Intelligent agent allocation method and system based on customer dialect recognition

Abstract

The present invention relates to an intelligent agent allocation method and system based on customer dialect recognition, and relates to the technical field of dialect recognition. The present application pre-establishes an agent dialect capability database: the agent dialect capability database records the dialect types, dialect proficiency information, business scope, service quality evaluation and idle status of each agent; collects the target voice of the customer; performs noise reduction and frame processing on the target voice to obtain a target voice frame, and stacks several continuous target voice frames to form an input S k , input S k The target dialect type is input into the pre-trained dialect recognition model for dialect recognition; the target agent is automatically assigned to the customer by comprehensively considering the dialect matching degree, dialect proficiency information, agent's idle status, and business matching degree; after the service is completed, the customer's evaluation feedback on the target agent is obtained; and the relevant information in the agent dialect capability database is dynamically updated based on the evaluation feedback.

Description

Technical Field

[0001] The present invention relates to the technical field of agent allocation based on dialect recognition, and in particular to an intelligent agent allocation method and system based on customer dialect recognition. Background Art

[0002] At present, insurance agents need to communicate effectively with customers in order to provide them with better services. In the process of insurance agency services, they often encounter customers who communicate in dialects. The differences between dialects and Mandarin are often large. At present, in the process of insurance agency services, especially telephone services, insurance agents who do not understand the customer's dialect often cannot understand the customer's intentions well.

[0003] Dialect recognition refers to identifying dialect categories based on speech. x-vector is a method that uses deep neural networks to extract features from speech signals. It can map high-dimensional speech features to a vector of fixed length to represent speech information. Unlike traditional speech feature extraction methods, x-vector is obtained through deep learning model training and can better express the nonlinear characteristics of speech signals. The x-vector architecture pays more attention to time series information. From the perspective of channel enhancement, researchers have proposed an x-vector architecture that emphasizes channel attention, propagation and aggregation (Emphasized Channel Attention, Propagation and Aggregation TDNN, referred to as ECAPA-TDNN). ECAPA-TDNN has made significant progress in speaker recognition. When migrating ECAPA-TDNN to the dialect recognition task, the original ECAPA-TDNN architecture has problems in dialect recognition, such as insufficient receptive field, inflexible context information extraction, and insufficient attention to local channel information. As a result, ECAPA-TDNN is not as effective in the dialect recognition task as in the speaker recognition field. Summary of the invention

[0004] In order to solve the above technical problems or at least partially solve the above technical problems, the present invention provides an intelligent agent allocation method and system based on customer dialect recognition.

[0005] In a first aspect, the present invention provides an intelligent agent allocation method based on customer dialect recognition, comprising: pre-establishing an agent dialect capability database: the agent dialect capability database records the type of dialect mastered by each agent, dialect proficiency information, business scope, service quality evaluation and idle status;

[0006] Collect the target voice of the customer when communicating with the AI ​​service robot;

[0007] The target speech is subjected to noise reduction, and the target speech is scanned and framed using a first sliding window with a time width of t according to a set first step length to obtain a plurality of target speech frames, wherein the first step length is less than the time width t; and a frame width value of C is used. 1 The second sliding window scans the target speech frame according to the set second step length to obtain each group of input S k , where the second step length is less than the frame width value C 1 , input S k Inputting the target speech into a pre-trained dialect recognition model for dialect recognition, wherein the dialect recognition model outputs a target dialect type corresponding to the target speech;

[0008] Based on the target dialect type used by the customer and the agent's dialect ability, the target agent is automatically assigned to the customer by comprehensively considering the dialect matching, dialect proficiency information, agent's idle status, and business matching. After the customer service is completed, the customer's evaluation feedback on the target agent is obtained. The content of the evaluation feedback includes service attitude evaluation, oral evaluation, comprehension ability evaluation and business professionalism evaluation; the target agent information in the agent dialect ability database is dynamically updated according to the customer's evaluation feedback.

[0009] Furthermore, the target agent information in the agent dialect capability database is dynamically updated according to the customer's evaluation feedback, including:

[0010] Update the dialect proficiency information of the target agent; the dialect proficiency information includes two priority indicators: able to listen and speak and able to listen only. The priority of being able to listen and speak is higher than that of being able to listen only, and a more refined scoring quantification is performed for being able to listen and speak and being able to listen only: if the target agent is the type who can listen and speak, the target agent's dialect speaking ability and dialect comprehension ability are quantified, and the weighted sum is used to obtain a comprehensive quantitative score of being able to listen and speak; if the target agent is the type who can only listen, the target agent's dialect comprehension ability is quantified; among them, the target agent's dialect speaking ability is quantified by taking the weighted sum of the company's oral assessment score for the target agent and the customer's oral evaluation of the target agent, and the agent's dialect comprehension ability is quantified by taking the weighted sum of the company's dialect listening test score for the target agent and the customer's evaluation of the target agent's comprehension ability;

[0011] Update the target agent's service quality evaluation; weighted sum the customer's evaluation of the target agent's business professionalism and service attitude to obtain the service quality score of this service, and calculate the overall service quality average score as the service quality evaluation.

[0012] Furthermore, the dialect recognition model includes: kOne-dimensional convolution layer, ReLU activation function layer and batch normalization layer are used to map the channel dimension; the batch normalized features are divided into two parts according to the channel to obtain the features and Features feature It is input into the local feature extraction network to extract local features. b S k The local features are input into the global feature extraction network to extract global features, and the local features and global features are concatenated and input into the classification network. The classification network includes: a linear layer, an ASP module, a batch normalization layer, a fully connected layer, and a Softmax layer.

[0013] Furthermore, the local feature extraction network includes: three groups of stacked multi-scale channel adaptive residual modules; the global feature extraction network includes: three groups of stacked attention weighted global filtering modules;

[0014] The outputs of the first two groups of multi-scale channel adaptive residual modules are divided into three branches, one of which is used as part of the local feature, one is used to be weighted summed with the output of the attention weighted global filter module at the same level and then input into the next group of multi-scale channel adaptive residual modules, and the last one is used to be weighted summed with the output of the attention weighted global filter module at the same level and then input into the next group of attention weighted global filter modules; the output of the last group of multi-scale channel adaptive residual modules is used as part of the local feature;

[0015] The outputs of the first two groups of attention-weighted global filter modules are divided into three branches, one of which is used as part of the global feature, one is used to weighted sum the outputs of the multi-scale channel adaptive residual modules at the same level and then input into the next group of multi-scale channel adaptive residual modules, and the last one is used to weighted sum the outputs of the multi-scale channel adaptive residual modules at the same level and then input into the next group of attention-weighted global filter modules; the output of the last group of attention-weighted global filter modules is used as part of the global feature.

[0016] Furthermore, the multi-scale channel adaptive residual module of the i-th group converts the input features After being processed by the linear layer, ReLU activation function and batch normalization layer, it is divided into two branches. One branch is processed by the SE module to extract channel attention, and the other branch is divided into N groups according to the channel to obtain sub-features:

[0017]

[0018] Sub-features are iteratively processed through N-1 multi-scale convolutional SE modules, ReLU activation functions, and batch normalization layers:

[0019]

[0020] When n>1, Finally, when n=N-1,

[0021] in, They represent the results of the 2nd, nth, n+1th, N-1th, and Nth multi-scale convolutional SE modules after being processed by the ReLU activation function and the batch normalization layer. BN() represents batch normalization, and MSC-SE 1 (), MSC-SE n ()MSC-SE (N-1) () represents the 1st, nth and N-1th multi-scale convolutional SE modules;

[0022] Will After splicing, the linear layer, ReLU activation function and batch normalization layer are processed and the weighted channel attention is applied to obtain the output of the i-th group of multi-scale channel adaptive residual modules.

[0023] Furthermore, the multi-scale convolution SE module includes four parallel convolutional layers with convolution kernel sizes of 1, 3, 5, and 7. The number of output channels of each convolutional layer is one-fourth of the number of input channels. The output of each convolutional layer is weighted with its own channel weight extracted by the SE module and then spliced ​​in the channel dimension. The spliced ​​result is processed by a linear layer, a ReLU activation function, and a batch normalization layer as the output.

[0024] Furthermore, the attention weighted global filtering module of the i-th group takes the input features After being processed by the linear layer, ReLU activation function and batch normalization layer, it is divided into two branches.

[0025] Input one of the branches into the global average pooling layer, the fully connected layer, the ReLU activation function layer, the fully connected layer and the Softmax layer to obtain the attention weights corresponding to several global filters, and add the learnable global filter and the corresponding attention weights to obtain the weighted global filter;

[0026] Perform one-dimensional Fourier transform on each channel of the other branch to obtain the corresponding frequency band spectrum, combine the frequency band spectrum into a spectrum, divide the spectrum equally according to the time dimension and take the first half, perform Hadamard product on the first half of the spectrum and the weighted global filter, and then perform inverse Fourier transform. The result of the inverse Fourier transform is processed by the linear layer, ReLU activation function and batch normalization layer as the output.

[0027] Furthermore, Mandarin and various dialects are collected in a balanced manner, and expanded by speed changing and adding noise to the speech, and the dialect labels are marked to obtain dialect recognition data. The dialect recognition data is divided into a training set and a validation set, and the training set is used to train the dialect recognition model. The training uses the cross entropy loss between the real dialect label and the predicted dialect type as the loss function, selects the stochastic gradient descent method, and uses the ADAM optimizer. After the training dialect recognition model converges, the validation set is used to verify whether the dialect recognition model is overfitting.

[0028] In a second aspect, the present invention provides an intelligent agent allocation device based on customer dialect recognition, comprising: at least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the intelligent agent allocation method based on customer dialect recognition is implemented.

[0029] In a third aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the intelligent agent assignment method based on customer dialect recognition is implemented.

[0030] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0031] This patent establishes a dynamic agent dialect ability database, recording the dialect types, dialect proficiency information, business scope, service quality evaluation and idle status of each agent; through the dialect recognition model, the customer dialect is recognized in real time, and the target agent is automatically assigned to the customer based on the dialect matching, dialect proficiency information, agent idle status and business matching; the accuracy and efficiency of the assignment are greatly improved, and the optimal agent assignment is achieved. Assigning agents with corresponding dialect skills to customers can ensure smooth communication between customers and agents and improve customer satisfaction.

[0032] The multi-scale convolution SE module of the multi-scale channel adaptive residual module obtains features of different local scales through convolution kernels of different scales, and then splices them after attention weighting, so as to more richly extract the local features of speech and reflect the short-term local contextual characteristics of the dialect. The attention-weighted global filtering module obtains multiple contextual connections in the long-term range of speech through multiple groups of weighted global filters, which is conducive to better modeling the long-term global contextual characteristics of speech. In summary, the multi-scale channel adaptive residual module of the local feature extraction network in the dialect recognition model and the attention-weighted global filtering module of the global feature extraction network cooperate with each other to extract the local and global contextual associations of speech to perform dialect classification, and the classification is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0035] Figure 1 A flowchart of an intelligent agent allocation method based on customer dialect recognition provided by an embodiment of the present invention;

[0036] Figure 2 An architectural diagram of a dialect recognition model provided by an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of a multi-scale channel adaptive residual module provided by an embodiment of the present invention;

[0038] Figure 4 A schematic diagram of a multi-scale convolution SE module provided in an embodiment of the present invention;

[0039] Figure 5 A schematic diagram of an attention-weighted global filtering module provided in an embodiment of the present invention;

[0040] Figure 6 A schematic diagram of an apparatus for allocating intelligent agents based on customer dialect recognition provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including 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, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0043] Example 1

[0044] like Figure 1 As shown, the technology of the present invention implements an intelligent agent allocation method based on customer dialect recognition, including:

[0045] An agent dialect ability database is pre-established: the agent dialect ability database records the types of dialects mastered by each agent, dialect proficiency information, business scope, service quality evaluation and idle status.

[0046] During the specific implementation process, the agent dialect ability database records the agent number, agent name, types of dialects mastered by the agent, dialect proficiency information, agent business scope, agent service quality evaluation and idle status.

[0047] The setting parameters in the agent dialect capability database support dynamic updating based on customer evaluation feedback.

[0048] During the specific implementation process, after serving the customer, we obtain the customer's evaluation feedback on the target agent. The content of the evaluation feedback includes service attitude evaluation, oral evaluation, comprehension ability evaluation and business professionalism evaluation.

[0049] Dynamically update the target agent information in the agent dialect capability database based on customer evaluation feedback, including:

[0050] Update the target agent's dialect proficiency information. The dialect proficiency information includes two priority indicators: able to listen and speak and able to listen only. The priority of being able to listen and speak is higher than that of being able to listen only, and a more refined scoring quantification is performed for being able to listen and speak and being able to listen only: if the target agent is the type who can listen and speak, the target agent's dialect speaking ability and dialect comprehension ability are quantified, and the weighted sum is used to obtain a comprehensive quantitative score for being able to listen and speak; if the target agent is the type who can only listen, the target agent's dialect comprehension ability is quantified. Among them, the target agent's dialect speaking ability is quantified by taking the weighted sum of the company's oral assessment score for the target agent and the customer's oral evaluation of the target agent, and the agent's dialect comprehension ability is quantified by taking the weighted sum of the company's dialect listening test score for the target agent and the customer's evaluation of the target agent's comprehension ability.

[0051] Update the target agent's service quality evaluation. The customer's evaluation of the target agent's professionalism and service attitude is weighted and summed to obtain the service quality score of this service, and the overall service quality average score is used as the service quality evaluation.

[0052] Collect the target voice of the customer during the communication process with the AI ​​service robot. The AI ​​service robot is used to provide simple services to customers.

[0053] After the target speech is subjected to noise reduction and frame division processing, a number of target speech frames with a time length of t per frame are obtained. m=1,2……M; stack several consecutive target speech frames as input S k , Among them C 1 is the number of target speech frames stacked in S, k is the sequence number of the input, and R represents the real number domain.

[0054] In the specific implementation process, the first sliding window with a time width of t is used to scan the target speech according to the set first step length to obtain a number of target speech frames, wherein the first step length is less than the time width t; and the frame width value is C 1 The second sliding window scans the target speech frame according to the set second step length to obtain each group of input S k , where the second step length is less than the frame width value C 1 .

[0055] Enter S k The target dialect is input into a pre-trained dialect recognition model for dialect recognition, and the dialect recognition model outputs a target dialect type corresponding to the target speech.

[0056] In the specific implementation process, Figure 2 As shown, the dialect recognition model includes:

[0057] For input S kOne-dimensional convolution layer, ReLU activation function layer and batch normalization layer for channel dimension mapping, where the input channel of the one-dimensional convolution layer is C 1 , the output channel is C, and the batch normalized features are divided into two parts according to the channel to obtain the features and Features feature It is input into the local feature extraction network to extract local features. It is input into the global feature extraction network to extract global features.

[0058] Specifically, the local feature extraction network includes: three groups of stacked multi-scale channel adaptive residual modules. The global feature extraction network includes: three groups of stacked attention weighted global filter modules. The outputs of the first two groups of multi-scale channel adaptive residual modules are divided into three branches, one of which is used as part of the final local feature, one is used to be weighted summed with the output of the attention weighted global filter module at the same level and then input into the next group of multi-scale channel adaptive residual modules, and the last is used to be weighted summed with the output of the attention weighted global filter module at the same level and then input into the next group of attention weighted global filter module; the output of the last group of multi-scale channel adaptive residual modules is used as part of the final local feature. The outputs of the first two groups of attention weighted global filter modules are divided into three branches, one of which is used as part of the final global feature, one is used to be weighted summed with the output of the multi-scale channel adaptive residual module at the same level and then input into the next group of multi-scale channel adaptive residual modules, and the last is used to be weighted summed with the output of the multi-scale channel adaptive residual module at the same level and then input into the next group of attention weighted global filter modules; the output of the last group of attention weighted global filter modules is used as part of the final global feature.

[0059] After the final local features and global features are spliced, they are input into the classification network, which includes: a linear layer, an ASP module, a batch normalization layer, a fully connected layer and a Softmax layer. The spliced ​​local features and global features are mapped to the ASP module through the linear layer. After the output of the ASP module is batch normalized, it is input into the fully connected layer, which is followed by the Softmax layer. The dialect type of the target speech is classified according to the global features and local features.

[0060] like Figure 3 As shown, the multi-scale channel adaptive residual module of the i-th group converts the input features After being processed by the linear layer, ReLU activation function and batch normalization layer, it is divided into two branches. One branch is processed by the SE module to extract channel attention, and the other branch is divided into N groups according to the channel to obtain sub-features:

[0061]

[0062] Sub-features are iteratively processed through N-1 multi-scale convolutional SE modules, ReLU activation functions, and batch normalization layers:

[0063]

[0064] When n>1, Finally, when n=N-1,

[0065] in, They represent the results of the 2nd, nth, n+1th, N-1th, and Nth multi-scale convolutional SE modules after being processed by the ReLU activation function and the batch normalization layer. BN() represents batch normalization, and MSC-SE 1 (), MSC-SE n ()MSC-SE (N-1) () represents the 1st, nth and N-1th multi-scale convolutional SE modules.

[0066] Will After splicing, the linear layer, ReLU activation function and batch normalization layer are processed and the weighted channel attention is applied to obtain the output of the i-th group of multi-scale channel adaptive residual modules.

[0067] like Figure 4 As shown, the multi-scale convolution SE module includes parallel convolution layers with convolution kernel sizes of 1, 3, 5, and 7. The number of output channels of each convolution layer is one-fourth of the number of input channels. The output of each convolution layer is weighted with its own channel weight extracted by the SE module, and then spliced ​​in the channel dimension. The spliced ​​result is processed by a linear layer, a ReLU activation function, and a batch normalization layer. Features of different local scales are obtained through convolution kernels of different scales, and then spliced ​​after attention weighting, so as to extract richer local features of speech and reflect the short-term local contextual characteristics of dialects.

[0068] The SE module is Squeeze-and-Excitation Block, a compression excitation module, which is a commonly used attention mechanism.

[0069] like Figure 5 As shown, the attention weighted global filtering module of the i-th group takes the input features After being processed by the linear layer, ReLU activation function and batch normalization layer, it is divided into two branches.

[0070] Input one of the branches into the global average pooling layer, the fully connected layer, the ReLU activation function layer, the fully connected layer and the Softmax layer to obtain the attention weights corresponding to several global filters, and add the learnable global filter and the corresponding attention weights to obtain the weighted global filter;

[0071] Each channel of the other branch is subjected to a one-dimensional Fourier transform to obtain the corresponding frequency band spectrum, and the frequency band spectrum is combined into a spectrum. The spectrum is evenly divided according to the time dimension and the first half is taken. The first half of the spectrum is subjected to a Hadamard product with a weighted global filter, and then an inverse Fourier transform is performed. The result of the inverse Fourier transform is processed by a linear layer, a ReLU activation function, and a batch normalization layer as an output. The attention weighted global filtering module obtains multiple contextual connections in a long-term range in speech through multiple groups of weighted global filters, which is conducive to better modeling the long-term global contextual characteristics of speech.

[0072] Mandarin and various dialects are collected in a balanced manner, expanded by speed changing and adding noise to the speech, and dialect recognition data is obtained by marking the dialect labels. The dialect recognition data is divided into a training set and a validation set. The training set is used to train the dialect recognition model. The cross entropy loss between the real dialect label and the predicted dialect type is used as the loss function. The stochastic gradient descent method is selected, and the ADAM optimizer is used. After the training dialect recognition model converges, the validation set is used to verify whether the dialect recognition model is overfitting.

[0073] Based on the target dialect type used by the customer and the agent's dialect ability, the target agent is automatically assigned to the customer by comprehensively considering the dialect matching, dialect proficiency information, agent availability, and business matching.

[0074] In the specific implementation process, after the target dialect is identified, the agent dialect ability database is combined with the customer's business needs and the target dialect to match the agent who can understand the target dialect and is responsible for the relevant business and is in an idle state as a candidate agent, and the dialect proficiency information of the candidate agent is compared to select the target agent. Specifically, if there are candidate agents who can listen and speak, the candidate agent with the highest comprehensive quantitative score is selected as the target agent. If there are no candidate agents who can listen and speak, and only the candidate agent who can listen, the candidate agent with the highest dialect comprehension ability is selected.

[0075] After the customer service is completed, the customer's evaluation feedback on the target agent is obtained. The content of the evaluation feedback includes service attitude evaluation, oral evaluation, comprehension ability evaluation and business professionalism evaluation; the target agent information in the agent dialect ability database is dynamically updated based on the customer's evaluation feedback.

[0076] Example 2

[0077] See also Figure 6As shown, an embodiment of the present invention provides a device for intelligent agent allocation based on customer dialect recognition, including: at least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit is a computer-readable storage medium, and can be used to store software programs, computer executable programs and modules, such as the software programs, computer executable programs and modules corresponding to an intelligent agent allocation method based on customer dialect recognition in an embodiment of the present invention. The processing unit implements the above-mentioned intelligent agent allocation method based on customer dialect recognition by running the software programs, computer executable programs and modules stored in the storage unit, including:

[0078] Pre-establishing an agent dialect ability database: the agent dialect ability database records the dialect types, dialect proficiency information, business scope, service quality evaluation and idle status of each agent;

[0079] Collect the target voice of the customer when communicating with the AI ​​service robot;

[0080] The target speech is subjected to noise reduction, and the target speech is scanned and framed using a first sliding window with a time width of t according to a set first step length to obtain a plurality of target speech frames, wherein the first step length is less than the time width t; and a frame width value of C is used. 1 The second sliding window scans the target speech frame according to the set second step length to obtain each group of input S k , where the second step length is less than the frame width value C 1 , input S k Inputting the target speech into a pre-trained dialect recognition model for dialect recognition, wherein the dialect recognition model outputs a target dialect type corresponding to the target speech;

[0081] Based on the target dialect type used by the customer and the agent's dialect ability, the target agent is automatically assigned to the customer by comprehensively considering the dialect matching, dialect proficiency information, agent's idle status, and business matching. After the customer service is completed, the customer's evaluation feedback on the target agent is obtained. The content of the evaluation feedback includes service attitude evaluation, oral evaluation, comprehension ability evaluation and business professionalism evaluation; the target agent information in the agent dialect ability database is dynamically updated according to the customer's evaluation feedback.

[0082] Of course, the computer program stored in the storage unit of the device for allocating intelligent agents based on customer dialect recognition provided in an embodiment of the present invention is not limited to the operation of the method described above, but can also execute related operations in the method for allocating intelligent agents based on customer dialect recognition provided in any embodiment of the present invention.

[0083] Example 3

[0084] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the intelligent agent assignment method based on customer dialect recognition is implemented, including:

[0085] Pre-establishing an agent dialect ability database: the agent dialect ability database records the dialect types, dialect proficiency information, business scope, service quality evaluation and idle status of each agent;

[0086] Collect the target voice of the customer when communicating with the AI ​​service robot;

[0087] The target speech is subjected to noise reduction, and the target speech is scanned and framed using a first sliding window with a time width of t according to a set first step length to obtain a plurality of target speech frames, wherein the first step length is less than the time width t; and a frame width value of C is used. 1 The second sliding window scans the target speech frame according to the set second step length to obtain each group of input S k , where the second step length is less than the frame width value C 1 , input S k Inputting the target speech into a pre-trained dialect recognition model for dialect recognition, wherein the dialect recognition model outputs a target dialect type corresponding to the target speech;

[0088] Based on the target dialect type used by the customer and the agent's dialect ability, the target agent is automatically assigned to the customer by comprehensively considering the dialect matching, dialect proficiency information, agent's idle status, and business matching. After the customer service is completed, the customer's evaluation feedback on the target agent is obtained. The content of the evaluation feedback includes service attitude evaluation, oral evaluation, comprehension ability evaluation and business professionalism evaluation; the target agent information in the agent dialect ability database is dynamically updated according to the customer's evaluation feedback.

[0089] A computer-readable storage medium provided in an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in an intelligent agent assignment method based on customer dialect recognition provided in any embodiment of the present invention.

[0090] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0091] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0092] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0093] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those 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 claimed herein.

Claims

1. An intelligent agent allocation method based on customer dialect recognition, characterized in that: include: Pre-establishing an agent dialect ability database: the agent dialect ability database records the dialect types, dialect proficiency information, business scope, service quality evaluation and idle status of each agent; Collect the target voice of the customer when communicating with the AI ​​service robot; Denoising the target speech, scanning the target speech according to a set first step length using a first sliding window with a time width t to perform framing, and obtaining a plurality of target speech frames, wherein the first step length is less than the time width t; The second sliding window with a frame width of C1 is used to scan the target speech frame according to the set second step length to obtain each group of input S k , where the second step length is less than the frame width value C1, the input S k The target dialect type corresponding to the target speech is output into a pre-trained dialect recognition model for dialect recognition, wherein the dialect recognition model outputs the target dialect type corresponding to the target speech; wherein the dialect recognition model includes: k One-dimensional convolution layer, ReLU activation function layer and batch normalization layer are used to map the channel dimension; the batch normalized features are divided into two parts according to the channel to obtain the features and Features feature It is input into the local feature extraction network to extract local features. The local features are input into the global feature extraction network to extract global features, and the local features and the global features are spliced ​​and then input into the classification network. The classification network includes: a linear layer, an ASP module, a batch normalization layer, a fully connected layer and a Softmax layer; the local feature extraction network includes: three groups of stacked multi-scale channel adaptive residual modules; the global feature extraction network includes: three groups of stacked attention weighted global filtering modules; the outputs of the first two groups of multi-scale channel adaptive residual modules are divided into three branches, one of which is used as part of the local features, one is used to be weighted summed with the output of the attention weighted global filtering module at the same level and then input into the next group of multi-scale channel adaptive residual modules, and the last one is used to be weighted summed with the output of the attention weighted global filtering module at the same level. The outputs of the attention-weighted global filter modules of one level are weighted and summed and then input into the next group of attention-weighted global filter modules; the output of the last group of multi-scale channel adaptive residual modules is used as part of the local features; the outputs of the first two groups of attention-weighted global filter modules are divided into three branches, one of which is used as part of the global features, one is used to weight the outputs of the multi-scale channel adaptive residual modules of the same level and then input into the next group of multi-scale channel adaptive residual modules, and the last one is used to weight the outputs of the multi-scale channel adaptive residual modules of the same level and then input into the next group of attention-weighted global filter modules; the output of the last group of attention-weighted global filter modules is used as part of the global features; Based on the target dialect type used by the customer and the agent's dialect ability, the target agent is automatically assigned to the customer by comprehensively considering the dialect matching, dialect proficiency information, agent's idle status, and business matching; after the customer service is completed, the customer's evaluation feedback on the target agent is obtained, and the content of the evaluation feedback includes service attitude evaluation, oral evaluation, comprehension ability evaluation and business professionalism evaluation; the target agent information in the agent dialect ability database is dynamically updated according to the customer's evaluation feedback.

2. The intelligent agent allocation method based on customer dialect recognition according to claim 1 is characterized in that: The dynamically updating the target agent information in the agent dialect capability database according to the customer's evaluation feedback includes: Update the dialect proficiency information of the target agent; the dialect proficiency information includes two priority indicators: able to listen and speak and able to listen only. The priority of being able to listen and speak is higher than that of being able to listen only, and a more refined scoring quantification is performed for being able to listen and speak and being able to listen only: if the target agent is the type who can listen and speak, the target agent's dialect speaking ability and dialect comprehension ability are quantified, and the weighted sum is used to obtain a comprehensive quantitative score of being able to listen and speak; if the target agent is the type who can only listen, the target agent's dialect comprehension ability is quantified; among them, the target agent's dialect speaking ability is quantified by taking the weighted sum of the company's oral assessment score for the target agent and the customer's oral evaluation of the target agent, and the agent's dialect comprehension ability is quantified by taking the weighted sum of the company's dialect listening test score for the target agent and the customer's evaluation of the target agent's comprehension ability; Update the target agent's service quality evaluation; weighted sum the customer's evaluation of the target agent's business professionalism and service attitude to obtain the service quality score of this service, and calculate the overall service quality average score as the service quality evaluation.

3. The intelligent agent allocation method based on customer dialect recognition according to claim 1 is characterized in that: The multi-scale channel adaptive residual module of the i-th group takes the input features After being processed by the linear layer, ReLU activation function and batch normalization layer, it is divided into two branches. One branch is processed by the SE module to extract channel attention, and the other branch is divided into N groups according to the channel to obtain sub-features: Sub-features are iteratively processed through N-1 multi-scale convolutional SE modules, ReLU activation functions, and batch normalization layers: When n>1, Finally, when n=N-1, in, They represent the results of the 2nd, nth, n+1th, N-1th, and Nth multi-scale convolutional SE modules after being processed by the ReLU activation function and the batch normalization layer, respectively. BN() represents batch normalization, MSC-SE1(), MSC-SE n ()MSC-SE (N-1) () represents the 1st, nth and N-1th multi-scale convolutional SE modules; Will After splicing, the linear layer, ReLU activation function and batch normalization layer are processed and the weighted channel attention is applied to obtain the output of the i-th group of multi-scale channel adaptive residual modules.

4. The intelligent agent allocation method based on customer dialect recognition according to claim 3 is characterized in that: The multi-scale convolution SE module includes four parallel convolutional layers with convolution kernel sizes of 1, 3, 5, and 7. The number of output channels of each convolutional layer is one-fourth of the number of input channels. The output of each convolutional layer is weighted with its own channel weight extracted by the SE module and then spliced ​​in the channel dimension. The spliced ​​result is processed by a linear layer, a ReLU activation function, and a batch normalization layer as output.

5. The intelligent agent allocation method based on customer dialect recognition according to claim 1 is characterized in that: The attention weighted global filtering module of the i-th group takes the input features After being processed by the linear layer, ReLU activation function and batch normalization layer, it is divided into two branches. Input one of the branches into the global average pooling layer, the fully connected layer, the ReLU activation function layer, the fully connected layer and the Softmax layer to obtain the attention weights corresponding to several global filters, and add the learnable global filter and the corresponding attention weights to obtain the weighted global filter; Perform one-dimensional Fourier transform on each channel of the other branch to obtain the corresponding frequency band spectrum, combine the frequency band spectrum into a spectrum, divide the spectrum equally according to the time dimension and take the first half, perform Hadamard product on the first half of the spectrum and the weighted global filter, and then perform inverse Fourier transform. The result of the inverse Fourier transform is processed by the linear layer, ReLU activation function and batch normalization layer as the output.

6. The intelligent agent allocation method based on customer dialect recognition according to claim 1 is characterized in that: We evenly collected Mandarin and various dialects, expanded them by speed-changing and adding noise, and marked dialect labels to obtain dialect recognition data. We divided the dialect recognition data into a training set and a validation set, and used the training set to train the dialect recognition model. The cross entropy loss between the real dialect label and the predicted dialect type was used as the loss function. We selected the stochastic gradient descent method and used the ADAM optimizer. After the training dialect recognition model converges, the validation set is used to verify whether the dialect recognition model is overfitting.

7. An intelligent agent allocation device based on customer dialect recognition, characterized in that: include: At least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the intelligent agent allocation method based on customer dialect recognition as described in any one of claims 1-6 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the intelligent agent allocation method based on customer dialect recognition as described in any one of claims 1-6.

Images