Intelligent dialogue management method of outbound system

Abstract

The invention discloses an intelligent dialogue management method of an outbound system, and relates to the technical field of intelligent management.In the system operation process, multi-dimensional data of clients are collected in real time, a client voice feature file is established through the voice recognition technology, cleaning, noise reduction, feature extraction and standardization are conducted on the data, and a preprocessed feature matrix is generated; according to an analysis result, a deep learning algorithm is applied, and after calculation, a timbre adaptation coefficient Zys and an intonation matching coefficient Zyt are obtained; a timbre and intonation model highly matched with customer features is dynamically generated; voice is synthesized in real time and dynamically adjusted and output is performed; a current scene is judged and an outbound content and an intonation style are optimized; selecting an intonation style according to a scene demand, collecting real-time feedback in an outbound process, continuously optimizing a tone and intonation matching model through deep learning, calculating to obtain a comprehensive adaptation index Zsa, comparing the comprehensive adaptation index Zsa with a preset adaptation threshold A, and judging whether adaptation succeeds or not.

Description

Technical Field

[0001] The present invention relates to the field of intelligent management technology, and in particular to a conversation intelligent management method for an outbound call system. Background Art

[0002] As an important tool in marketing promotion, outbound call systems have been widely used in many fields such as telemarketing, customer service, and market research. Traditional outbound call systems usually rely on fixed script templates and standardized voice output, but with the increasing diversification and personalization of customer needs, the single voice interaction mode has gradually shown its limitations. When receiving outbound call services, customers usually hope that the system can show a more natural, friendly and close to their own needs. The existing outbound call system faces significant shortcomings when dealing with complex dialogue scenarios, including the lack of emotional expression in voice output, the inability to accurately identify customer scenario needs, and the difficulty of interactive tone matching customer expectations. These deficiencies make it easy for customers to resist when interacting with the system, which is manifested in high hang-up rates, low conversation completion rates, and decreased customer satisfaction. In addition, the traditional system lacks the ability to dynamically adapt to customer emotions and intonations, resulting in a single and stiff voice style output in different scenarios, which is difficult to meet the diverse needs of customers.

[0003] The shortcomings of the existing outbound call system are mainly due to the multiple contradictions between technology and application requirements. On the one hand, the limitations of speech recognition and natural language processing technology make the system lack precision in capturing and understanding customer characteristics; on the other hand, fixed voice templates and scripted dialogue structures are difficult to adapt to the dynamic changes in customers' emotions and intonations. These shortcomings not only reduce the adaptability and flexibility of the outbound call system, but also lead to a series of abnormal effects in actual applications, such as customers' dissatisfaction with the stiff or mechanical voice output of the system, thereby increasing the hang-up rate and reducing customer engagement. In addition, in key scenarios (such as promotion invitations or customer satisfaction surveys), the mismatch between intonation and scenario requirements may also lead to the loss of potential customers and even have a negative impact on the corporate brand image. Therefore, there is an urgent need for an intelligent management method that can dynamically adapt to voice tone and intonation to fundamentally solve the problem of insufficient matching in outbound call systems in voice interaction. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method for intelligent conversation management of an outbound call system, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for intelligent dialogue management of an outbound call system, comprising the following steps: Step 1: Collect multi-dimensional data on the customer's voice, emotions, intonation, and conversational behavior in real time, use speech recognition technology to extract the customer's age, gender, and emotional characteristics, and establish a customer voice feature profile; Step 2: Analyze the conversation content through natural language processing (NLP), clean, reduce noise, extract features and standardize the data, generate a preprocessed feature matrix, and input it into the subsequent steps; Step 3: Based on the analysis results, by using the preset timbre and intonation model, including the voice template matching the age group and gender, and applying the deep learning algorithm, the timbre adaptation coefficient Zys and the intonation matching coefficient Zyt are calculated and obtained, and the timbre and intonation model that is highly consistent with the customer's characteristics is dynamically generated; Step 4: Based on the generated timbre and intonation model, the speech is synthesized in real time and the output is dynamically adjusted, and the speech speed, intonation curve and pause rhythm are adjusted in real time; Step 5: Based on the conversation context and scenario requirements, determine the current scenario and optimize the outbound call content and voice tone style, calculate the scenario adaptation coefficient Zsc, and select the voice tone style according to the scenario requirements, including enthusiasm, gentleness, and calmness; Step 6: Collect real-time feedback during the outbound call process, including hang-up rate, customer satisfaction, and conversation interruption points, continuously optimize the voice and intonation matching model through deep learning, calculate the comprehensive adaptation index Zsa, compare it with the preset adaptation threshold A, determine whether the adaptation is successful, and update the matching model; Step 7: Monitor and manage the running status of each step of the system, record abnormal events, provide an adjustable control interface, support flexible configuration of timbre, intonation, speech speed and scene thresholds, and allow manual adjustment of parameters or forced switching strategies under special circumstances.

[0006] Preferably, step one includes: Collect the customer's real-time voice data, including the frequency, amplitude, timbre characteristics and background noise characteristics of the voice signal, and complete the digital conversion of the audio input through a high-precision microphone array; The built-in speech analysis module extracts age, gender, intonation and emotion features from speech signals, and uses the fundamental frequency BGFR and formant frequency FRMT feature values ​​to establish a multi-dimensional customer speech feature vector; Collect speech features: fundamental frequency BGFR, formant frequency FRMT, acoustic feature MFCC, timbre feature TMBC, speech rate feature SPRT; Collect emotional features: emotion identification feature value EMFI, emotion change rate EMTR; Collect intonation characteristics: intonation curve characteristics PCVT, pause interval distribution PDIN, sentence end intonation change rate EPSR; Collect scenario data: conversation goal achievement rate CVRT, semantic matching degree SMAT, customer context perception characteristics CASC; Collect response behavior: feedback time distribution RTDP; Based on the extracted speech feature vector, the customer's speech feature file is generated and stored, and the file is structured and managed for subsequent timbre and intonation matching optimization.

[0007] Preferably, step 2 includes: Filter out redundant information and noise in the collected data, including the removal of static noise and the smoothing of dynamic signals, extract features from the cleaned data, including multi-dimensional indicators such as timbre feature MFCC and intonation change rate EPSR, generate a standardized feature matrix, provide a unified data format for subsequent calculations, perform dimensionless processing on the parameters in the feature matrix, normalize the dimensions and value ranges of different features, and eliminate scale differences between data.

[0008] Preferably, step three includes: Based on the customer's timbre features in the feature matrix, the timbre adaptation coefficient Zys is obtained after calculation, and the customer's timbre matching score is generated; Based on the customer's intonation characteristics, the intonation matching coefficient Zyt is calculated in combination with the intonation matching formula, the degree of adaptation of the customer's intonation characteristics is dynamically determined, and the intonation model parameters are output; The timbre adaptation score and intonation matching score are combined to dynamically generate a timbre and intonation model; the model includes timbre parameters, intonation curves, and speech speed adjustment suggestions.

[0009] Preferably, the timbre adaptation coefficient Zys is calculated by the following formula: ; In the formula, BGFR represents fundamental frequency, FRMT represents formant frequency, MFCC represents acoustic features, TMBC represents timbre features, and SPRT represents speech rate features. They represent the proportional coefficients of fundamental frequency BGFR, formant frequency FRMT, acoustic feature MFCC, timbre feature TMBC and speech rate feature SPRT respectively; The intonation matching coefficient Zyt is calculated by the following formula: ; In the formula, EMFI represents the emotion recognition feature value, EMTR represents the emotion change rate, PCVT represents the intonation curve characteristics, PDIN represents the pause interval distribution, EPSR represents the sentence end intonation change rate, They respectively represent the proportional coefficients of emotion recognition feature value EMFI, emotion change rate EMTR, intonation curve characteristic PCVT, pause interval distribution PDIN and sentence end intonation change rate EPSR.

[0010] Preferably, step four includes: Based on the timbre and intonation model, natural speech is synthesized through TTS technology. The speech content and the intonation model correspond one-to-one to ensure the adaptability and fluency of the output speech. In the speech generation process, the intonation parameters are adjusted in real time, including the fluctuation range of the intonation curve, the rate of change of the tone at the end of the sentence, and the pause rhythm, to ensure that the intonation changes are smooth and conform to the rules of natural conversation. According to the speed parameters in the timbre and intonation model, the playback speed of the output speech is dynamically adjusted to achieve personalized speed matching in different conversation situations.

[0011] Preferably, step five includes: According to the current dialogue context, the semantic matching degree SMAT and dialogue goal achievement rate CVRT scenario parameters are analyzed to determine the specific scenario requirements of the outbound call. The scenario adaptation coefficient Zsc is obtained by combining the speech speed adjustment index SSPR and the feedback time distribution RTDP. The scene adaptation coefficient Zsc is calculated by the following formula: ; In the formula, CVRT represents the dialogue goal achievement rate, SMAT represents the semantic matching degree, CASC represents the customer context perception characteristics, RTDP represents the feedback time distribution, They represent the proportional coefficients of conversation goal achievement rate CVRT, semantic matching degree SMAT, customer context perception characteristic CASC and feedback time distribution RTDP respectively; By comparing the scene adaptation coefficient Zsc with the preset threshold D, the matching degree between the current tone and the scene requirements is evaluated. According to the scene adaptation results, the corresponding tone style is selected and optimized, including three types: enthusiastic, gentle and calm. The scenario adaptation coefficient Zsc>preset threshold D indicates that the current intonation is highly matched with the scenario requirements, the customer accepts the intonation well, the hang-up rate is low, and the dialogue goal achievement rate is high. The intonation curve characteristic PCVT is optimized again to add more natural processing, adjust the intonation slightly upward at the end of the sentence, enhance emotional expression, and optimize the pause interval distribution PDIN to avoid emotional loss caused by fast or slow speech speed; If the scenario adaptation coefficient Zsc ≤ the preset threshold D, it indicates that there is a deviation between the current tone and the scenario requirements. Customers are disgusted with or not used to the tone, the hang-up rate is high, and the conversation completion rate is low. Analyze the feedback data of low-adaptation scenarios, focus on the incompatibility points, change the current tone from enthusiastic to gentle, and reduce the customer's psychological pressure. When the customer is indifferent or resistant, use a calm tone to weaken the emotional intensity of the system voice and avoid irritating the customer.

[0012] Preferably, step six includes: Monitor the outbound call process in real time, collect feedback data on hang-up rate, customer satisfaction, and conversation interruption points, and record the collection results in the feedback database. Iterate and optimize the timbre and intonation matching model through deep learning algorithms; Based on the feedback data, the comprehensive adaptation index Zsa is calculated and compared with the preset adaptation threshold A to determine the effectiveness of the current timbre and intonation matching.

[0013] Preferably, the comprehensive adaptation index Zsa is calculated by the following formula: ; Where Zys represents the timbre adaptation coefficient, Zyt represents the intonation matching coefficient, and Zsc represents the scene adaptation coefficient. They represent the proportional coefficients of the timbre adaptation coefficient Zys, the intonation matching coefficient Zyt, and the scene adaptation coefficient Zsc respectively; By comparing the comprehensive adaptation index Zsa with the preset threshold A, the evaluation strategy solution is obtained: The comprehensive adaptation index Zsa ≥ the preset adaptation threshold A+Δ. The comprehensive adaptation index Zsa is higher than the preset threshold A and exceeds the tolerance range ΔA. The customer feedback data shows good performance, including low hang-up rate, high conversation completion rate, and customer satisfaction at the first level. The timbre and intonation model is fine-tuned to further optimize by increasing the emotional richness of the intonation and enhancing the natural changes of the intonation curve. The preset threshold value A−ΔA≤comprehensive adaptation index Zsa<preset threshold value A+ΔA. The comprehensive adaptation index Zsa is close to the preset threshold value A and fluctuates within the tolerance range ΔA. Customer feedback data shows that some hang-ups or conversation interruptions occur, and customer satisfaction is at the second level. For scenarios with a high hang-up rate, a calm or gentle tone of voice is preferred to avoid strong tone that may cause customer resistance. The comprehensive adaptation index Zsa<preset threshold A−ΔA, the comprehensive adaptation index Zsa is lower than the preset threshold A, lower than the tolerance range ΔA, and the customer feedback data performance is unqualified, including high hang-up rate, frequent conversation interruptions and customer satisfaction lower than normal. Through cluster analysis of hang-up points and reasons for conversation interruptions, the causes of timbre, intonation or scene adaptation problems are located, a simulation scenario is established, and the adaptation effect of the new model is verified through manual intervention.

[0014] Preferably, step seven includes: Real-time monitoring of the running status of each step, including data collection, feature analysis, speech synthesis and feedback optimization process, recording abnormal events and generating running logs, providing a flexible control interface, allowing users to manually set the timbre, intonation, speech speed and scene threshold parameters, and supporting manual intervention operations in special circumstances; In case of sudden exceptions or system failures, the current strategy is forced to switch to ensure system stability and high availability.

[0015] The present invention provides a method for intelligent management of conversations in an outbound call system, which has the following beneficial effects: (1) When the system is running, it collects multi-dimensional data of customers in real time, uses speech recognition technology to establish customer speech feature files, cleans, de-noises, extracts features and standardizes the data, generates a pre-processed feature matrix, applies a deep learning algorithm based on the analysis results, and obtains the following after calculation: timbre adaptation coefficient Zys and intonation matching coefficient Zyt, dynamically generates a timbre and intonation model that is highly consistent with customer characteristics, synthesizes speech in real time and dynamically adjusts the output, determines the current scenario and optimizes the outbound call content and intonation style, calculates the scenario adaptation coefficient Zsc, selects the intonation style according to the scenario requirements, collects real-time feedback during the outbound call process, continuously optimizes the timbre and intonation matching model through deep learning, calculates the comprehensive adaptation index Zsa, compares it with the preset adaptation threshold A, and determines whether the adaptation is successful.

[0016] (2) Through the systematic design of the above seven steps, the outbound call system has completed key tasks such as real-time collection of customer voice data, feature analysis, model generation, speech synthesis and optimization feedback. The system can dynamically extract the customer's voice features, emotional features and scenario requirements, and combine the timbre and intonation model with deep learning algorithms to provide precise support for generating voice output that is highly consistent with customer needs during the outbound call process. This multi-dimensional dynamic adjustment mechanism not only achieves a high degree of matching between voice output and customer personalized needs, but also enhances the naturalness and fluency of customer interaction, effectively reduces customer resistance and hang-up rate, and significantly improves conversation completion rate and customer satisfaction.

[0017] (3) Compared with the fixed, template-based voice output method in traditional outbound call systems, this method significantly enhances the system's dynamic perception of customer characteristics and scenario needs by introducing advanced technologies such as speech recognition, natural language processing (NLP) and deep learning. Traditional technologies are difficult to effectively capture customer emotions and intonation characteristics, while this system realizes real-time data analysis and personalized voice generation, and can provide customized intonation styles for different customers and scenarios, including enthusiasm, gentleness, calmness and other types. By introducing the comprehensive adaptation index Zsa and the scenario adaptation coefficient Zsc, the system has achieved significant improvements in voice output accuracy and adaptability, effectively reducing customers' resistance to the voice system.

[0018] (4) By dynamically adjusting voice parameters, the hang-up rate is greatly reduced, and the conversation completion rate and customer satisfaction are significantly improved. Especially in complex scenarios, such as promotional invitations or customer complaint handling, by accurately matching the tone of voice, the system can effectively guide customer emotions and promote the achievement of business goals. At the same time, the system's feedback learning mechanism supports the continuous optimization of the model to ensure that the outbound call effect continues to improve as customer needs change. This closed-loop optimization intelligent management method not only brings higher marketing conversion rates to companies, but also improves customer experience, helps accumulate brand value in the long term, and further consolidates the core position of the outbound call system in marketing promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The present invention is a schematic diagram of the steps of a method for intelligent management of conversations in an outbound call system. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Example 1 The present invention provides a method for intelligent management of conversations in an outbound call system. Figure 1 , including the following steps: Step 1: Collect multi-dimensional data on the customer's voice, emotions, intonation, and conversational behavior in real time, use speech recognition technology to extract the customer's age, gender, and emotional characteristics, and establish a customer voice feature profile; Step 2: Analyze the conversation content through natural language processing (NLP), clean, reduce noise, extract features and standardize the data, generate a preprocessed feature matrix, and input it into the subsequent steps; Step 3: Based on the analysis results, by using the preset timbre and intonation model, including the voice template matching the age group and gender, and applying the deep learning algorithm, the timbre adaptation coefficient Zys and the intonation matching coefficient Zyt are calculated and obtained, and the timbre and intonation model that is highly consistent with the customer's characteristics is dynamically generated; Step 4: Based on the generated timbre and intonation model, the speech is synthesized in real time and the output is dynamically adjusted, and the speech speed, intonation curve and pause rhythm are adjusted in real time; Step 5: Based on the conversation context and scenario requirements, determine the current scenario and optimize the outbound call content and voice tone style, calculate the scenario adaptation coefficient Zsc, and select the voice tone style according to the scenario requirements, including enthusiasm, gentleness, and calmness; Step 6: Collect real-time feedback during the outbound call process, including hang-up rate, customer satisfaction, and conversation interruption points, continuously optimize the voice and intonation matching model through deep learning, calculate the comprehensive adaptation index Zsa, compare it with the preset adaptation threshold A, determine whether the adaptation is successful, and update the matching model; Step 7: Monitor and manage the running status of each step of the system, record abnormal events, provide an adjustable control interface, support flexible configuration of timbre, intonation, speech speed and scene thresholds, and allow manual adjustment of parameters or forced switching strategies under special circumstances.

[0022] In this embodiment, multi-dimensional data of the customer's voice, emotion, tone and conversation behavior are collected in real time, and the customer's age, gender and emotional characteristics are extracted by voice recognition technology to establish a customer voice feature file. The conversation content is analyzed by natural language processing (NLP), and the data is cleaned, denoised, feature extracted and standardized to generate a pre-processed feature matrix, which is input into the subsequent steps. According to the analysis results, by using a preset timbre and tone model, including a voice template matching the age group and gender, a deep learning algorithm is applied to obtain the following after calculation: the timbre adaptation coefficient Zys and the tone matching coefficient Zyt, and a timbre and tone model that is highly consistent with the customer's characteristics is dynamically generated. According to the generated timbre and tone model, the voice is synthesized in real time and the output is dynamically adjusted, and the speech speed, Intonation curve and pause rhythm, based on the dialogue context and scene requirements, judge the current scene and optimize the outbound call content and intonation style, calculate and obtain the scene adaptation coefficient Zsc, select the intonation style according to the scene requirements, including enthusiasm, gentleness and calmness, collect real-time feedback during the outbound call process, including hang-up rate, customer satisfaction and dialogue interruption points, continuously optimize the timbre and intonation matching model through deep learning, calculate and obtain the comprehensive adaptation index Zsa, compare it with the preset adaptation threshold A, judge whether the adaptation is successful, update the matching model, monitor and manage the running status of each step of the system, record abnormal events, provide an adjustable control interface, support flexible configuration of timbre, intonation, speaking speed and scene thresholds, and allow manual adjustment of parameters or forced switching strategies under special circumstances.

[0023] Example 2 This embodiment is explained in Example 1, please refer to Figure 1 , specifically: Step 1 includes: Collect the customer's real-time voice data, including the frequency, amplitude, timbre characteristics and background noise characteristics of the voice signal, and complete the digital conversion of the audio input through a high-precision microphone array; The built-in speech analysis module extracts age, gender, intonation and emotion features from speech signals, and uses the fundamental frequency BGFR and formant frequency FRMT feature values ​​to establish a multi-dimensional customer speech feature vector; Collect speech features: fundamental frequency BGFR, formant frequency FRMT, acoustic feature MFCC, timbre feature TMBC, speech rate feature SPRT; Collect emotional features: emotion identification feature value EMFI, emotion change rate EMTR; Collect intonation characteristics: intonation curve characteristics PCVT, pause interval distribution PDIN, sentence end intonation change rate EPSR; Collect scenario data: conversation goal achievement rate CVRT, semantic matching degree SMAT, customer context perception characteristics CASC; Collect response behavior: feedback time distribution RTDP; Based on the extracted speech feature vector, the customer's speech feature file is generated and stored, and the file is structured and managed for subsequent timbre and intonation matching optimization.

[0024] Step 2 includes: Filter out redundant information and noise in the collected data, including the removal of static noise and the smoothing of dynamic signals, extract features from the cleaned data, including multi-dimensional indicators such as timbre feature MFCC and intonation change rate EPSR, generate a standardized feature matrix, provide a unified data format for subsequent calculations, perform dimensionless processing on the parameters in the feature matrix, normalize the dimensions and value ranges of different features, and eliminate scale differences between data.

[0025] Step three includes: Based on the customer's timbre features in the feature matrix, the timbre adaptation coefficient Zys is obtained after calculation, and the customer's timbre matching score is generated; Based on the customer's intonation characteristics, the intonation matching coefficient Zyt is calculated in combination with the intonation matching formula, the degree of adaptation of the customer's intonation characteristics is dynamically determined, and the intonation model parameters are output; The timbre adaptation score and intonation matching score are combined to dynamically generate a timbre and intonation model; the model includes timbre parameters, intonation curves, and speech speed adjustment suggestions.

[0026] In this embodiment, through the implementation of the above method, the system can comprehensively collect and efficiently process the multi-dimensional voice feature data of the customer, dynamically generate personalized timbre and intonation models, and significantly improve the intelligence level and adaptability of the outbound call system. Step 1 ensures the comprehensiveness and accuracy of voice data collection. By extracting core parameters such as frequency, amplitude, emotion and intonation, a structured customer voice feature file is established, providing a high-quality data basis for subsequent analysis. Step 2 effectively eliminates data redundancy and dimensional differences through cleaning and standardization, generates a unified feature matrix, and ensures the efficiency and reliability of subsequent calculations. Step 3 combines deep learning algorithms to calculate the timbre adaptation coefficient Zys and the intonation matching coefficient Zyt, and generates a dynamic timbre and intonation model that is highly consistent with customer characteristics, which not only meets the personalized needs of different customers, but also optimizes the voice interaction effect of the system, significantly reduces customer resistance, and improves the conversation completion rate and customer satisfaction.

[0027] Example 3 This embodiment is explained in Example 1, please refer to Figure 1 Specifically: the timbre adaptation coefficient Zys is calculated by the following formula: ; In the formula, BGFR represents fundamental frequency, FRMT represents formant frequency, MFCC represents acoustic features, TMBC represents timbre features, and SPRT represents speech rate features. They represent the proportional coefficients of fundamental frequency BGFR, formant frequency FRMT, acoustic feature MFCC, timbre feature TMBC and speech rate feature SPRT respectively; The intonation matching coefficient Zyt is calculated by the following formula: ; In the formula, EMFI represents the emotion recognition feature value, EMTR represents the emotion change rate, PCVT represents the intonation curve characteristics, PDIN represents the pause interval distribution, EPSR represents the sentence end intonation change rate, They respectively represent the proportional coefficients of emotion recognition feature value EMFI, emotion change rate EMTR, intonation curve characteristic PCVT, pause interval distribution PDIN and sentence end intonation change rate EPSR.

[0028] Step 4 includes: Based on the timbre and intonation model, natural speech is synthesized through TTS technology. The speech content and the intonation model correspond one-to-one to ensure the adaptability and fluency of the output speech. In the speech generation process, the intonation parameters are adjusted in real time, including the fluctuation range of the intonation curve, the rate of change of the tone at the end of the sentence, and the pause rhythm, to ensure that the intonation changes are smooth and conform to the rules of natural conversation. According to the speed parameters in the timbre and intonation model, the playback speed of the output speech is dynamically adjusted to achieve personalized speed matching in different conversation situations.

[0029] In this embodiment, through the accurate calculation of the timbre adaptation coefficient Zys and the intonation matching coefficient Zyt, and the real-time speech synthesis and adjustment based on the timbre and intonation model, the intelligence level and customer experience of the system in the outbound call dialogue have been significantly improved. The calculation of the timbre adaptation coefficient Zys integrates key parameters such as the fundamental frequency BGFR, the formant frequency FRMT, and the acoustic feature MFCC to ensure the naturalness and matching degree of the output voice timbre; the intonation matching coefficient Zyt combines the multi-dimensional data such as the emotion recognition feature value EMFI and the intonation curve characteristic PCVT to dynamically judge the degree of adaptation of the intonation characteristics to customer needs. Through the timbre and intonation models generated by these calculation results, the system can adjust the intonation curve, the end-of-sentence change rate and the pause rhythm of the voice output in real time to ensure the smoothness and naturalness of the voice interaction. In addition, the dynamic adjustment of the speech rate parameter further improves the fluency and emotional matching effect of the dialogue, and realizes the accurate output of personalized voice services. The above optimization not only enhances the customer's sense of participation and satisfaction, but also effectively reduces the hang-up rate and improves the achievement rate of the dialogue goal, providing strong technical support for the marketing and promotion of the outbound call system.

[0030] Example 4 This embodiment is explained in Example 1, please refer to Figure 1 , specifically: Step 5 includes: According to the current dialogue context, the semantic matching degree SMAT and dialogue goal achievement rate CVRT scenario parameters are analyzed to determine the specific scenario requirements of the outbound call. The scenario adaptation coefficient Zsc is obtained by combining the speech speed adjustment index SSPR and the feedback time distribution RTDP. The scene adaptation coefficient Zsc is calculated by the following formula: ; In the formula, CVRT represents the dialogue goal achievement rate, SMAT represents the semantic matching degree, CASC represents the customer context perception characteristics, RTDP represents the feedback time distribution, They represent the proportional coefficients of conversation goal achievement rate CVRT, semantic matching degree SMAT, customer context perception characteristic CASC and feedback time distribution RTDP respectively; By comparing the scene adaptation coefficient Zsc with the preset threshold D, the matching degree between the current tone and the scene requirements is evaluated. According to the scene adaptation results, the corresponding tone style is selected and optimized, including three types: enthusiastic, gentle and calm. The scenario adaptation coefficient Zsc>preset threshold D indicates that the current intonation is highly matched with the scenario requirements, the customer accepts the intonation well, the hang-up rate is low, and the dialogue goal achievement rate is high. The intonation curve characteristic PCVT is optimized again to add more natural processing, adjust the intonation slightly upward at the end of the sentence, enhance emotional expression, and optimize the pause interval distribution PDIN to avoid emotional loss caused by fast or slow speech speed; If the scenario adaptation coefficient Zsc ≤ the preset threshold D, it indicates that there is a deviation between the current tone and the scenario requirements. Customers are disgusted with or not used to the tone, the hang-up rate is high, and the conversation completion rate is low. Analyze the feedback data of low-adaptation scenarios, focus on the incompatibility points, change the current tone from enthusiastic to gentle, and reduce the customer's psychological pressure. When the customer is indifferent or resistant, use a calm tone to weaken the emotional intensity of the system voice and avoid irritating the customer.

[0031] Step six includes: Monitor the outbound call process in real time, collect feedback data on hang-up rate, customer satisfaction, and conversation interruption points, and record the collection results in the feedback database. Iterate and optimize the timbre and intonation matching model through deep learning algorithms; Based on the feedback data, the comprehensive adaptation index Zsa is calculated and compared with the preset adaptation threshold A to determine the effectiveness of the current timbre and intonation matching.

[0032] In this embodiment, through the implementation of steps five and six, the system can accurately identify the needs of the dialogue scene and dynamically adjust the tone style, and optimize the model performance through feedback data, thereby significantly improving the adaptability and customer experience of the outbound call system. In step five, through in-depth analysis of scene parameters (such as semantic matching degree SMAT, dialogue goal achievement rate CVRT, etc.), and calculation of the scene adaptation coefficient Zsc, the system can effectively evaluate the matching degree between the tone and the scene. When Zsc>D, the system further optimizes the tone curve and pause rhythm to enhance the naturalness of emotional expression; when Zsc≤D, it switches to a gentle or calm tone to reduce customer resistance and ensure that voice interaction is more in line with customer emotional needs. Step six realizes the continuous improvement of the timbre and tone matching model through the collection of real-time feedback data and iterative optimization of the deep learning model. By comparing the calculation of the comprehensive adaptation index Zsa with the preset threshold A, the system can promptly discover adaptation problems and take targeted measures. The above optimization measures significantly reduced the hang-up rate, improved the conversation completion rate and customer satisfaction, and at the same time achieved precise adaptation to scenario requirements, laying a solid foundation for the efficient application of the outbound call system in diverse scenarios.

[0033] Example 5 This embodiment is explained in Example 1, please refer to Figure 1 Specifically, the comprehensive adaptation index Zsa is calculated by the following formula: ; Where Zys represents the timbre adaptation coefficient, Zyt represents the intonation matching coefficient, and Zsc represents the scene adaptation coefficient. They represent the proportional coefficients of the timbre adaptation coefficient Zys, the intonation matching coefficient Zyt, and the scene adaptation coefficient Zsc respectively; By comparing the comprehensive adaptation index Zsa with the preset threshold A, the evaluation strategy solution is obtained: The comprehensive adaptation index Zsa ≥ the preset adaptation threshold A+Δ. The comprehensive adaptation index Zsa is higher than the preset threshold A and exceeds the tolerance range ΔA. The customer feedback data shows good performance, including low hang-up rate, high conversation completion rate, and customer satisfaction at the first level. The timbre and intonation model is fine-tuned to further optimize by increasing the emotional richness of the intonation and enhancing the natural changes of the intonation curve. The preset threshold value A−ΔA≤comprehensive adaptation index Zsa<preset threshold value A+ΔA. The comprehensive adaptation index Zsa is close to the preset threshold value A and fluctuates within the tolerance range ΔA. Customer feedback data shows that some hang-ups or conversation interruptions occur, and customer satisfaction is at the second level. For scenarios with a high hang-up rate, a calm or gentle tone of voice is preferred to avoid strong tone that may cause customer resistance. The comprehensive adaptation index Zsa<preset threshold A−ΔA, the comprehensive adaptation index Zsa is lower than the preset threshold A, lower than the tolerance range ΔA, and the customer feedback data performance is unqualified, including high hang-up rate, frequent conversation interruptions and customer satisfaction lower than normal. Through cluster analysis of hang-up points and reasons for conversation interruptions, the causes of timbre, intonation or scene adaptation problems are located, a simulation scenario is established, and the adaptation effect of the new model is verified through manual intervention.

[0034] Step seven includes: Real-time monitoring of the running status of each step, including data collection, feature analysis, speech synthesis and feedback optimization process, recording abnormal events and generating running logs, providing a flexible control interface, allowing users to manually set the timbre, intonation, speech speed and scene threshold parameters, and supporting manual intervention operations in special circumstances; In case of sudden exceptions or system failures, the current strategy is forced to switch to ensure system stability and high availability.

[0035] In this embodiment, through the calculation and evaluation strategy of the comprehensive adaptation index Zsa, as well as the monitoring of the running status of each step and the implementation of the flexible control interface, the system effectively improves the accuracy, reliability and dynamic adaptation ability of outbound interactions. In the comprehensive adaptation evaluation, by comparing Zsa with the preset threshold A, the system can adjust the intonation style and optimization strategy in a timely manner according to the adaptation level. When Zsa ≥ A + Δ, the system further improves the interaction effect in high-matching scenarios by fine-tuning the emotional richness of the intonation curve; when A - ΔA ≤ Zsa < A + ΔA, the system preferentially selects a calm or gentle intonation for scenarios with a high hang-up rate to optimize the customer experience; and when Zsa < A - Δ, the system verifies the new model through feedback analysis and manual intervention in low-adaptation scenarios, and gradually repairs the adaptation problems. At the same time, through the real-time monitoring and flexible control in Step 7, the system can quickly record and respond to abnormal events, support the flexible configuration and manual adjustment of various operating parameters, automatically switch strategies in case of sudden anomalies or system failures, and ensure overall stability and high availability. The above optimization measures effectively reduce the hang-up rate and conversation interruption rate, improve the conversation completion rate and customer satisfaction, and provide a strong guarantee for the intelligent operation of the outbound system in complex scenarios.

[0036] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent conversation management of an outbound call system, characterized in that: The following steps are involved: Step 1: Collect multi-dimensional data on the customer's voice, emotions, intonation, and conversational behavior in real time, use speech recognition technology to extract the customer's age, gender, and emotional characteristics, and establish a customer voice feature profile; Step 2: Analyze the conversation content through natural language processing (NLP), clean, reduce noise, extract features and standardize the data, generate a preprocessed feature matrix, and input it into the subsequent steps; Step 3: Based on the analysis results, by using the preset timbre and intonation model, including the voice template matching the age group and gender, and applying the deep learning algorithm, the timbre adaptation coefficient Zys and the intonation matching coefficient Zyt are calculated and obtained, and the timbre and intonation model that is highly consistent with the customer's characteristics is dynamically generated; Step 4: Based on the generated timbre and intonation model, the speech is synthesized in real time and the output is dynamically adjusted, and the speech speed, intonation curve and pause rhythm are adjusted in real time; Step 5: Based on the conversation context and scenario requirements, determine the current scenario and optimize the outbound call content and voice tone style, calculate the scenario adaptation coefficient Zsc, and select the voice tone style according to the scenario requirements, including enthusiasm, gentleness, and calmness; Step 6: Collect real-time feedback during the outbound call process, including hang-up rate, customer satisfaction, and conversation interruption points, continuously optimize the voice and intonation matching model through deep learning, calculate the comprehensive adaptation index Zsa, compare it with the preset adaptation threshold A, determine whether the adaptation is successful, and update the matching model; Step 7: Monitor and manage the running status of each step of the system, record abnormal events, provide an adjustable control interface, support flexible configuration of timbre, intonation, speech speed and scene thresholds, and allow manual adjustment of parameters or forced switching strategies under special circumstances.

2. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step 1 includes: Collect the customer's real-time voice data, including the frequency, amplitude, timbre characteristics and background noise characteristics of the voice signal, and complete the digital conversion of the audio input through a high-precision microphone array; The built-in speech analysis module extracts age, gender, intonation and emotion features from speech signals, and uses the fundamental frequency BGFR and formant frequency FRMT feature values ​​to establish a multi-dimensional customer speech feature vector; Collect speech features: fundamental frequency BGFR, formant frequency FRMT, acoustic feature MFCC, timbre feature TMBC, speech rate feature SPRT; Collect emotional features: Emotion identification feature value EMFI, emotion change rate EMTR; Collect intonation characteristics: intonation curve characteristics PCVT, pause interval distribution PDIN, sentence end intonation change rate EPSR; Collect scenario data: conversation goal achievement rate CVRT, semantic matching degree SMAT, customer context perception characteristics CASC; Collect response behavior: feedback time distribution RTDP; Based on the extracted speech feature vector, the customer's speech feature file is generated and stored, and the file is structured and managed for subsequent timbre and intonation matching optimization.

3. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step 2 includes: Filter out redundant information and noise in the collected data, including the removal of static noise and the smoothing of dynamic signals, extract features from the cleaned data, including multi-dimensional indicators such as timbre feature MFCC and intonation change rate EPSR, generate a standardized feature matrix, provide a unified data format for subsequent calculations, perform dimensionless processing on the parameters in the feature matrix, normalize the dimensions and value ranges of different features, and eliminate scale differences between data.

4. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step three includes: Based on the customer's timbre features in the feature matrix, the timbre adaptation coefficient Zys is obtained after calculation, and the customer's timbre matching score is generated; Based on the customer's intonation characteristics, the intonation matching coefficient Zyt is calculated in combination with the intonation matching formula, the degree of adaptation of the customer's intonation characteristics is dynamically determined, and the intonation model parameters are output; The timbre adaptation score and intonation matching score are combined to dynamically generate a timbre and intonation model; the model includes timbre parameters, intonation curves, and speech speed adjustment suggestions.

5. The method for intelligent conversation management of an outbound call system according to claim 4, characterized in that: The timbre adaptation coefficient Zys is calculated by the following formula: ； In the formula, BGFR represents fundamental frequency, FRMT represents formant frequency, MFCC represents acoustic features, TMBC represents timbre features, and SPRT represents speech rate features. They represent the proportional coefficients of fundamental frequency BGFR, formant frequency FRMT, acoustic feature MFCC, timbre feature TMBC and speech rate feature SPRT respectively; The intonation matching coefficient Zyt is calculated by the following formula: ； In the formula, EMFI represents the emotion recognition feature value, EMTR represents the emotion change rate, PCVT represents the intonation curve characteristics, PDIN represents the pause interval distribution, EPSR represents the sentence end intonation change rate, They respectively represent the proportional coefficients of emotion recognition feature value EMFI, emotion change rate EMTR, intonation curve characteristic PCVT, pause interval distribution PDIN and sentence end intonation change rate EPSR.

6. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step 4 includes: Based on the timbre and intonation model, natural speech is synthesized through TTS technology. The speech content and the intonation model correspond one-to-one to ensure the adaptability and fluency of the output speech. In the speech generation process, the intonation parameters are adjusted in real time, including the fluctuation range of the intonation curve, the rate of change of the tone at the end of the sentence, and the pause rhythm, to ensure that the intonation changes are smooth and conform to the rules of natural conversation. According to the speed parameters in the timbre and intonation model, the playback speed of the output speech is dynamically adjusted to achieve personalized speed matching in different conversation situations.

7. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step five includes: According to the current dialogue context, the semantic matching degree SMAT and dialogue goal achievement rate CVRT scenario parameters are analyzed to determine the specific scenario requirements of the outbound call. The scenario adaptation coefficient Zsc is obtained by combining the speech speed adjustment index SSPR and the feedback time distribution RTDP. The scene adaptation coefficient Zsc is calculated by the following formula: ； In the formula, CVRT represents the dialogue goal achievement rate, SMAT represents the semantic matching degree, CASC represents the customer context perception characteristics, RTDP represents the feedback time distribution, They represent the proportional coefficients of conversation goal achievement rate CVRT, semantic matching degree SMAT, customer context perception characteristic CASC and feedback time distribution RTDP respectively; By comparing the scene adaptation coefficient Zsc with the preset threshold D, the matching degree between the current tone and the scene requirements is evaluated. According to the scene adaptation results, the corresponding tone style is selected and optimized, including three types: enthusiastic, gentle and calm. The scenario adaptation coefficient Zsc>preset threshold D indicates that the current intonation is highly matched with the scenario requirements, the customer accepts the intonation well, the hang-up rate is low, and the dialogue goal achievement rate is high. The intonation curve characteristic PCVT is optimized again to add more natural processing, adjust the intonation slightly upward at the end of the sentence, enhance emotional expression, and optimize the pause interval distribution PDIN to avoid emotional loss caused by fast or slow speech speed; If the scenario adaptation coefficient Zsc ≤ the preset threshold D, it indicates that there is a deviation between the current tone and the scenario requirements. Customers are disgusted with or not used to the tone, the hang-up rate is high, and the conversation completion rate is low. Analyze the feedback data of low-adaptation scenarios, focus on the incompatibility points, change the current tone from enthusiastic to gentle, and reduce the customer's psychological pressure. When the customer is indifferent or resistant, use a calm tone to weaken the emotional intensity of the system voice and avoid irritating the customer.

8. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step six includes: Monitor the outbound call process in real time, collect feedback data on hang-up rate, customer satisfaction, and conversation interruption points, and record the collection results in the feedback database. Iterate and optimize the timbre and intonation matching model through deep learning algorithms; Based on the feedback data, the comprehensive adaptation index Zsa is calculated and compared with the preset adaptation threshold A to determine the effectiveness of the current timbre and intonation matching.

9. The method for intelligent conversation management of an outbound call system according to claim 8, characterized in that: The comprehensive adaptation index Zsa is calculated by the following formula: ； Where Zys represents the timbre adaptation coefficient, Zyt represents the intonation matching coefficient, and Zsc represents the scene adaptation coefficient. They represent the proportional coefficients of the timbre adaptation coefficient Zys, the intonation matching coefficient Zyt, and the scene adaptation coefficient Zsc respectively; By comparing the comprehensive adaptation index Zsa with the preset threshold A, the evaluation strategy solution is obtained: The comprehensive adaptation index Zsa ≥ the preset adaptation threshold A+Δ. The comprehensive adaptation index Zsa is higher than the preset threshold A and exceeds the tolerance range ΔA. The customer feedback data shows good performance, including low hang-up rate, high conversation completion rate, and customer satisfaction at the first level. The timbre and intonation model is fine-tuned to further optimize by increasing the emotional richness of the intonation and enhancing the natural changes of the intonation curve. The preset threshold value A−ΔA≤comprehensive adaptation index Zsa<preset threshold value A+ΔA. The comprehensive adaptation index Zsa is close to the preset threshold value A and fluctuates within the tolerance range ΔA. Customer feedback data shows that some hang-ups or conversation interruptions occur, and customer satisfaction is at the second level. For scenarios with a high hang-up rate, a calm or gentle tone of voice is preferred to avoid strong tone that may cause customer resistance. The comprehensive adaptation index Zsa<preset threshold A−ΔA, the comprehensive adaptation index Zsa is lower than the preset threshold A, lower than the tolerance range ΔA, and the customer feedback data performance is unqualified, including high hang-up rate, frequent conversation interruptions and customer satisfaction lower than normal. Through cluster analysis of hang-up points and reasons for conversation interruptions, the causes of timbre, intonation or scene adaptation problems are located, a simulation scenario is established, and the adaptation effect of the new model is verified through manual intervention.

10. The method for intelligent conversation management of an outbound call system according to claim 1, characterized in that: Step seven includes: Real-time monitoring of the running status of each step, including data collection, feature analysis, speech synthesis and feedback optimization process, recording abnormal events and generating running logs, providing a flexible control interface, allowing users to manually set the timbre, intonation, speech speed and scene threshold parameters, and supporting manual intervention operations in special circumstances; In case of sudden exceptions or system failures, the current strategy is forced to switch to ensure system stability and high availability.

Images