Talk skill generation method and device, computer equipment and storage medium

Abstract

The invention discloses a verbal skill generation method and device, computer equipment and a storage medium, relates to the technical field of artificial intelligence, and is applied to an insurance service continuous receiving scene. According to the method, the pre-trained continuous reception evaluation model is introduced to classify the historical continuous reception verbal skills, and the high-quality iterative training data set is constructed based on the continuous reception success and failure samples, so that the deep optimization training of the initial continuous reception verbal skill generation model is realized, and the deep continuous reception verbal skill generation model obtained through training has the advantages that the efficiency is high, and the efficiency is high. According to the method and the device, the target resuming call which conforms to the resuming specification, has a good communication effect and has a high success rate can be automatically generated according to the current dialogue situation. According to the method and the system, intelligence, standardization and individuation of verbal skill generation are realized, systematicness and efficiency of resumption work are greatly improved, service quality fluctuation caused by personnel difference is effectively reduced, and communication experience and resumption willingness of customers are remarkably improved, so that improvement of the resumption rate of enterprises and optimization of customer relationships are assisted.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and specifically to a speech generation method, apparatus, computer equipment and storage medium. Background Art

[0002] Premium renewal collection is a key link in the insurance service process. It refers to the service action taken by the insurance company to remind the policyholder to pay the next premium on time through various means such as text messages, phone calls, and APP notifications before the expiration of the payment period stipulated in the insurance contract, so as to maintain the continued validity of the policy.

[0003] Currently, premium renewal collection methods primarily include phone reminders, text message reminders, home visits, WeChat push notifications, and bank debits. For example, staff contact customers by phone to remind them to renew their premiums, or send renewal notifications via WeChat. Meanwhile, some institutions utilize apps or mini-programs to implement online renewal functions. However, these renewal collection methods often rely on the experience and personal abilities of renewal personnel and lack systematic and standardized guidance, resulting in low work efficiency and low customer satisfaction. The renewal process is prone to problems such as untimely information transmission, inconsistent service attitudes, and non-standardized reminder content, further exacerbating the risk of customer churn. Summary of the Invention

[0004] The purpose of the embodiments of this application is to propose a speech generation method, device, computer equipment and storage medium to solve the technical problems that the current premium renewal collection method mainly relies on personal experience, lacks standardization, is inefficient, and has a poor customer experience.

[0005] In order to solve the above technical problems, the present application provides a method for generating speech scripts, which adopts the following technical solutions:

[0006] A method for generating speech techniques, comprising:

[0007] Obtain historical speech and generate the historical speech text corresponding to the historical speech;

[0008] Input the historical speech text into the pre-trained speech evaluation model and output the speech evaluation results corresponding to the historical speech;

[0009] Based on the speech evaluation results, historical speech is classified into speech categories, wherein the speech categories include successful speech and failed speech;

[0010] Construct an iterative training dataset for the initial speech generation model based on the speech classification results;

[0011] Iteratively train the initial speech generation model using the iterative training dataset to obtain a deep speech generation model;

[0012] Receive the speech generation instruction, obtain the current conversation text, and input the current conversation text into the deep speech generation model to obtain the target speech.

[0013] In order to solve the above technical problems, the present application also provides a speech generation device, which adopts the following technical solution:

[0014] A speech generating device, comprising:

[0015] The historical speech module is used to obtain historical speech and generate the historical speech text corresponding to the historical speech;

[0016] The speech evaluation module is used to input historical speech text into the pre-trained speech evaluation model and output the speech evaluation results corresponding to the historical speech;

[0017] The speech classification module is used to classify historical speech based on speech evaluation results, where the speech classification includes successful speech and failed speech;

[0018] The dataset construction module is used to construct an iterative training dataset for the initial speech generation model based on the speech classification results;

[0019] An iterative training module is used to iteratively train the initial speech generation model using an iterative training dataset to obtain a deep speech generation model;

[0020] The speech generation module is used to receive speech generation instructions, obtain the current dialogue text, and input the current dialogue text into the deep speech generation model to obtain the target speech.

[0021] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0022] A computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the speech generation method as described in any one of the above items are implemented.

[0023] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0024] A computer-readable storage medium having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by a processor, implement the steps of the speech generation method as described in any one of the above.

[0025] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0026] The present application discloses a speech generation method, device, computer equipment and storage medium, which relate to the field of artificial intelligence technology and are applied to insurance business renewal scenarios. The present application classifies historical speech by introducing a pre-trained speech evaluation model, and constructs a high-quality iterative training data set based on samples of successful and failed speech, thereby achieving deep optimization training of the initial speech generation model. The deep speech generation model obtained through training can automatically generate target renewal speech that meets renewal standards, has good communication effects, and has a high success rate according to the current dialogue context. The present application realizes the intelligence, standardization and personalization of speech generation, greatly improves the systematicness and efficiency of renewal operations, effectively reduces the fluctuations in service quality caused by personnel differences, and significantly improves customers' communication experience and willingness to renew, thereby helping companies to improve renewal rates and optimize customer relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the solutions in this application, a brief introduction will be given below to the drawings required for use in the description of the embodiments of this application. Obviously, the drawings described below are some embodiments of this application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0028] Figure 1 shows an exemplary system architecture diagram in which the present application can be applied;

[0029] Figure 2 A flowchart of an embodiment of a method for generating speech techniques according to the present application is shown;

[0030] Figure 3 A schematic structural diagram of an embodiment of a speech generation device according to the present application is shown;

[0031] Figure 4 A schematic structural diagram of an embodiment of a computer device according to the present application is shown. DETAILED DESCRIPTION

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0033] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0034] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0035] like Figure 1 As shown, system architecture 100 may include a terminal device 101, a network 102, and a server 103. Terminal device 101 may be a laptop computer 1011, a tablet computer 1012, or a mobile phone 1013. Network 102 is a medium for providing a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0036] The user can use the terminal device 101 to interact with the server 103 via the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0037] The terminal device 101 can be various electronic devices with a display screen and supporting web browsing. In addition to the laptop computer 1011, tablet computer 1012 or mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV) player, a laptop computer and a desktop computer, etc.

[0038] The server 103 may be a server that provides various services, such as a background server that provides support for web pages displayed on the terminal device 101 .

[0039] It should be noted that the speech generation method provided in the embodiment of the present application is generally executed by a server / terminal device, and accordingly, the speech generation device is generally set in the server / terminal device.

[0040] It should be understood that Figure 1The numbers of terminal devices, networks and servers in the embodiment are merely illustrative. The above system may have any number of terminal devices, networks and servers according to implementation requirements.

[0041] Continue to refer Figure 2 , shows a flow chart of an embodiment of a method for generating speech techniques according to the present application. The speech technique generation method comprises the following steps:

[0042] S201, obtaining historical speech and generating historical speech text corresponding to the historical speech;

[0043] Specifically, acquiring historical conversations involves collecting historical communication data related to premium renewals, loan collection, and other matters. Data sources can include transcribed text from phone recordings, WeChat communication records, APP collection reminders, and text message content. To ensure data quality, the raw data must be initially cleaned, including removing irrelevant information, eliminating noisy text (such as garbled characters, advertisements, and non-renewal content), and unifying the text format. Subsequently, the historical conversations are standardized through natural language processing technology, such as extracting effective conversation content in segments and removing redundant greetings, thereby generating structured historical conversation text that can be used for model learning.

[0044] For example, the system extracted conversations between renewal agents and customers from phone recordings collected during the premium renewal process of an insurance company over the past two years. Using automated speech recognition (ASR) technology, the recordings were transcribed into text, forming a preliminary historical renewal transcript. For example, a recording was transcribed as follows: "Hello, I am from XX Insurance Company. I would like to remind you that your policy is due for renewal this month. If you have any questions, please feel free to contact us." The system then cleans the transcript, removing background noise, repetitive phrases, and irrelevant small talk, retaining only the effective conversational language related to renewal. The text is also uniformly formatted, such as by standardizing titles and simplifying lengthy sentences. Through this process, a standardized historical renewal transcript was successfully generated.

[0045] S202: Input the historical speech text into the pre-trained speech evaluation model and output the speech evaluation results corresponding to the historical speech;

[0046] Specifically, the pre-trained speech evaluation model is a model that has been trained in advance based on a large amount of historical speech data, and has strong speech understanding and evaluation capabilities. In this step, the generated historical text is used as input, and after text vectorization processing (such as pre-trained language models such as BERT to extract text features), it is input into the speech evaluation model for semantic understanding and analysis. The model will comprehensively consider the emotional tendency of the text, compliance expression, collection intensity, customer response tendency and other multi-dimensional features, and automatically judge the comprehensive performance of each renewal speech. The final output of the speech evaluation result can be a quantitative score (such as compliance score, emotional score, success expectation score), or a classification label (such as compliance / violation, positive / negative).

[0047] For example, a cleaned historical renewal text reads: "Hello, a reminder: Your policy has expired. Timely renewal protects your rights. If necessary, please contact me for assistance." The system inputs this text into a pre-trained renewal evaluation model. Based on a large number of renewal texts, the model learns features such as the compliance, emotional orientation, and renewal success probability of the script. After model processing, the output renewal evaluation results are: a compliance score of 92, an emotional friendliness score of 88, and a success expectation score of 75. The comprehensive evaluation results indicate that the script uses standard language, friendly expression, and a high renewal success rate.

[0048] For example, consider a historical renewal message that reads, "If you don't renew your policy, the consequences will be your own. We've contacted you multiple times. Please address this as soon as possible!" The system feeds this message into a pre-trained renewal evaluation model for analysis. After analyzing sentiment recognition, wording detection, and compliance assessment, the model outputs a renewal evaluation score of 45 for compliance, 30 for emotional friendliness, and 20 for success expectations. The results indicate that this renewal message uses threatening language and exhibits emotional instability, severely impacting the customer experience and potentially leading to complaints. Based on this evaluation, the message is classified as a failed renewal message.

[0049] S203, classifying the historical speech techniques based on the speech technique evaluation results, wherein the speech technique classification includes successful speech techniques and failed speech techniques;

[0050] Specifically, based on the evaluation results of the sales pitch output in step S202, combined with actual renewal success rate data (such as customer repayment records, renewal completion status, etc.), historical sales pitches are bidirectionally classified: one category is successful sales pitches, that is, effective sales pitches that facilitate customers to complete renewal or repayment during the actual renewal process; the other category is failed sales pitches, that is, sales pitches that fail to effectively promote collection and even cause customer disgust or loss of contact. During the classification process, the judgment can be made based on the set threshold rules (such as a success rate greater than a certain ratio is a successful sales pitch, otherwise it is a failed sales pitch), or through the assistance of a machine learning classifier.

[0051] S204, constructing an iterative training dataset for an initial speech generation model based on the speech classification results;

[0052] Specifically, based on the classification results of successful and failed speech, they are organized into positive sample sets and negative sample sets respectively. Each sample should contain information such as the original renewal text, the corresponding renewal label (successful or failed), and a summary of customer reactions (such as customer reply content or behavior records) to ensure the richness and diversity of the training data. When constructing the training data set, further data enhancement is required, such as synonymous sentence replacement, word order adjustment, partial content filling, etc., to improve the model's robustness to different expressions. In addition, according to business needs, a certain proportion of negative sampling strategies can be set to balance the number of positive and negative samples to avoid category skew during model training.

[0053] S205, iteratively training the initial speech generation model using the iterative training dataset to obtain a deep speech generation model;

[0054] Specifically, the constructed iterative training data set is input into the initial speech generation model for targeted iterative training. During the training process, the model's generation strategy is first optimized through the positive sample set to strengthen the model's ability to generate successful renewal speech and improve the success rate of generated speech in actual applications; then the model is trained through the negative sample set to identify and avoid inefficient and failed speech expressions to prevent the model from outputting inappropriate content. The training adopts a supervised learning method, combined with loss function optimization, gradient update and other mechanisms, to gradually adjust the model parameters, so that the model performs better in understanding the context of the conversation, grasping the timing of renewal and emotional guidance. After multiple rounds of iterative training, the final deep renewal speech generation model not only has high speech generation capabilities, but can also dynamically adapt to the communication styles of different customer groups, significantly improving the intelligence level of actual renewal operations.

[0055] S206, receiving a speech generation instruction, obtaining the current dialogue text, and inputting the current dialogue text into a deep speech generation model to obtain the target speech.

[0056] Specifically, during the system operation phase, it first receives a speech generation instruction from a business person or automatically triggered by the system, and simultaneously obtains the current conversation text information in real time, including the customer's historical communication content, recent communication records, and customer behavior characteristics. Then, these current conversation texts are input into the trained deep speech generation model. Based on the existing conversation context and customer profile, the model automatically infers and generates target speech that conforms to the renewal strategy. The generated speech will comprehensively consider factors such as the current conversation sentiment, customer repayment willingness, and historical collection records to ensure that the speech is both targeted and personalized, and meets the renewal compliance and communication specifications. The final generated target renewal speech can be directly pushed to the renewal staff for reference, or automatically sent to the customer by the system in the form of SMS, WeChat, telephone scripts, etc., thereby effectively improving the renewal success rate and customer experience.

[0057] In the above embodiment, the present application classifies historical speech by introducing a pre-trained speech evaluation model, and constructs a high-quality iterative training data set based on successful samples and failed samples, thereby achieving deep optimization training of the initial speech generation model. The deep speech generation model obtained through training can automatically generate target speech that meets the renewal standards, has good communication effects, and has a high success rate based on the current dialogue context. The present application realizes the intelligent, standardized, and personalized generation of speech, greatly improves the systematicness and efficiency of renewal operations, effectively reduces the fluctuations in service quality caused by personnel differences, and significantly improves customers' communication experience and willingness to renew, thereby helping companies improve renewal rates and optimize customer relationships.

[0058] Furthermore, the steps of obtaining historical speech and generating historical speech text corresponding to the historical speech specifically include:

[0059] Identify the data format of historical conversations, including audio, text, image, and video formats;

[0060] If the data format of the historical speech is in audio format, the audio format data is transcribed into text format using automatic speech recognition technology to obtain the historical speech text;

[0061] If the data format of the historical speech is text format, directly extract the text format data to obtain the historical speech text;

[0062] If the data format of the historical speech is in image format or video format, optical character recognition technology is used to recognize the text content in the image format data or video format data to obtain the historical speech text.

[0063] In this embodiment, for historical collection renewal speech data from different sources and in different storage forms, a unified data format recognition is first performed to ensure the consistency and standardization of subsequent processing. Specifically, when the historical collection renewal data is stored in the form of an audio file, the system uses automatic speech recognition (ASR) technology to transcribe the audio content and extract valid collection renewal text; when the historical collection renewal data itself is in text format (such as chat records, collection reminder text message content), the text is directly extracted without transcription processing; when the historical collection renewal data is in image or video format (such as collection renewal notification screenshots, video collection renewal clips), the system uses optical character recognition (OCR) technology to automatically extract and recognize text information in the image or video frame. In order to improve the recognition accuracy, the system can also combine semantic correction, denoising processing and other technologies in the transcription and recognition process to ensure that the generated historical collection renewal text content is complete, the expression is coherent, and there is less noise interference.

[0064] Through the above steps, the unified extraction and conversion of follow-up data in various formats such as audio, text, pictures, and videos is achieved, effectively ensuring the data standardization and accuracy of subsequent follow-up speech analysis and model training.

[0065] Furthermore, before the step of inputting the historical text into the pre-trained speech evaluation model and outputting the speech evaluation results corresponding to the historical speech, the method further includes:

[0066] Get the processing results corresponding to the historical words and get the historical processing results;

[0067] Input the historical processing results into the pre-trained speech evaluation model.

[0068] In this embodiment, before the historical renewal text is input into the pre-trained renewal evaluation model, the renewal results corresponding to the historical renewal speech are first obtained to form structured historical renewal result data. The renewal results may include actual business feedback information such as whether the customer has completed the renewal, the customer's response intention (such as positive response, refusal to communicate, clear refusal), and customer status changes (such as contact information update, complaint record). These historical renewal results can be extracted synchronously from channels such as the CRM system, renewal operation platform, and customer service system, and associated and matched with the corresponding speech text. Subsequently, the historical renewal text and the corresponding historical renewal results are input into the pre-trained renewal evaluation model, so that the model can make a comprehensive judgment based on the understanding of the speech content and the actual renewal results, thereby improving the accuracy and reliability of the renewal evaluation. By integrating real renewal effectiveness data, the model not only focuses on the text features of the speech itself, but also learns which speech features can promote positive results in actual business.

[0069] Through the above steps, the renewal evaluation model can be trained and judged based on actual renewal results, further improving the accuracy of renewal speech classification and its actual business guidance value.

[0070] Furthermore, the speech evaluation model is a classifier model. The steps of inputting historical text into the pre-trained speech evaluation model and outputting speech evaluation results corresponding to the historical speech specifically include:

[0071] Perform feature extraction on the historical processing results to obtain the first feature;

[0072] Extracting text semantic features from historical speech texts to obtain second features, where the second features include emotional features, compliance features, and speech skill features;

[0073] Import the first feature into the speech evaluation model and construct a classification center combination based on the first feature;

[0074] The second feature is classified using the classification center combination, and the speech evaluation result corresponding to the historical speech is determined based on the feature classification result.

[0075] In this embodiment, the renewal evaluation model acts as a classifier model, and first extracts features from historical renewal results to generate a first renewal feature. This feature includes customer feedback information directly related to the renewal results, such as repayment amount, renewal, customer behavior patterns, etc. These features help the model understand the correlation between the sales pitch and the actual results. Next, text semantic feature extraction is performed on the historical renewal text to obtain a second renewal feature, which includes emotional features (such as tone, attitude, etc.), compliance features (such as whether there is improper urging, legal risks, etc.) and renewal skill features (such as the attractiveness, effectiveness, patience, etc. of the sales pitch). After these first and second renewal features are input into the pre-trained renewal evaluation model, the model will construct a classification center combination based on the first renewal feature, and further use these combinations to classify the second renewal feature, thereby judging the effectiveness of each renewal speech. Finally, through the classification results, the model can output the renewal evaluation results of each historical renewal speech, helping to analyze its performance in actual applications.

[0076] Through the above steps, the renewal evaluation model combines historical results and text semantic features to effectively improve the accuracy and business adaptability of renewal script evaluation.

[0077] Furthermore, the classification center combination includes an emotion feature classification center, a compliance feature classification center, and a speech skill feature classification center. The steps of using the classification center combination to perform feature classification on the second feature, and determining the speech skill evaluation result corresponding to the historical speech skill based on the feature classification result specifically include:

[0078] Calling the emotion feature scoring benchmark in the emotion feature classification center, classifying and scoring the emotion features, and obtaining the emotion feature score;

[0079] Call the compliance feature scoring benchmark in the compliance feature classification center to classify and score the compliance features to obtain a compliance feature score;

[0080] Call the speech skill feature scoring benchmark in the speech skill feature classification center to classify and score the speech skill features to obtain the speech skill feature score;

[0081] Obtain the preset feature score weights, and based on the feature score weights, perform a weighted sum of the emotional feature score, compliance feature score, and speech skill feature score to obtain the total historical speech skill score;

[0082] Based on the total score of historical speech, the speech evaluation results corresponding to the historical speech are generated.

[0083] In this embodiment, the classification center combination includes an emotional feature classification center, a compliance feature classification center, and a renewal skill feature classification center. Each classification center uses the corresponding scoring benchmark to classify and score the features to evaluate the quality of each historical renewal speech. First, the emotional feature classification center analyzes the emotional features (such as tone, speech speed, attitude, etc.) in the historical renewal text according to the emotional scoring benchmark, and scores them to generate an emotional feature score; then, the compliance feature classification center uses the compliance scoring benchmark to classify the compliance issues in the dialogue, such as whether there is misleading language, threatening expressions, etc., to generate a compliance feature score; finally, the renewal skill feature classification center scores the strategic, attractive and persuasive nature of the dialogue according to the renewal skill scoring benchmark to obtain a renewal skill feature score. Then, the system extracts the corresponding weight value from the preset feature score weight, and performs a weighted summation of the emotional feature score, compliance feature score and renewal skill feature score to obtain the total score of the historical renewal speech. Finally, based on the total score of historical renewal scripts, the model will generate a renewal evaluation result, which can comprehensively reflect the effectiveness of the scripts, help renewal personnel understand the advantages and disadvantages of the scripts, and then optimize the script content and renewal strategies.

[0084] For example, suppose that during a renewal process at an insurance company, the conversation between the renewal agent and the customer is as follows:

[0085] Renewal text: "Hello, thank you for your continued support! Your policy is about to expire. Please renew as soon as possible, otherwise it may affect your protection. Can you arrange payment now? If you have any questions, I would be happy to answer them for you."

[0086] Step 1: The emotion feature classification center classifies and scores the emotion features.

[0087] The emotional feature classification center will evaluate the emotional features of the passage based on the tone, attitude, and sentiment analysis in the text. For example:

[0088] Tone: Friendly, gentle;

[0089] Attitude: Positive and caring;

[0090] Emotion: Positive and warm.

[0091] Assume that the system assigns an emotional characteristic score of 85 (out of 100) based on the scoring benchmark because the tone is friendly and the attitude is positive.

[0092] Step 2: The Compliance Characteristics Classification Center classifies and scores the compliance characteristics.

[0093] The Compliance Feature Classification Center will detect whether the speech has compliance issues, such as whether there are threatening, misleading, illegal words, etc. In this example:

[0094] The words used are not threatening to the customer and are not misleading or inappropriate;

[0095] Therefore, the compliance score is high.

[0096] Assume that the system gives a compliance feature score of 95 based on the scoring benchmark because there are no violations and the compliance standards are met.

[0097] Step 3: The renewal skill feature classification center classifies and scores the renewal skill features.

[0098] The Renewal Skills and Characteristics Classification Center will analyze the strategy, appeal, and persuasiveness of the sales pitch. For example:

[0099] Strategic: Use a caring tone to guide customers to pay;

[0100] Appeal: Use the phrase "impact guarantee" to arouse customers' sense of urgency;

[0101] Persuasion: Reduce customer resistance by saying, "If you have any questions, I'm happy to answer them."

[0102] Assume that the system gives the renewal skill feature a score of 80 based on the scoring benchmark. Although the sales pitch is reasonable, it lacks further incentives or clear payment guidance, and fails to maximize the renewal effect.

[0103] Step 4: Weighted sum of feature scores.

[0104] The system performs a weighted summation based on the preset feature score weights. For example, assuming the weights of sentiment, compliance, and follow-up skills are 30%, 40%, and 30%, respectively, the total score is calculated as follows:

[0105] Emotional characteristic score: 85*30%=25.5;

[0106] Compliance feature score: 95*40%=38;

[0107] Renewal skill feature score: 80*30%=24;

[0108] Total score = 25.5 + 38 + 24 = 87.5.

[0109] Step 5: Output the follow-up evaluation results.

[0110] Based on the final total score, the system will output the follow-up evaluation results. Assuming the score range is 0 to 100 points:

[0111] Follow-up evaluation results: Excellent, score 87.5.

[0112] This shows that this renewal script performs well in terms of emotion, compliance and skills, and is suitable for continued use in actual renewal.

[0113] Through the above steps, the system can comprehensively evaluate multi-dimensional factors such as emotions, compliance and skills of historical renewal scripts, generate comprehensive and accurate renewal evaluation results, and support the optimization of renewal strategies and improvement of their effectiveness.

[0114] Furthermore, the iterative training data set includes a first data set and a second data set. The steps of constructing the iterative training data set of the initial speech generation model according to the speech classification results specifically include:

[0115] Add the successful words to the first data set, and use the first data set as a positive sample set for iterative training;

[0116] Add the failed words to the second dataset and use the second dataset as a negative sample set for iterative training;

[0117] Combine the positive sample set and the negative sample set to obtain the iterative training data set.

[0118] In this embodiment, the construction of the iterative training data set is first based on the classification results of the renewal speech, and the successful speech and the failed speech are classified into different data sets. Specifically, the system adds all the speech that are judged to be successful in renewal to the first data set. This data set will serve as the positive sample set in the iterative training process, representing high-quality speech with efficient renewal effect; at the same time, the speech classified as failed renewal is added to the second data set. This data set serves as a negative sample set, representing inefficient speech that failed to successfully promote renewal or other goals in the actual renewal process. Afterwards, the system combines the positive sample set with the negative sample set to obtain a complete iterative training data set, which contains both positive and negative samples, making it easier for the model to learn how to distinguish the characteristics of successful speech and failed speech, and gradually optimize the initial renewal speech generation model through the iterative training process.

[0119] Through the above steps, the system can use comparative training of successful and failed scripts to improve the accuracy of the initial follow-up script generation model, making it more efficient in actual application and optimizing the follow-up script strategy.

[0120] Furthermore, the initial speech generation model is iteratively trained using the iterative training dataset to obtain the deep speech generation model, specifically including the following steps:

[0121] Use the training data in the positive sample set to perform the first iteration training on the initial speech generation model to generate the model's positive generation parameters;

[0122] Use the training data in the negative sample set to perform a second iteration of training on the initial speech generation model to generate negative correction parameters for the model;

[0123] The parameters of the initial speech generation model are updated using positive generation parameters and negative correction parameters to obtain a deep speech generation model.

[0124] In this embodiment, an initial renewal script generation model is iteratively trained using an iterative training dataset to continuously optimize the model's generation performance. First, the initial model undergoes a first iteration of training using training data from the positive sample set. The data in the positive sample set represents successful renewal cases, so the model learns how to generate scripts that better meet the criteria for successful renewals based on these positive examples. Through training, the model adjusts its generation parameters to better reflect the characteristics of successful scripts. Next, the model undergoes a second iteration of training using training data from the negative sample set, which contains failed renewals. Through this process, the model learns how to identify and correct biases in generating failed scripts. The model is then calibrated based on negative examples, generating negative correction parameters. Finally, the positive generation parameters are combined with the negative correction parameters to update the overall model parameters, generating an optimized deep renewal script generation model. Through continuous iterative training, this deep model can generate renewal scripts that better meet actual business needs, improving the efficiency of renewal personnel and optimizing the customer experience.

[0125] Positive generation parameters and negative correction parameters represent different optimization aspects the model relies on when generating sales pitches. For example, positive generation parameters include tone intensity, the frequency of friendly vocabulary, and the expression of positive emotions, helping the model generate more persuasive and engaging sales pitches. Negative correction parameters, on the other hand, include avoiding negative or threatening language and repetitive, ineffective persuasion. These parameters help correct for sales pitch characteristics that may lead to sales recurrence failures, ensuring compliance with the generated pitches and preventing customer frustration.

[0126] Suppose an insurance company is collecting premium renewals and the model uses iterative training to optimize the generation of renewal scripts. First, the system uses a set of positive examples, which include past successful renewal cases, such as conversations where customers ultimately renewed their policies on time or paid off their outstanding balances. The scripts in these positive examples often exhibit a gentle and persuasive tone, such as, "Hello, your policy is about to expire. We hope to continue to provide you with protection, and you can easily complete the renewal process." During this first round of training, the model uses this data to adjust its positive generation parameters, such as the positivity of the tone, the warmth of the emotions expressed, and the use of friendly vocabulary, thereby generating more persuasive and customer-friendly renewal scripts.

[0127] Next, the system utilizes a set of negative examples, consisting of conversations involving failed renewals, such as when customers don't renew or pay. These negative examples may involve overly forceful, impatient conversations, or phrases that put customers in a state of pressure, such as "If you don't renew soon, you will lose your coverage." By analyzing these negative examples, the model learns to identify the language that is likely to lead to renewal failures and, based on the feedback generated from these errors, conducts a second round of iterative training. At this point, the model adjusts negative correction parameters, such as avoiding excessive pressure, reducing threatening language, and adopting a gentle and understanding tone. Ultimately, through these two rounds of iteration, the model combines positive generation parameters with negative correction parameters to optimize its ability to generate renewal scripts, ensuring that the newly generated scripts better meet customer needs and company requirements in terms of tone, compliance, and persuasiveness.

[0128] Through the above steps, the system can continuously optimize the subsequent speech generation model through repeated iterative training of positive and negative samples, thereby improving the quality and accuracy of speech generation.

[0129] In the above embodiment, the present application discloses a method for generating speech, which relates to the field of artificial intelligence technology and is applied to the renewal scenario of insurance business. The present application classifies historical renewal speech by introducing a pre-trained renewal evaluation model, and constructs a high-quality iterative training data set based on renewal success and failure samples, thereby achieving deep optimization training of the initial renewal speech generation model. The deep renewal speech generation model obtained through training can automatically generate target renewal speech that meets renewal specifications, has good communication effects, and has a high success rate according to the current dialogue context. The present application realizes the intelligence, standardization, and personalization of speech generation, greatly improves the systematicness and efficiency of renewal operations, effectively reduces the service quality fluctuations caused by personnel differences, and significantly improves customers' communication experience and renewal willingness, thereby helping companies to improve renewal rates and optimize customer relationships.

[0130] In this embodiment, the electronic device on which the speech generation method is executed (eg Figure 1 The server shown in the figure can receive instructions or obtain data through a wired connection or a wireless connection. It should be noted that the above-mentioned wireless connection methods may include but are not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (Ultra Wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0131] It should be emphasized that in order to further ensure the privacy and security of the above-mentioned follow-up speech information, the above-mentioned follow-up speech information can also be stored in a node of a blockchain.

[0132] The blockchain referred to in this application refers to a new application model for computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Blockchain is essentially a decentralized database, a series of data blocks generated using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of the information (to prevent counterfeiting) and generate the next block. Blockchain can include the blockchain underlying platform, the platform product service layer, and the application service layer.

[0133] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial Intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve optimal results.

[0134] Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0135] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware via computer-readable instructions. The computer-readable instructions can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes in the above-described method embodiments. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0136] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0137] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a speech generation device, which is similar to Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0138] like Figure 3 As shown, the speech generation device 300 described in this embodiment includes:

[0139] The historical speech module 301 is used to obtain historical speech and generate historical speech text corresponding to the historical speech;

[0140] The speech evaluation module 302 is used to input the historical speech text into the pre-trained speech evaluation model and output the speech evaluation results corresponding to the historical speech;

[0141] The speech classification module 303 is used to classify the historical speech based on the speech evaluation results, wherein the speech classification includes successful speech and failed speech;

[0142] A data set construction module 304 is used to construct an iterative training data set for the initial speech generation model based on the speech classification results;

[0143] Iterative training module 305, used to iteratively train the initial speech generation model using the iterative training data set to obtain a deep speech generation model;

[0144] The speech generation module 306 is used to receive speech generation instructions, obtain the current dialogue text, and input the current dialogue text into the deep speech generation model to obtain the target speech.

[0145] Furthermore, the historical speech module 301 specifically includes:

[0146] A format recognition unit, configured to recognize the data format of the historical speech, wherein the data format includes audio format, text format, image format, and video format;

[0147] A first format processing unit is configured to, if the data format of the historical speech is in audio format, apply automatic speech recognition technology to transcribe the audio format data into text format to obtain the historical speech text;

[0148] A second format processing unit is configured to directly extract the text format data to obtain the text of the historical speech if the data format of the historical speech is text format;

[0149] The third format processing unit is used to use optical character recognition technology to recognize text content in the image format data or video format data to obtain the historical speech text if the data format of the historical speech is in image format or video format.

[0150] Furthermore, the speech generation device 300 further includes:

[0151] The processing result module is used to obtain the processing results corresponding to the historical speech and obtain the historical processing results;

[0152] The result input module is used to input historical processing results into the pre-trained speech evaluation model.

[0153] Furthermore, the speech evaluation model is a classifier model, and the speech evaluation module 302 specifically includes:

[0154] A first feature unit is used to extract features from the historical processing results to obtain a first feature;

[0155] The second feature unit is used to extract text semantic features from the historical speech text to obtain second features, wherein the second features include emotional features, compliance features, and speech skill features;

[0156] A classification center construction unit, configured to import the first feature into the speech evaluation model and construct a classification center combination based on the first feature;

[0157] The feature classification unit is used to classify the second feature using the classification center combination, and determine the speech evaluation result corresponding to the historical speech based on the feature classification result.

[0158] Furthermore, the classification center combination includes an emotion feature classification center, a compliance feature classification center, and a speech skill feature classification center. The feature classification units specifically include:

[0159] The emotion feature classification and scoring subunit is used to call the emotion feature scoring benchmark in the emotion feature classification center, classify and score the emotion features, and obtain the emotion feature score;

[0160] The compliance feature classification and scoring subunit is used to call the compliance feature scoring benchmark in the compliance feature classification center to classify and score the compliance features to obtain a compliance feature score;

[0161] The speech skill feature classification and scoring subunit is used to call the speech skill feature scoring benchmark in the speech skill feature classification center, classify and score the speech skill features, and obtain the speech skill feature score;

[0162] The feature score weighted summation sub-unit is used to obtain the preset feature score weights and perform weighted summation of the emotional feature score, compliance feature score, and speech skill feature score based on the feature score weights to obtain the total historical speech skill score;

[0163] The evaluation result generation sub-unit is used to generate the speech evaluation results corresponding to the historical continued speech based on the total score of the historical speech.

[0164] Furthermore, the data set construction module 304 specifically includes:

[0165] A positive sample set unit, configured to add successful speech patterns to the first data set and use the first data set as a positive sample set for iterative training;

[0166] A negative sample set unit, configured to add failed speech patterns to a second data set and use the second data set as a negative sample set for iterative training;

[0167] The sample set combination unit is used to combine the positive sample set and the negative sample set to obtain an iterative training data set.

[0168] Furthermore, the iterative training module 305 specifically includes:

[0169] A forward parameter unit, configured to perform a first iteration of training on the initial speech generation model using the training data in the positive sample set to generate the forward generation parameters of the model;

[0170] A negative parameter unit is used to perform a second iteration of training on the initial speech generation model using the training data in the negative sample set to generate negative correction parameters of the model;

[0171] The model parameter updating unit is used to update the parameters of the initial speech generation model using positive generation parameters and negative correction parameters to obtain a deep speech generation model.

[0172] In the above embodiment, the present application discloses a speech generation device, which relates to the field of artificial intelligence technology and is applied to insurance business renewal scenarios. The present application classifies historical speech by introducing a pre-trained speech evaluation model, and constructs a high-quality iterative training data set based on samples of successful and failed speech, thereby achieving deep optimization training of the initial speech generation model. The deep speech generation model obtained through training can automatically generate target renewal speech that meets renewal specifications, has good communication effects, and has a high success rate according to the current dialogue context. The present application realizes the intelligence, standardization, and personalization of speech generation, greatly improves the systematicness and efficiency of renewal operations, effectively reduces the fluctuations in service quality caused by personnel differences, and significantly improves customers' communication experience and willingness to renew, thereby helping companies to improve renewal rates and optimize customer relationships.

[0173] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0174] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with a memory 41, a processor 42, and a network interface 43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to a microprocessor, an application-specific integrated circuit (ASIC), a programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, etc.

[0175] The computer device may be a desktop computer, notebook computer, PDA, cloud server, etc. The computer device may interact with the user via a keyboard, mouse, remote control, touchpad, or voice control device.

[0176] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk equipped on the computer device 4, a smart memory card (SMC), a secure digital (SD) card, a flash card, etc. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of the speech generation method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0177] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions or process data stored in the memory 41, such as computer-readable instructions for executing the speech generation method.

[0178] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0179] The present application also provides an embodiment, namely, a computer device, comprising a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the above-mentioned user insurance demand assessment method are implemented, namely, the following steps are implemented:

[0180] A method for generating speech techniques, comprising:

[0181] Obtain historical speech and generate the historical speech text corresponding to the historical speech;

[0182] Input the historical speech text into the pre-trained speech evaluation model and output the speech evaluation results corresponding to the historical speech;

[0183] Based on the speech evaluation results, historical speech is classified into speech categories, wherein the speech categories include successful speech and failed speech;

[0184] Construct an iterative training dataset for the initial speech generation model based on the speech classification results;

[0185] Iteratively train the initial speech generation model using the iterative training dataset to obtain a deep speech generation model;

[0186] Receive the speech generation instruction, obtain the current conversation text, and input the current conversation text into the deep speech generation model to obtain the target speech.

[0187] The present application also provides another embodiment, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor to cause the at least one processor to perform the steps of the above-mentioned speech generation method, namely, to achieve:

[0188] A method for generating speech techniques, comprising:

[0189] Obtain historical speech and generate the historical speech text corresponding to the historical speech;

[0190] Input the historical speech text into the pre-trained speech evaluation model and output the speech evaluation results corresponding to the historical speech;

[0191] Based on the speech evaluation results, historical speech is classified into speech categories, wherein the speech categories include successful speech and failed speech;

[0192] Construct an iterative training dataset for the initial speech generation model based on the speech classification results;

[0193] Iteratively train the initial speech generation model using the iterative training dataset to obtain a deep speech generation model;

[0194] Receive the speech generation instruction, obtain the current conversation text, and input the current conversation text into the deep speech generation model to obtain the target speech.

[0195] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0196] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0197] It should be noted that the non-Company's software tools or components appearing in the various embodiments of this application are merely examples and do not represent actual use.

[0198] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.

Claims

1. A method for generating speech, characterized in that: include: Obtain historical speech and generate historical speech text corresponding to the historical speech; Input the historical speech text into the pre-trained speech evaluation model, and output the speech evaluation result corresponding to the historical speech; Classifying the historical speech techniques based on the speech technique evaluation results, wherein the speech technique classification includes successful speech techniques and failed speech techniques; Construct an iterative training dataset for the initial speech generation model based on the speech classification results; Iteratively train the initial speech generation model using the iterative training data set to obtain a deep speech generation model; Receive a speech generation instruction, obtain the current dialogue text, and input the current dialogue text into the deep speech generation model to obtain the target speech.

2. The speech generation method according to claim 1, wherein: The step of obtaining historical speech and generating historical speech text corresponding to the historical speech specifically includes: Identifying the data format of the historical speech, wherein the data format includes audio format, text format, image format and video format; If the data format of the historical speech is in audio format, the audio format data is transcribed into text format using automatic speech recognition technology to obtain the text of the historical speech; If the data format of the historical speech is text format, directly extract the text format data to obtain the historical speech text; If the data format of the historical speech is a picture format or a video format, optical character recognition technology is used to recognize the text content in the picture format data or the video format data to obtain the historical speech text.

3. The method for generating speech as claimed in claim 1, wherein: Before the step of inputting the historical speech text into the pre-trained speech evaluation model and outputting the speech evaluation result corresponding to the historical speech, the method further includes: Obtain the processing results corresponding to the historical speech to obtain the historical processing results; The historical processing results are input into the pre-trained speech evaluation model.

4. The method for generating speech as claimed in claim 3, wherein: The speech evaluation model is a classifier model. The step of inputting the historical speech text into the pre-trained speech evaluation model and outputting the speech evaluation result corresponding to the historical speech specifically includes: Performing feature extraction on the historical processing result to obtain a first feature; Extracting text semantic features from the historical speech text to obtain second features, wherein the second features include emotional features, compliance features, and speech skill features; Importing the first feature into the speech evaluation model, and constructing a classification center combination based on the first feature; The second feature is classified using the classification center combination, and the speech evaluation result corresponding to the historical speech is determined according to the feature classification result.

5. The method for generating speech as claimed in claim 4, characterized in that: The classification center combination includes an emotion feature classification center, a compliance feature classification center, and a speech skill feature classification center. The step of using the classification center combination to perform feature classification on the second feature and determining the speech skill evaluation result corresponding to the historical speech skill based on the feature classification result specifically includes: calling the emotion feature scoring benchmark in the emotion feature classification center, classifying and scoring the emotion feature, and obtaining an emotion feature score; calling the compliance feature scoring benchmark in the compliance feature classification center to classify and score the compliance features to obtain a compliance feature score; Calling the speech skill feature scoring benchmark in the speech skill feature classification center, classifying and scoring the speech skill features, and obtaining a speech skill feature score; Obtaining preset feature score weights, and performing weighted summation of the emotional feature score, the compliance feature score, and the speech skill feature score based on the feature score weights to obtain a total historical speech skill score; Based on the total score of the historical speech, a speech evaluation result corresponding to the historical speech is generated.

6. The method for generating speech as claimed in claim 1, wherein: The iterative training data set includes a first data set and a second data set. The step of constructing the iterative training data set of the initial speech generation model according to the speech classification results specifically includes: Adding the successful speech script to the first data set, and using the first data set as a positive sample set for iterative training; Adding the failed speech to the second data set, and using the second data set as a negative sample set for iterative training; The positive sample set and the negative sample set are combined to obtain the iterative training data set.

7. The method for generating speech as claimed in claim 6, characterized in that: The step of iteratively training the initial speech generation model using the iterative training dataset to obtain a deep speech generation model specifically includes: Performing a first iteration of training on the initial speech generation model using the training data in the positive sample set to generate positive generation parameters of the model; Performing a second iteration of training on the initial speech generation model using the training data in the negative sample set to generate negative correction parameters of the model; The parameters of the initial speech generation model are updated using the positive generation parameters and the negative correction parameters to obtain the deep speech generation model.

8. A speech generation device, characterized in that: include: A historical speech module is used to obtain historical speech and generate historical speech text corresponding to the historical speech; A speech evaluation module is used to input the historical speech text into a pre-trained speech evaluation model and output the speech evaluation result corresponding to the historical speech; A speech classification module, configured to classify the historical speech based on the speech evaluation result, wherein the speech classification includes successful speech and failed speech; The dataset construction module is used to construct an iterative training dataset for the initial speech generation model based on the speech classification results; An iterative training module, configured to iteratively train the initial speech generation model using the iterative training data set to obtain a deep speech generation model; The speech generation module is used to receive speech generation instructions, obtain the current dialogue text, and input the current dialogue text into the deep speech generation model to obtain the target speech.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, it implements the steps of the speech generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the speech generation method according to any one of claims 1 to 7.

Images