such as a WI-FI interface, Bluetooth, a traffic card, etc.).

[0208] Among them, the processor 301 may include one or more processing cores. The processor 301 uses various interfaces and lines to connect the various parts of the entire multi-tenant mode data acquisition system 300, and executes various functions and processes data of the multi-tenant mode data acquisition system 300 by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Optionally, the processor 301 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 301 can integrate one or more combinations of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to handle wireless communication. It can be understood that the above-mentioned modem may not be integrated into the processor 301, and it can be implemented by a single chip.

[0209] Among them, the memory 305 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read-Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may optionally be at least one storage device located away from the aforementioned processor 301. As a computer storage medium, the memory 305 may include an operating system, a network communication module, a user interface module and an application. In Figure 3 In the system 300 shown, the user interface 303 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 301 can be used to call the application stored in the memory 305 and specifically execute the following steps: Figure 1 The specific process can be referred to Figure 1 As shown, no further description is given here.

[0210] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium along with the main application program. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be cloud storage, a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0211] The features or feature combinations mentioned above in the specification, drawings and claims, as long as they are meaningful within the scope of the present invention and do not contradict each other, can be used in any combination or alone. The advantages and features described for the system provided by the present invention are applicable to the method and computer-readable storage medium provided by the present invention in a corresponding manner, and vice versa.

[0212] In summary, this application eliminates the problem that traditional small models have limited ability to understand long texts and cannot effectively learn more contextual information. Secondly, the language recognition learning model based on multiple large models can have a more thorough understanding of diverse vocabulary and grammatical rules. In addition, it can better handle ambiguity and have a clearer understanding of more contextual information. By introducing language recognition learning models of multiple large models, the accuracy and stability of classification can be improved; the scoring comparison method can fully consider the prediction results of each model, reduce the influence of subjectivity, and improve the objectivity and stability of the prediction.

[0213] It should also be noted that the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, product or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, product or apparatus. In the absence of further restrictions, the elements defined by the sentence "comprises one..." do not exclude the presence of other identical elements in the process, method, product or apparatus that includes the elements.

[0214] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A speech processing method based on the LLM model, characterized in that: include: Obtain multiple language recognition learning models on the same hardware platform to ensure that all language recognition learning models run in a fair resource environment; Acquire voiced appeal text data, and input the voiced appeal text data into each language recognition learning model respectively; All language recognition learning models are run in parallel, and the utterance demand text data is compared, analyzed and predicted with the demand labels in the demand label library of the language recognition learning model, and each language recognition learning model outputs its own prediction result, wherein the prediction result includes the weight information of the prediction result in the demand label library of the language recognition learning model; The prediction results output by each language recognition learning model are statistically scored based on the weight information, and the most matching voice demand text data recognition result is obtained based on the score.

2. The method according to claim 1, characterized in that The specific method of comparing, analyzing and predicting the uttered appeal text data with the appeal labels in the appeal label library of the language recognition learning model, and each language recognition learning model outputting its own prediction result includes: Perform data cleaning on the demand labels in the demand label library, and perform feature extraction after data cleaning; Cluster the feature vectors and obtain preliminary preprocessing, determine the cluster center labels, adjust the cluster boundaries, and then iteratively update the cluster centers and boundaries; The weight of each cluster is calculated based on the number of labels in the cluster. The more labels there are in the cluster, the greater the weight of the corresponding cluster center label, and vice versa.

3. The method according to claim 2, characterized in that: The step of clustering the feature vectors and obtaining preliminary clustering results, determining cluster center labels, adjusting cluster boundaries, and then iteratively updating cluster centers and boundaries includes: The preprocessed label feature vectors are clustered using the K-means algorithm to determine the optimal number of clusters K and obtain preliminary clustering results; The centroid of each cluster is selected as the representative label of the cluster, that is, the cluster center label; A greedy algorithm is used to gradually adjust the cluster boundaries to optimize the clustering results.

4. The method according to claim 3, characterized in that The method of using a greedy algorithm to gradually adjust the clustering boundaries to optimize the clustering results includes: a1. Determine the initial cluster center through K-means algorithm; a2. Determine the initialization boundary of each cluster based on the initial clustering results; a3. Obtain the similarity between all points in each cluster and its cluster center; a4. Obtain the similarity of cluster centers between different clusters; a5. Adjust and iterate the labels; a6. Adjust cluster boundaries by moving selected labels from one cluster to another; a7. After the label is moved, recalculate the cluster center of each cluster; a8. Repeat a5 to a7 until the preset number of iterations is reached and the impact of cluster boundary adjustment on the total similarity is less than a certain threshold.

5. The method according to claim 3, characterized in that: The method for clustering the preprocessed label feature vectors by the K-means algorithm to determine the optimal clustering number K includes: running the K-means algorithm for different K values, calculating the total square error corresponding to each K value, obtaining a relationship diagram between the total square error and K, selecting a point where the total average error decreases as the K value, or obtaining a silhouette coefficient for each sample for different K values, selecting a value with the highest silhouette coefficient average as the K value, or calculating the total average error of the K-means clustering and the total average error of a reference data set for each K value, wherein the reference data set is generated by adding random noise to the original data set, then calculating the difference between the two, and selecting the one with the largest difference as the K value.

6. The method according to claim 1, characterized in that: The format of the utterance request text data needs to be unified when input into each language recognition learning model.

7. The method according to claim 1, characterized in that The prediction results output by each language recognition learning model are statistically scored according to the weight information, and the specific method of obtaining the most matching voice appeal text data recognition result according to the score includes: When the content of the appeal labels output by each language recognition learning model is different, the appeal labels output by each model are weighted, the weight information of each label is counted, and the appeal label with the highest appeal label weight is selected as the recognition result label that best matches the current utterance appeal text data; When the contents of the appeal labels output by each language recognition learning model are the same, the appeal labels with the same content output by each model are weighted, the weight information of the appeal labels with the same content is summed up, and then the summed weight information is compared. The appeal label corresponding to the weight information with the highest score is selected as the recognition result label that best matches the current voice appeal text data.

8. A speech processing system based on the LLM model, characterized in that: include: A language recognition learning model acquisition unit is used to acquire multiple language recognition learning models on the same hardware platform to ensure that all language recognition learning models run in a fair resource environment; A voiced appeal text data acquisition unit, used to acquire voiced appeal text data, and input the voiced appeal text data into each language recognition learning model; A prediction result output unit is used to run all language recognition learning models in parallel, compare and analyze the voiced demand text data with the demand labels in the demand label library of each language recognition learning model, and each language recognition learning model outputs its own prediction result, wherein the prediction result includes the weight information of the prediction result in the demand label library of the language recognition learning model; The unit for obtaining the most matching voiced demand text data recognition result performs statistical scoring on the prediction results output by each language recognition learning model according to the weight information, and obtains the most matching voiced demand text data recognition result according to the score.

9. A computer storage medium, characterized in that: The computer storage medium stores a plurality of instructions, which are suitable for being loaded by a processor and executing the method steps according to any one of claims 1-5.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps according to any one of claims 1 to 5.

Images