A method and system for evaluating the quality of voice generated by a smart phone customer service

Abstract

The application discloses a kind of intelligent telephone customer service generation voice quality evaluation method and system, belong to artificial intelligence technical field, method includes: obtaining all dialogue scene, the dialogue scene of the query probability greater than preset threshold is set to independent intent;The data of each independent intent is obtained, the feature pooling vector of each message is extracted by pre-training language model, and the feature pooling vector of each message is input into language intent recognition model;By pre-training language model, it is judged whether the query intent input by user meets minimum query condition;If yes, the query intent is identified by language intent recognition model, the query condition of query intent is extracted by entity extraction model, and the extracted query condition is input into voice dialing test system, dialing recording and answering recording are collected by voice dialing test system, dialing recording and answering recording are evaluated by audio processing center, and output voice quality evaluation result.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for evaluating the quality of voice generated by intelligent telephone customer service. Background Technology

[0002] Evaluating the quality of voice generated by intelligent telephone customer service is crucial for improving customer experience, increasing efficiency, reducing costs, enhancing brand reputation, and driving business growth. It is a key component of building a successful customer service strategy, helping to provide better service, meet customer needs, and increase the company's competitiveness.

[0003] Traditional methods for evaluating the quality of voice generated by intelligent telephone customer service often rely on manual measurement by test engineers. Test engineers can score the voice based on clarity, naturalness, fluency, and other aspects. However, this method requires a large amount of manpower and is easily affected by subjective factors such as work experience and testing environment, resulting in poor evaluation accuracy. Summary of the Invention

[0004] To address the technical problems of existing technologies that rely on manual measurement by test engineers, which requires a large human resource investment and is easily affected by subjective factors such as work experience and testing environment, resulting in poor evaluation accuracy, this invention provides a quality evaluation method and system for intelligent telephone customer service voice generation.

[0005] First aspect

[0006] This invention provides a method for evaluating the quality of voice generated by intelligent telephone customer service, including:

[0007] S101: Obtain all dialogue scenarios and set dialogue scenarios with a query probability greater than a preset threshold as independent intents;

[0008] S102: Obtain data for each independent intent, construct a sample dataset, the sample dataset including multiple messages, extract the feature pooling vector of each message through a pre-trained language model, and input the feature pooling vector of each message into the language intent recognition model;

[0009] S103: Using the pre-trained language model and the language intent recognition model, the output probability of each message is predicted by the activation function, and the output probability is used as a soft label for the message;

[0010] S104: Using the pre-trained language model, determine whether the user's input query intent meets the minimum query conditions; if yes, proceed to S105; otherwise, proceed to S106.

[0011] S105: The query intent is identified by the language intent recognition model, the query conditions of the query intent are extracted by the entity extraction model, and the extracted query conditions are input into the voice dialing test system. The voice dialing test system collects dialing and answering recordings, and the audio processing center evaluates the dialing and answering recordings and outputs the voice quality evaluation results.

[0012] S106: Prompt the user to add and / or modify query conditions;

[0013] S107: Input the user-added and / or modified query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

[0014] Second aspect

[0015] This invention provides a quality evaluation system for intelligent telephone customer service voice generation, comprising:

[0016] The settings module is used to obtain all dialogue scenarios and set dialogue scenarios with a query probability greater than a preset threshold as independent intents.

[0017] An extraction module is used to acquire data for each independent intent, construct a sample dataset, which includes multiple messages, extract the feature pooling vector of each message through a pre-trained language model, and input the feature pooling vector of each message into a language intent recognition model;

[0018] The prediction module is used to predict the output probability of each message using the pre-trained language model and the language intent recognition model, and to use the output probability as a soft label for the message.

[0019] The judgment module is used to determine whether the user's input query intent meets the minimum query conditions through the pre-trained language model; if so, it jumps to the first output module; otherwise, it jumps to the prompt module.

[0020] The first output module is used to identify the query intent through the language intent recognition model, extract the query conditions of the query intent through the entity extraction model, input the extracted query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

[0021] The prompt module is used to prompt users to add and / or modify query conditions;

[0022] The second output module is used to input the query conditions supplemented and / or modified by the user into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

[0023] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0024] In this invention, the quality evaluation of voice generated by intelligent telephone customer service is automatically completed through the cooperation of a pre-trained language model, a language intent recognition model, and a voice dialing test system. No human intervention is required, saving a lot of manpower costs, avoiding the influence of subjective factors, and improving the accuracy of voice quality evaluation. Attached Figure Description

[0025] The preferred embodiments will now be described in a clear and easy-to-understand manner, in conjunction with the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of the present invention.

[0026] Figure 1 This is a flowchart illustrating a method for evaluating the quality of voice generated by intelligent telephone customer service, provided by the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of an intelligent telephone customer service voice quality evaluation system provided by the present invention. Detailed Implementation

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0029] To keep the drawings concise, each figure only schematically shows the parts relevant to the invention, and these do not represent the actual structure of the product. Furthermore, to facilitate understanding, in some figures, only one of components with the same structure or function is schematically depicted, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one."

[0030] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0031] In this document, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections. They can refer to mechanical connections or electrical connections. They can refer to direct connections or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0032] Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0033] Example 1

[0034] In one embodiment, refer to the appendix to the specification. Figure 1 The diagram shows a flowchart of a method for evaluating the quality of voice generated by intelligent telephone customer service provided by the present invention.

[0035] This invention provides a method for evaluating the quality of voice generated by intelligent telephone customer service, comprising:

[0036] S101: Obtain all dialogue scenarios and set dialogue scenarios with a query probability greater than a preset threshold as independent intents.

[0037] Those skilled in the art can set the size of the preset threshold according to the actual situation, and the present invention does not limit it.

[0038] Each dialogue scenario represents an interaction between a user and the telephone customer service system, which may include the user's questions, answers, requests, etc.

[0039] Furthermore, within each dialogue scenario, the system estimates whether it contains the user's query intent. This assessment primarily relies on query probability, which estimates whether the scenario is relevant to the user's query. If the query probability exceeds a pre-set threshold, the dialogue scenario is considered to have independent intent. In subsequent processing, these scenarios with independent intent will be processed separately, without being mixed with other non-query-related scenarios.

[0040] In this invention, setting dialogue scenarios with a query probability greater than a preset threshold as independent intents helps improve the efficiency, accuracy, and user experience of the customer service system, while reducing the risk of misunderstanding and confusion, making the system more intelligent and adaptable.

[0041] S102: Obtain data for each independent intent, construct a sample dataset, which includes multiple messages, extract the feature pooling vector of each message through a pre-trained language model, and input the feature pooling vector of each message into the language intent recognition model.

[0042] Pre-trained language models are deep learning models trained on large-scale text data to learn the grammar, semantics, and contextual information within the text. These models are typically based on neural networks, such as the Transformer architecture. The goal of a pre-trained language model is to predict the next word or token in a given context. Through this task, the model can learn the structure and contextual information of the language, thereby enabling it to generate coherent text or understand the meaning of text.

[0043] Language intent recognition models are machine learning models used to identify the intent or category in natural language text. They can classify text into predefined categories to determine the intent expressed by the text. These models are typically based on deep learning and can be implemented using architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or Transformers. Typically, additional neural network layers are added to a pre-trained language model for classification tasks. Training a language intent recognition model requires labeled training data, where each text sample is associated with its corresponding intent or category. The model learns to extract features from the text and map them to the correct intent or category.

[0044] In one possible implementation, the loss function of the language intent recognition model is:

[0045]

[0046] Where Loss represents the loss function, θ MsgEnc θ represents the network parameters of the message encoder. Ipred The network parameters of the intent predictor are represented by CE, which represents the cross-entropy operation, and f. Ipred f represents the intent predictor; f represents the message encoder, x represents the intent predictor; f represents the message encoder; ... message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x i Represents the i-th message, y i x represents i The corresponding intent, D represents the dataset, and dataset D contains x i and x i Corresponding intention y i .

[0047] The loss function includes cross-entropy loss and message encoder loss. Cross-entropy loss measures the difference between the predicted intent and the true intent. The message encoder loss ensures that the feature vectors generated by the message encoder are meaningful and can be used for intent prediction.

[0048] In this invention, this loss function and parameter determination method enables end-to-end training while simultaneously optimizing the message encoder θ. MsgEnc and intention predictor θ Ipred The parameters allow the entire language intent recognition model to be learned in a single training process, without the need to train the two components separately.

[0049] The parameters of the language intent recognition model are determined by minimizing the loss function of the language intent recognition model.

[0050] In this invention, by minimizing the loss function, the model's parameters are adjusted to minimize the difference between the predicted intent and the true intent. This helps improve the accuracy of intent recognition, enabling the model to better understand textual information and classify it into the correct intent category.

[0051] In one possible implementation, the feature pooling vector for each message is extracted, specifically as follows:

[0052] The feature pooling vector is extracted using the following formula:

[0053]

[0054]

[0055] Among them, vec pooling h represents the feature pooling vector. i α represents the vector of the i-th word after a message has been segmented by the tokenizer of a pre-trained language model. i h represents the maximum preserved feature parameter of the i-th word vector. j Let represent the vector of the i-th word after a message has been segmented by the tokenizer of the pre-trained language model, N represent the total number of word vectors after a message has been segmented by the tokenizer of the pre-trained language model, d represent the model dimension of the pre-trained language model, and <h i ,h j > represents the inner product of the i-th word vector and the j-th word vector, max represents the maximum value function, and sigmoid represents the sigmoid function.

[0056] Among them, α is used i As weights for each word vector, an attention mechanism can be implemented, allowing the model to focus more on words that significantly contribute to intent recognition. This helps improve model performance, especially in long texts.

[0057] By applying the Sigmoid function, the inner product of word vectors can be mapped to the range (0,1), introducing a non-linear transformation that helps to capture the complex relationships and importance between words.

[0058] It should be noted that by applying feature pooling to each word vector, the information of the entire message can be encoded into a fixed-dimensional feature pooling vector, which helps to reduce the dimensionality of the data and allows messages of different lengths to be mapped to feature vectors of the same dimension.

[0059] S103: By using a pre-trained language model and a language intent recognition model, the output probability of each message is predicted using an activation function, and the output probability is used as a soft label for the message.

[0060] The activation function can be the Sigmoid function.

[0061] In this invention, using soft labels helps the model better handle uncertainty and noise. Because soft labels are probability distributions that can reflect multiple possible categories or intentions, the model is more robust and can handle various complex text inputs.

[0062] S104: Using a pre-trained language model, determine whether the user's query intent meets the minimum query conditions. If yes, proceed to S105. Otherwise, proceed to S106.

[0063] Minimum query criteria refer to the most basic requirements or conditions that a user's input query intent must meet in order for the system to process it or provide a response. These conditions are usually determined by business needs or system design to ensure that the information provided by the user is clear and complete enough so that the system can effectively understand and satisfy the user's request. Minimum query criteria can vary depending on different applications and scenarios.

[0064] For example, the system may require users' queries to contain specific keywords or phrases to clarify their intent. For instance, a restaurant reservation system might require queries to include keywords such as "reservation" or "booking." Furthermore, some queries may require additional information to be accepted by the system. For example, if a user wants to order a product, the system might require the user to provide necessary information such as the order number, quantity, model, or delivery address.

[0065] S105: The query intent is identified through the language intent recognition model, the query conditions of the query intent are extracted through the entity extraction model, and the extracted query conditions are input into the voice dialing test system. The voice dialing test system collects dialing and answering recordings, and the audio processing center evaluates the dialing and answering recordings and outputs the voice quality evaluation results.

[0066] Among them, the voice dialing system can automatically collect call recordings and answer recordings.

[0067] In one possible implementation, call recordings and answer recordings are collected through a voice testing system, and the call and answer recordings are evaluated by an audio processing center, which then outputs voice quality evaluation results, specifically including:

[0068] Multiple sets of caller and receiver audio data are obtained through the voice dialing system.

[0069] The audio processing center performs Fast Fourier Transform on the audio data to obtain spectral data. It then performs peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis on the spectral data to obtain and output speech quality evaluation results.

[0070] In this invention, by performing Fast Fourier Transform and various audio analyses, including spectrum, peak, tempo, loudness, energy, and timbre analysis, the quality of speech can be comprehensively evaluated. This enables the system to detect a variety of problems, such as noise, distortion, and audio loss.

[0071] In one possible implementation, peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis are performed on the spectral data, specifically including:

[0072] We constructed a ResNet50 based on a convolutional neural network and imported weights pre-trained based on an audio analysis dataset. We configured the input layer according to the input size, and the output of the input layer was used as the input of ResNet50. Then, we flattened the output feature vector of ResNet50 using an unfolding layer, adjusted the dimension of the feature vector by connecting fully connected layers, and finally connected a Softmax layer as a classifier.

[0073] ResNet50 is a deep convolutional neural network architecture and belongs to the ResNet series of models.

[0074] A clustering layer based on the Layer class in Keras is constructed. The input of the classifier is also used as the input of the clustering layer, and the clustering results generated by the clustering layer are used as the reference labels of the classifier to train the neural network.

[0075] Keras is an open-source deep learning framework designed to make building and training deep learning models easier and faster.

[0076] In this invention, by constructing a clustering layer and using the classifier's input as the clustering layer's input, clustering results can be automatically generated and used as the classifier's reference labels. This method allows the model to automatically learn the clustering structure of audio data without manual annotation.

[0077] The trained neural network is used to analyze and cluster the speech quality evaluation results to obtain the spectral energy value of the spectral data, the spectral energy range corresponding to the spectral energy value, and the audio analysis results data.

[0078] The audio analysis results are summarized and analyzed to obtain the speech quality evaluation results.

[0079] In this invention, the audio analysis results are summarized and analyzed to obtain speech quality evaluation results, making the evaluation results more comprehensive. This helps to better understand the quality of audio data and take corresponding measures to improve or optimize speech quality.

[0080] S106: Prompt the user to add and / or modify query conditions.

[0081] S107: Input the user-added and / or modified query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

[0082] For details on how to obtain the voice quality evaluation results, please refer to the content related to S105 above. To avoid repetition, this invention will not repeat the details.

[0083] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0084] In this invention, the quality evaluation of voice generated by intelligent telephone customer service is automatically completed through the cooperation of a pre-trained language model, a language intent recognition model, and a voice dialing test system. No human intervention is required, saving a lot of manpower costs, avoiding the influence of subjective factors, and improving the accuracy of voice quality evaluation.

[0085] Example 2

[0086] Reference manual attached Figure 2 The diagram shows a structural schematic of a quality evaluation system for intelligent telephone customer service voice generation provided by the present invention.

[0087] In one embodiment, the present invention provides a quality evaluation system 20 for intelligent telephone customer service voice generation, comprising:

[0088] The setting module 201 is used to obtain all dialogue scenarios and set dialogue scenarios with a query probability greater than a preset threshold as independent intents.

[0089] The extraction module 202 is used to acquire data for each independent intent, construct a sample dataset, which includes multiple messages. Through a pre-trained language model, the feature pooling vector of each message is extracted and input into the language intent recognition model.

[0090] The prediction module 203 is used to predict the output probability of each message by using an activation function through a pre-trained language model and a language intent recognition model, and to use the output probability as a soft label for the message.

[0091] The judgment module 204 is used to determine whether the user's query intent meets the minimum query conditions through a pre-trained language model; if so, it jumps to the first output module; otherwise, it jumps to the prompt module.

[0092] The first output module 205 is used to identify the query intent through a language intent recognition model, extract the query conditions of the query intent through an entity extraction model, input the extracted query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

[0093] The prompt module 206 is used to prompt the user to add and / or modify query conditions;

[0094] The second output module 207 is used to input the query conditions supplemented and / or modified by the user into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

[0095] In one possible implementation, the loss function of the language intent recognition model is:

[0096]

[0097] Where Loss represents the loss function, θ MsgEnc θ represents the network parameters of the message encoder. Ipred The network parameters of the intent predictor are represented by CE, which represents the cross-entropy operation, and f. Ipred f represents the intent predictor; f represents the message encoder, x represents the intent predictor; f represents the message encoder; ... message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x i Represents the i-th message, y i x represents i The corresponding intent, D represents the dataset, and dataset D contains x i and x i Corresponding intention y i .

[0098] The parameters of the language intent recognition model are determined by minimizing the loss function of the language intent recognition model.

[0099] In one possible implementation, the feature pooling vector for each message is extracted, specifically as follows:

[0100] The feature pooling vector is extracted using the following formula:

[0101]

[0102]

[0103] Among them, vec pooling h represents the feature pooling vector. i α represents the vector of the i-th word after a message has been segmented by the tokenizer of a pre-trained language model. i h represents the maximum preserved feature parameter of the i-th word vector. j Let represent the vector of the i-th word after a message has been segmented by the tokenizer of the pre-trained language model, N represent the total number of word vectors after a message has been segmented by the tokenizer of the pre-trained language model, d represent the model dimension of the pre-trained language model, and <h i ,h j > represents the inner product of the i-th word vector and the j-th word vector, max represents the maximum value function, and sigmoid represents the sigmoid function.

[0104] In one possible implementation, call recordings and answer recordings are collected through a voice testing system, and the call and answer recordings are evaluated by an audio processing center, which then outputs voice quality evaluation results, specifically including:

[0105] Multiple sets of caller and receiver audio data were obtained through the voice dialing system.

[0106] The audio processing center performs Fast Fourier Transform on the audio data to obtain spectral data. It then performs peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis on the spectral data to obtain and output speech quality evaluation results.

[0107] In one possible implementation, peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis are performed on the spectral data, specifically including:

[0108] We build a ResNet50 based on a convolutional neural network and import weights pre-trained based on an audio analysis dataset. We configure the input layer according to the input size, and use the output of the input layer as the input of ResNet50. Then, we use an unfolding layer to flatten the output feature vector of ResNet50, and then adjust the dimension of the feature vector by connecting fully connected layers. Finally, we connect a Softmax layer as a classifier.

[0109] A clustering layer based on the Layer class in Keras is constructed. The input of the classifier is also used as the input of the clustering layer. The clustering results generated by the clustering layer are used as the reference labels of the classifier to train the neural network.

[0110] The trained neural network is used to analyze and cluster the speech quality evaluation results to obtain the spectral energy value of the spectral data, the spectral energy range corresponding to the spectral energy value, and the audio analysis results data.

[0111] The audio analysis results are summarized and analyzed to obtain the speech quality evaluation results.

[0112] The intelligent telephone customer service voice quality evaluation system provided by the present invention can realize the steps and effects of the intelligent telephone customer service voice quality evaluation method in Embodiment 1 above. To avoid repetition, the present invention will not repeat them.

[0113] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0114] In this invention, the quality evaluation of voice generated by intelligent telephone customer service is automatically completed through the cooperation of a pre-trained language model, a language intent recognition model, and a voice dialing test system. No human intervention is required, saving a lot of manpower costs, avoiding the influence of subjective factors, and improving the accuracy of voice quality evaluation.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for evaluating the quality of voice generated by intelligent telephone customer service, characterized in that, include: S101: Obtain all dialogue scenarios and set dialogue scenarios with a query probability greater than a preset threshold as independent intents; S102: Obtain data for each independent intent, construct a sample dataset, the sample dataset including multiple messages, extract the feature pooling vector of each message through a pre-trained language model, and input the feature pooling vector of each message into the language intent recognition model; S103: Using the pre-trained language model and the language intent recognition model, the output probability of each message is predicted by the activation function, and the output probability is used as a soft label for the message; S104: Using the pre-trained language model, determine whether the user's input query intent meets the minimum query conditions; If so, execute S105; Otherwise, proceed to S106; S105: The query intent is identified by the language intent recognition model, the query conditions of the query intent are extracted by the entity extraction model, and the extracted query conditions are input into the voice dialing test system. The voice dialing test system collects dialing and answering recordings, and the audio processing center evaluates the dialing and answering recordings and outputs the voice quality evaluation results. S106: Prompt the user to add and / or modify query conditions; S107: Input the user-added and / or modified query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

2. The method for evaluating the quality of voice generated by intelligent telephone customer service according to claim 1, characterized in that, The loss function of the language intent recognition model is: Where Loss represents the loss function, θ MsgEnc θ represents the network parameters of the message encoder. Ipred The network parameters of the intent predictor are represented by CE, which represents the cross-entropy operation, and f. Ipred f represents the intent predictor; f represents the message encoder, x represents the intent predictor; f represents the message encoder; ... message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x i Represents the i-th message, y i x represents i The corresponding intent, D represents the dataset, and dataset D contains x i and x i Corresponding intention y i ; The parameters of the language intent recognition model are determined by minimizing the loss function of the language intent recognition model.

3. The method for evaluating the quality of voice generated by intelligent telephone customer service according to claim 1, characterized in that, The extraction of the feature pooling vector for each message is specifically as follows: The feature pooling vector is extracted using the following formula: Among them, vec pooling h represents the feature pooling vector. i α represents the vector of the i-th word after a message has been segmented by the tokenizer of a pre-trained language model. i h represents the maximum preserved feature parameter of the i-th word vector. j Let represent the vector of the i-th word after a message has been segmented by the tokenizer of the pre-trained language model, N represent the total number of word vectors after a message has been segmented by the tokenizer of the pre-trained language model, d represent the model dimension of the pre-trained language model, and <h i ,h j > represents the inner product of the i-th word vector and the j-th word vector, max represents the maximum value function, and sigmoid represents the sigmoid function.

4. The method for evaluating the quality of voice generated by intelligent telephone customer service according to claim 1, characterized in that, The process of collecting call and answer recordings through a voice testing system, evaluating the call and answer recordings through an audio processing center, and outputting voice quality evaluation results specifically includes: The voice dialing system obtains multiple sets of caller and receiver audio data. The audio processing center performs a fast Fourier transform on the audio data to obtain spectral data. It then performs peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis on the spectral data to obtain and output the speech quality evaluation results.

5. The method for evaluating the quality of voice generated by intelligent telephone customer service according to claim 4, characterized in that, The process of performing peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis on the spectral data specifically includes: We build a ResNet50 based on a convolutional neural network and import weights pre-trained based on an audio analysis dataset. We configure the input layer according to the input size, and use the output of the input layer as the input of ResNet50. Then, we use an unfolding layer to flatten the output feature vector of ResNet50, and then adjust the dimension of the feature vector by connecting fully connected layers. Finally, we connect a Softmax layer as a classifier. A clustering layer based on the Layer class in Keras is constructed. The input of the classifier is also used as the input of the clustering layer. The clustering result generated by the clustering layer is used as the reference label of the classifier to train the neural network. The trained neural network is used to analyze and cluster the speech quality evaluation results to obtain the spectral energy value of the spectral data, the spectral energy range corresponding to the spectral energy value, and the audio analysis result data. The audio analysis results are summarized and analyzed to obtain the speech quality evaluation results.

6. A quality evaluation system for intelligent telephone customer service voice generation, characterized in that, include: The settings module is used to obtain all dialogue scenarios and set dialogue scenarios with a query probability greater than a preset threshold as independent intents. An extraction module is used to acquire data for each independent intent, construct a sample dataset, which includes multiple messages, extract the feature pooling vector of each message through a pre-trained language model, and input the feature pooling vector of each message into a language intent recognition model; The prediction module is used to predict the output probability of each message using the pre-trained language model and the language intent recognition model, and to use the output probability as a soft label for the message. The judgment module is used to determine whether the user's input query intent meets the minimum query conditions through the pre-trained language model; if so, it jumps to the first output module. Otherwise, redirect to the prompt module; The first output module is used to identify the query intent through the language intent recognition model, extract the query conditions of the query intent through the entity extraction model, input the extracted query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results. The prompt module is used to prompt users to add and / or modify query conditions; The second output module is used to input user-added and / or modified query conditions into the voice dialing test system, collect dialing and answering recordings through the voice dialing test system, evaluate the dialing and answering recordings through the audio processing center, and output the voice quality evaluation results.

7. The intelligent telephone customer service voice quality evaluation system according to claim 6, characterized in that, The loss function of the language intent recognition model is: Where Loss represents the loss function, θ MsgEnc θ represents the network parameters of the message encoder. Ipred The network parameters of the intent predictor are represented by CE, which represents the cross-entropy operation, and f. Ipred f represents the intent predictor; f represents the message encoder, x represents the intent predictor; f represents the message encoder; ... message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x represents the message encoder; x i Represents the i-th message, y i x represents i The corresponding intent, D represents the dataset, and dataset D contains x i and x i Corresponding intention y i ; The parameters of the language intent recognition model are determined by minimizing the loss function of the language intent recognition model.

8. The intelligent telephone customer service voice quality evaluation system according to claim 6, characterized in that, The extraction of the feature pooling vector for each message is specifically as follows: The feature pooling vector is extracted using the following formula: Among them, vec pooling h represents the feature pooling vector. i α represents the vector of the i-th word after a message has been segmented by the tokenizer of a pre-trained language model. i h represents the maximum preserved feature parameter of the i-th word vector. j Let represent the vector of the i-th word after a message has been segmented by the tokenizer of the pre-trained language model, N represent the total number of word vectors after a message has been segmented by the tokenizer of the pre-trained language model, d represent the model dimension of the pre-trained language model, and <h i ,h j > represents the inner product of the i-th word vector and the j-th word vector, max represents the maximum value function, and sigmoid represents the sigmoid function.

9. The intelligent telephone customer service voice quality evaluation system according to claim 6, characterized in that, The process of collecting call and answer recordings through a voice testing system, evaluating the call and answer recordings through an audio processing center, and outputting voice quality evaluation results specifically includes: The voice dialing system obtains multiple sets of caller and receiver audio data. The audio processing center performs a fast Fourier transform on the audio data to obtain spectral data. It then performs peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis on the spectral data to obtain and output the speech quality evaluation results.

10. The intelligent telephone customer service voice quality evaluation system according to claim 9, characterized in that, The process of performing peak detection, rhythm point detection, loudness detection, energy analysis, and timbre analysis on the spectral data specifically includes: We build a ResNet50 based on a convolutional neural network and import weights pre-trained based on an audio analysis dataset. We configure the input layer according to the input size, and use the output of the input layer as the input of ResNet50. Then, we use an unfolding layer to flatten the output feature vector of ResNet50, and then adjust the dimension of the feature vector by connecting fully connected layers. Finally, we connect a Softmax layer as a classifier. A clustering layer based on the Layer class in Keras is constructed. The input of the classifier is also used as the input of the clustering layer. The clustering result generated by the clustering layer is used as the reference label of the classifier to train the neural network. The trained neural network is used to analyze and cluster the speech quality evaluation results to obtain the spectral energy value of the spectral data, the spectral energy range corresponding to the spectral energy value, and the audio analysis result data. The audio analysis results are summarized and analyzed to obtain the speech quality evaluation results.

Images