Virtual digital outbound intelligent customer service system based on generative AI

Abstract

The application discloses a virtual digital outbound intelligent customer service system based on generative AI, and relates to the technical field of virtual customer service management.The application solves the technical problems of lack of effective monitoring mechanism for service quality and single encryption processing mode for interactive data.The application monitors the overall state of the intelligent customer service in real time through the calculation of multi-dimensional indexes such as the knowledge anchor point answer rate, semantic ambiguity answer proportion and high-frequency complaint problem correlation rate, so as to facilitate the targeted optimization of the knowledge base, answer template or service process by the management personnel, improve the user service experience and system reliability, generate random modification features for the data packet byte size, dynamically bundle data packets and send them in random order, introduce a random delay transmission strategy, effectively interfere with the analysis of the traffic characteristics by attackers, prevent data eavesdropping and content tampering, and guarantee the safety and privacy of the user interactive information in the transmission process.

Description

Technical Field

[0001] This invention relates to the field of virtual customer service management technology, specifically to a virtual digital outbound intelligent customer service system based on generative AI. Background Technology

[0002] With the rapid development of artificial intelligence technology, virtual digital outbound call intelligent customer service systems have been widely used in multiple fields such as finance, e-commerce, and government affairs.

[0003] Traditional intelligent customer service systems mostly rely on fixed rules or simple machine learning models to achieve human-computer interaction through a pre-set script library. Speech recognition technology mainly uses a combination of traditional acoustic models and language models. In scenarios with complex accents and background noise, the accuracy of speech-to-text conversion is low. At the semantic understanding level, it is difficult to handle complex issues such as semantic ambiguity and context dependence, resulting in a lack of specificity and accuracy in the customer service system's responses.

[0004] Regarding data transmission, existing systems employ simplistic encryption methods for interactive data, making them vulnerable to attackers who can exploit this vulnerability through traffic analysis and packet sniffing to obtain sensitive information. Furthermore, the systems lack effective mechanisms for monitoring their own service quality, failing to promptly detect issues such as unsubstantiated responses and semantic ambiguity, thus hindering the assurance of user experience and service reliability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a virtual digital outbound call intelligent customer service system based on generative AI, which solves the problems of the system lacking an effective monitoring mechanism for its own service quality and having a single method for encrypting interactive data.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a virtual digital outbound call intelligent customer service system based on generative AI, comprising:

[0007] The speech conversion and recognition module is used to convert, recognize, and analyze the interactive information transmitted by the interactive information acquisition module to obtain text semantics. It determines the joint confidence by calculating the confidence of the text semantics and semantic information, and generates interactive response information by matching it with the threshold range. The information is then transmitted to the intelligent customer service comprehensive judgment module and the interactive transmission analysis module.

[0008] The intelligent customer service comprehensive judgment module is used to combine interactive response information and historical interaction records to monitor and analyze multi-dimensional indicators. It calculates the answer rate without knowledge anchors, the proportion of semantically ambiguous answers, and the correlation rate of high-frequency complaint issues, and judges the overall status of intelligent customer service by combining the three factors, generating normal or abnormal status information, and transmitting it to the intelligent management output module.

[0009] The interactive transmission analysis module is used to perform transmission encryption analysis on the acquired interactive response information, split the interactive response information into independent data packets, obtain the characteristics of the data packets, generate random byte segments based on the byte tail values ​​of the data packets, and combine them to obtain combined data packets.

[0010] Next, the combined data packets are bundled and grouped, and transmitted using out-of-order delivery and random delay to generate interactive transmission information, which is then transmitted to the intelligent management output module.

[0011] As a further aspect of the present invention, it also includes an interactive information acquisition module and an intelligent management output module;

[0012] The interactive information acquisition module is used to acquire interactive information, which is voice information, and transmit it to the voice conversion and recognition module.

[0013] The intelligent management output module is used to display normal or abnormal status information and interactive transmission information to the corresponding management personnel.

[0014] As a further aspect of the present invention, the specific method by which the speech conversion and recognition module obtains text semantics is as follows:

[0015] Interactive information is acquired and converted into text information using speech recognition technology. Natural language processing technology is then used to identify the semantics of the obtained text information. Finally, a text similarity algorithm is used to remove duplicate text from the semantics of the text to generate preprocessed semantics.

[0016] The system acquires preprocessed semantics, combines them with non-voice data from customer service scenarios, and aligns the preprocessed semantics with business data using a multimodal model. Simultaneously, it calculates the confidence scores for text semantics and voice information, respectively.

[0017] As a further aspect of the present invention, the specific method by which the speech conversion and recognition module calculates the text semantic confidence score and the speech information confidence score is as follows:

[0018] The semantic confidence score of the text is calculated by processing the audio frame by frame using an ASR model, obtaining the output character probability for each frame, and then calculating the joint product of all character probabilities according to the formula. The semantic confidence score (ASR) of the text is calculated, where N is the number of frames, specifically i = 1, 2, ..., N, P i，j It is the probability of recognizing character j in the i-th frame, and the average of the maximum probabilities in each frame is taken;

[0019] The confidence level of speech information is calculated, and an "acoustic feature-business semantics" matching degree model is constructed. The correlation between speech prosodic features and semantic intent is extracted, and the formula is used to... The matching degree is calculated, and then the confidence degree corresponding to the speech information is calculated according to the formula: confidence degree = matching degree × weight.

[0020] As a further aspect of the present invention, the specific method by which the speech conversion and recognition module generates interactive response information is as follows:

[0021] According to formula C z The joint confidence level C is calculated as α×ASR+(1-α)×P. z Where α is the text semantic confidence weight, P represents the speech information confidence, and the obtained joint confidence C is... z The system matches the data against the corresponding threshold range to determine the appropriate response strategy and generate interactive response information. The specific matching method is as follows:

[0022] If the joint confidence level is C z If the value is ≥0.85, the corresponding response strategy is to directly output the answer;

[0023] If 0.6 ≤ joint confidence level C z If the value is less than 0.85, the corresponding response strategy is to trigger follow-up questions for clarification.

[0024] If the joint confidence level is C z If the value is less than 0.6, the corresponding response strategy is to transfer the request to a human operator or refuse to answer.

[0025] As a further aspect of the present invention, the intelligent customer service comprehensive judgment module calculates the answer rate without knowledge anchors, the proportion of semantically ambiguous answers, and the correlation rate of high-frequency complaint issues in the following specific manner:

[0026] The rate of responses without knowledge anchors is calculated by setting a time period T, acquiring all response texts within that period, and filtering responses with similarity scores below a threshold using a semantic similarity algorithm. These responses are then recorded as responses without knowledge anchors. Calculate the response rate without knowledge anchors and compare it with the preset value of the anchor response rate. If it is lower than the preset value, it is normal; otherwise, it is abnormal.

[0027] The proportion of semantically ambiguous answers is calculated using an ambiguity detection model. The input is the answer text, and the output is the ambiguity probability. If the ambiguity probability is ≥0.5, the answer is considered ambiguous. The number of ambiguous answers is counted, and the result is calculated according to the formula... The percentage of semantically ambiguous answers is calculated and compared with a preset value for the percentage of ambiguous answers. If it is lower than the preset value, it is normal; otherwise, it is abnormal.

[0028] The correlation rate of frequently complained issues is calculated. Based on historical complaint data, a keyword database is constructed by analyzing the top K issues by complaint volume, with the value of K set by the operator. Keyword matching and semantic vector similarity detection are performed on the consultation text to identify related issues, according to the formula... The correlation rate of high-frequency complaint issues is calculated and compared with a preset correlation rate value. If it is higher than the preset value, it is abnormal; otherwise, it is normal.

[0029] As a further aspect of the present invention, the intelligent customer service comprehensive judgment module generates normal or abnormal status information in the following specific way:

[0030] The system analyzes the results of responses without knowledge anchors, the proportion of semantically ambiguous responses, and the correlation rate of frequently asked complaints. If any one of these groups is abnormal, it indicates that the overall state of the intelligent customer service is abnormal, and status adjustment information is generated. Conversely, if all three are normal, it indicates that the overall state of the intelligent customer service is normal, and status normal information is generated.

[0031] As a further aspect of the present invention, the interactive transmission analysis module obtains the combined data packets in the following specific manner:

[0032] The interactive response information is obtained and split into independent data packets, which are labeled as n, where n = 1, 2, ..., m, and m represents the number of data packets. At the same time, the characteristics of the corresponding data packets are obtained, the byte size of the data packets is obtained, and the byte tail value is obtained. Meanwhile, byte segments are randomly generated with double the byte tail value as the standard byte number. The generated byte segments are added to the data packet characteristics to obtain the combined data packets.

[0033] As a further aspect of the present invention, the interactive transmission analysis module generates interactive transmission information in the following specific manner:

[0034] All combined data packets are obtained, and h groups are randomly selected to bundle them to obtain bundled data packets. This process is repeated to obtain all bundled data packets. The obtained bundled data packets are sent out of order, and periodic interval transmission is performed using a random delay method. Interactive transmission information is generated.

[0035] This invention provides a virtual digital outbound call intelligent customer service system based on generative AI. Compared with existing technologies, it has the following advantages:

[0036] This invention utilizes an industrial-grade ASR model based on Transformer and pre-trained NLP models such as BERT, combined with multimodal data and CLIP-like multimodal frameworks, to achieve high-precision conversion and semantic understanding of speech information, enabling accurate recognition of user intent. By calculating the confidence scores of text semantics and speech information, and making a comprehensive judgment, a response matching strategy based on the joint confidence score is applied to ensure the accuracy and relevance of interactive responses.

[0037] This invention generates randomly modified characteristics for data packet byte size, dynamically bundles data packets, and sends them out of order. It also introduces a random delay transmission strategy to effectively interfere with attackers' analysis of traffic characteristics, prevent data eavesdropping and content tampering, and ensure the security and privacy of user interaction information during transmission. By calculating multi-dimensional indicators such as the rate of responses without knowledge anchors, the proportion of semantically ambiguous responses, and the correlation rate of high-frequency complaint issues, the overall status of the intelligent customer service is monitored in real time. Once abnormal indicators are detected, status adjustment information is generated promptly, allowing administrators to optimize the knowledge base, response templates, or service processes in a targeted manner, thereby improving user service experience and system reliability. Attached Figure Description

[0038] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] Please see Figure 1 This application provides a virtual digital outbound call intelligent customer service system based on generative AI, including: an interactive information collection module, a voice conversion and recognition module, an intelligent customer service comprehensive judgment module, an interactive transmission analysis module, and an intelligent management output module, and combined with... Figure 1 It can be seen that the above functional modules are connected electrically in one direction.

[0042] The interactive information acquisition module is used to acquire interactive information and transmit it to the speech conversion and recognition module. The interactive information acquired here is mainly speech information.

[0043] The speech conversion and recognition module is used to convert and analyze the obtained interactive information. It converts the speech information into text information through speech recognition technology. Relying on mature ASR models (such as industrial-grade solutions based on Transformer), it converts customer service calls and user consultation voices into text frame by frame. At the same time, it uses natural language processing (NLP) technology to recognize the text information to obtain text semantics. It uses pre-trained models such as BERT to parse the text semantics, identify the core intent, and remove duplicate corpora in the text semantics through text similarity algorithms to generate pre-processed semantics.

[0044] Next, preprocessed semantics are obtained, combined with non-voice data such as text logs and user profiles from customer service scenarios, and the preprocessed semantics are aligned with business data through a multimodal model. At the same time, the confidence scores of text semantics and voice information are calculated separately.

[0045] Integrate non-voice data such as customer service text logs (historical work orders, FAQ database) and user profiles (membership level, purchase history) to construct a multimodal semantic alignment model:

[0046] Data association: Using the user's order number as the association key, match preprocessed semantics with business data (such as the user's historical return count and whether the product meets the return and exchange policy).

[0047] Multimodal calibration: By using the CLIP-like multimodal framework, align semantic text with business data features (such as mapping the semantics of "return" with the knowledge base vector of "after-sales process") to ensure that semantic understanding conforms to customer service business rules.

[0048] The semantic confidence of the text is calculated by segmenting the audio into frames using an ASR (Speech-to-Text) model, obtaining the probability of each output character in each frame, and then calculating the joint product of all character probabilities according to the formula. The semantic confidence score (ASR) of the text is calculated, where N is the number of frames, specifically i = 1, 2, ..., N, P i，j It is the probability of recognizing character j in the i-th frame, and the average of the maximum probabilities in each frame is taken;

[0049] For example, if a speech segment is divided into 100 frames, and the average probability of the largest character in each frame is 0.92, then the ASR confidence level is 0.92, representing the basic reliability of speech-to-text conversion.

[0050] The confidence level of speech information is calculated, and an "acoustic feature-business semantics" matching degree model is constructed. The correlation between speech prosodic features (speech rate, pauses, emotional tone) and semantic intent is extracted, and then calculated according to the formula... The matching degree is calculated, and then the confidence degree corresponding to the speech information is calculated according to the formula: confidence degree = matching degree × weight.

[0051] For example, if a user angrily requests a refund, and the acoustic features of "urgent and high frequency" in the voice match the semantics of "emergency after-sales service" with a score of 0.85 and a business weight of 0.7, then the voice confidence score is 0.85 × 0.7 = 0.595.

[0052] The obtained text semantic confidence score (ASR) and speech information confidence score are combined and calculated according to formula C. z The joint confidence level C is calculated as α×ASR+(1-α)×P. z Where α is the text semantic confidence weight, P represents the speech information confidence, and the obtained joint confidence C is... zThe system matches the data against the corresponding threshold range to determine the appropriate response strategy and generate interactive response information. The specific matching method is as follows:

[0053] If the joint confidence level is C z If the value is ≥0.85, the corresponding response strategy is to directly output the answer;

[0054] If 0.6 ≤ joint confidence level C z If the value is less than 0.85, the corresponding response strategy is to trigger follow-up questions for clarification.

[0055] If the joint confidence level is C z If the value is less than 0.6, the corresponding response strategy is to transfer the request to a human operator or refuse to answer.

[0056] Simultaneously, the generated interactive response information is transmitted to the intelligent customer service comprehensive judgment module and the interactive transmission analysis module.

[0057] For example, if a user says "The charging port of my phone is broken", after the speech is converted to text, the BERT model recognizes the intent as "after-sales repair" (probability 0.78) and the entity as "phone charging port" (probability 0.92).

[0058] After sampling 50 times using MCDropout, the variance of the intent probability was 0.05 (low variance, improved confidence).

[0059] The knowledge graph verifies that "damaged charging port" is covered under warranty, with a 100% match rate and a corrected confidence level of 0.78 × 1.1 = 0.858. Based on the corresponding matching criteria, the response strategy can be determined as "directly replying with the repair process".

[0060] The intelligent customer service comprehensive judgment module is used to analyze the overall status of intelligent customer service based on the obtained interaction response information, obtain the corresponding historical interaction records of intelligent customer service, and perform multi-dimensional indicator monitoring and analysis based on the historical interaction records. The multi-dimensional indicators specifically include the answer rate without knowledge anchors, the proportion of semantically ambiguous answers, and the correlation rate of high-frequency complaint issues.

[0061] The rate of responses without knowledge anchors is calculated by retrieving all response texts within a time period T from historical interaction records. The specific value of time period T is set by the operator. Semantic similarity algorithms (such as BERT + cosine similarity) are used to match these responses with standard answers in the knowledge base. A threshold is set (e.g., similarity ≥ 0.7 is considered as having an anchor), and responses with similarity < 0.7 are considered "responses without knowledge anchors." Then, according to the formula... Calculate the response rate without knowledge anchors;

[0062] At the same time, the obtained no-knowledge anchor response rate is compared with the preset value of the anchor response rate. The specific value of the preset value of the anchor response rate is set by the operator. If the no-knowledge anchor response rate is less than the preset value of the anchor response rate, it means that the no-knowledge anchor response rate is normal. Otherwise, it means that the no-knowledge anchor response rate is abnormal.

[0063] For example, a customer service system processes 1,000 inquiries per day. Of these, 850 answers have a similarity of ≥0.7 with the answers in the knowledge base, and 150 do not match. The corresponding answer rate without knowledge anchors is calculated to be 15%. Further analysis reveals that the preset value for the anchor answer rate is 5%. Therefore, the comparison shows that the current answer rate without knowledge anchors is abnormal.

[0064] The proportion of semantically ambiguous answers is calculated using an ambiguity detection model (such as an LSTM-based semantic multi-solution classifier). The input is the answer text, and the output is whether it is ambiguous (a probability ≥ 0.5 indicates ambiguity). Ambiguity detection is performed on each answer, the number of ambiguous answers is counted, and the result is calculated according to the formula... The percentage of semantically ambiguous answers is calculated and compared with a preset value for the percentage of ambiguous answers. If the percentage of semantically ambiguous answers is greater than the preset value, it indicates that the percentage of semantically ambiguous answers is abnormal; otherwise, it indicates that the percentage of semantically ambiguous answers is normal.

[0065] For example, if a customer service representative answers, "Your application will take about 3 business days to process," the word "about" might be misinterpreted by users as "within 3 days" or "more than 3 days," and the model would judge it as ambiguous.

[0066] If 50 ambiguous answers are detected out of 1000, accounting for 5%, and the preset percentage of ambiguous answers is 8%, then after comparison, it is known that the percentage of semantically ambiguous answers is normal.

[0067] The correlation rate of high-frequency complaint issues is calculated. Based on historical complaint data, the top K issues in terms of complaint volume (such as "slow refund arrival" and "product quality issues") are statistically analyzed to form a keyword / semantic vector library, where K represents the ranking, and the specific value is set by the operator. Consultation texts containing keywords from the complaint library (such as "refund" and "quality") are initially marked as related. Implicit correlations (such as the semantic correlation between "funds not received" and "slow refund arrival") are identified through topic models (such as LDA) or semantic vector similarity. Then, the correlation is further calculated according to the formula... The correlation rate of high-frequency complaint issues was calculated and compared with the preset correlation rate value.

[0068] If the correlation rate of frequently complained issues is greater than the preset correlation rate, it indicates that the correlation rate of frequently complained issues is abnormal; conversely, if the correlation rate of frequently complained issues is less than the preset correlation rate, it indicates that the correlation rate of frequently complained issues is normal.

[0069] For example, a frequently complained issue in the past included "refund delay" (keywords: refund, delay, arrival of funds). One inquiry text was "I submitted my refund application 3 days ago and still haven't received the money." The issue was identified as related through keyword and semantic matching.

[0070] If 80 out of 1000 inquiries are related to frequently asked questions, the correlation rate for frequently asked questions is 8%, and the preset correlation rate is 10%. By comparison, it can be seen that the correlation rate for frequently asked questions is less than the preset correlation rate, then the correlation rate for frequently asked questions is normal.

[0071] The system analyzes the results of responses without knowledge anchors, the proportion of semantically ambiguous responses, and the correlation rate of high-frequency complaint issues. If any of these anomalies are found, the overall state of the intelligent customer service is considered abnormal, and status adjustment information is generated. Conversely, if all of these are normal, the overall state of the intelligent customer service is considered normal, and status normal information is generated. Both are then transmitted to the intelligent management output module.

[0072] The intelligent management output module is used to display the acquired status information (normal or abnormal) to the corresponding management personnel.

[0073] Example 2

[0074] As a second embodiment of the present invention, the focus is on combining the implementation processes of the first and second embodiments.

[0075] The interactive transmission analysis module acquires the interactive response information transmitted by the semantic conversion and recognition module, performs encrypted transmission analysis on it, and splits the interactive response information into independent data packets, labeled n (n = 1, 2, ..., m), where m represents the number of data packets. It also acquires the corresponding data packet features, including timestamps or protocol header fields. Key features of each data packet are extracted, including the timestamp (recording the data packet generation time) and protocol header fields (such as source port, destination port, and sequence number in TCP / UDP protocols). Taking a TCP-based data packet as an example, its source port 50001, destination port 8080, and sequence number 123456 are extracted. At the sending end, randomly modified features are generated based on the data packet's byte size, and the generation method is as follows:

[0076] Obtain the size of the data packet in bytes and its byte tail value. At the same time, randomly generate a byte segment with a standard byte length of twice the byte tail value. Add the generated byte segment to the data packet characteristics to obtain a combined data packet. Specifically, if the tail value is 4, the generated byte segment length is 8; if the tail value is 9, the generated byte segment length is 18; and if the tail value is 0, a byte segment of length 20 is always generated. For example, if the original data packet timestamp is 2025-06-17 10:30:00, after merging, it becomes 2025-06-17 10:30:00 + random byte segment.

[0077] All combined data packets are acquired, and h groups are randomly selected to bundle them into bundled data packets. The specific value of h is set by the operator and is within the range of 2-9. This process is repeated to acquire all bundled data packets. The bundled data packets are then sent out of order, and periodic interval transmission is performed using a random delay. Random delays are inserted at the session layer: a random wait of 0-50ms is performed before sending (to balance delay and call experience), which disrupts the "periodic packet sending" characteristic. Grouped out-of-order transmission is used: continuous voice packets are divided into groups (e.g., 5 packets per group), and the sending order within each group is randomly adjusted. The receiving end restores the order and generates interactive transmission information. Specifically, the receiving end generates corresponding transmission decryption rules, and the obtained interactive transmission information is decrypted in reverse according to the transmission decryption rules to obtain the corresponding interactive response information, which is then transmitted to the intelligent management output module.

[0078] Assuming there are 10 combined data packets, and h=3, the 1st, 2nd, and 3rd combined data packets are bundled into one group, the 4th, 5th, and 6th are bundled into another group, and so on. The original order is 1, 2, 3, but when sent, it becomes 3, 1, 2. The first bundled data packet is sent with a delay of 12ms, and the second is sent with a delay of 35ms, which disrupts the timing characteristics of the traffic.

[0079] The intelligent management output module is used to manage the transmission of interactive response information based on the generated interactive transmission information.

[0080] Example 3

[0081] As a third embodiment of the present invention, the focus is on combining the implementation processes of the first and second embodiments.

[0082] The data in the above formulas are all calculated using numerical values, without substituting the units of the parameters. In addition, the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0083] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A virtual digital outbound intelligent customer service system based on generative AI, characterized in that: include: The speech-to-speech recognition module is used to convert, recognize, and analyze the interactive information transmitted by the interactive information acquisition module to obtain text semantics. It then determines the joint confidence score by calculating the confidence scores of the text semantics and the semantic information, according to the formula... The joint confidence level C was calculated. z ,in The text semantic confidence weight is P, which represents the confidence of the speech semantic information. At the same time, it is matched with the threshold range to generate interactive response information and transmitted to the intelligent customer service comprehensive judgment module and the interactive transmission analysis module. The intelligent customer service comprehensive judgment module is used to monitor and analyze multi-dimensional indicators by combining interactive response information and historical interaction records. It calculates the rate of responses without knowledge anchors, the proportion of semantically ambiguous responses, and the correlation rate of high-frequency complaint issues, and combines the three to judge the overall status of intelligent customer service, generating normal or abnormal status information, and transmitting it to the intelligent management output module. The specific processing method is as follows: The rate of responses without knowledge anchors is calculated by setting a time period T, acquiring all response texts within that period, and filtering responses with similarity scores below a threshold using a semantic similarity algorithm. These responses are then recorded as responses without knowledge anchors. Calculate the response rate without knowledge anchors and compare it with the preset value of anchor response rate. If it is lower than the preset value, it is normal; otherwise, it is abnormal. The proportion of semantically ambiguous answers is calculated using an ambiguity detection model. The input is the answer text, and the output is the ambiguity probability. If the ambiguity probability is ≥0.5, the answer is considered ambiguous. The number of ambiguous answers is counted, and the result is calculated according to the formula... The percentage of semantically ambiguous answers is calculated and compared with a preset value for the percentage of ambiguous answers. If it is lower than the preset value, it is normal; otherwise, it is abnormal. The correlation rate of frequently complained issues is calculated. Based on historical complaint data, a keyword database is constructed by statistically analyzing the top K issues in terms of complaint volume, with the value of K set by the operator. Keyword matching and semantic vector similarity detection are performed on the consultation text to identify related issues, according to the formula... The correlation rate of high-frequency complaint issues is calculated and compared with a preset correlation rate value. If it is higher than the preset value, it is abnormal; otherwise, it is normal. The interactive transmission analysis module is used to perform transmission encryption analysis on the acquired interactive response information, split the interactive response information into independent data packets, obtain the characteristics of the data packets, generate random byte segments based on the byte tail values ​​of the data packets, and combine them to obtain combined data packets. Next, the combined data packets are bundled and grouped, and transmitted using out-of-order delivery and random delay to generate interactive transmission information, which is then transmitted to the intelligent management output module.

2. The virtual digital outbound intelligent customer service system based on generative AI according to claim 1, characterized in that, It also includes an interactive information acquisition module and an intelligent management output module; The interactive information acquisition module is used to acquire interactive information, which is voice information, and transmit it to the voice conversion and recognition module. The intelligent management output module is used to display normal or abnormal status information and interactive transmission information to the corresponding management personnel.

3. The virtual digital outbound intelligent customer service system based on generative AI according to claim 1, characterized in that, The specific method by which the speech conversion and recognition module obtains the text semantics is as follows: Interactive information is acquired and converted into text information using speech recognition technology. Natural language processing technology is then used to identify the semantics of the obtained text information. Finally, a text similarity algorithm is used to remove duplicate text from the semantics of the text to generate preprocessed semantics. The system acquires preprocessed semantics, combines them with non-voice data from customer service scenarios, and aligns the preprocessed semantics with business data using a multimodal model. It also calculates the confidence scores for text semantics and voice semantic information separately.

4. The virtual digital outbound intelligent customer service system based on generative AI according to claim 3, characterized in that, The specific method by which the speech conversion and recognition module calculates the text semantic confidence score and the speech semantic information confidence score is as follows: The semantic confidence score of the text is calculated by processing the audio frame by frame using an ASR model, obtaining the output character probability for each frame, and then calculating the joint product of all character probabilities according to the formula. The semantic confidence score (ASR) of the text is calculated, where N is the number of frames, specifically i = 1, 2, ..., N, and P... i，j It is the probability of recognizing character j in the i-th frame, and the average of the maximum probabilities in each frame is taken; The confidence level of speech semantic information is calculated, and an "acoustic feature-business semantic" matching degree model is constructed. The correlation between speech prosodic features and semantic intent is extracted, and the formula is used to calculate the confidence level of speech semantic information. The matching degree is calculated, and then the confidence degree corresponding to the speech information is calculated according to the formula: confidence degree = matching degree × weight.

5. The virtual digital outbound intelligent customer service system based on generative AI according to claim 1, characterized in that, The specific method by which the speech conversion and recognition module generates interactive response information is as follows: Simultaneously, the obtained joint confidence level C will be... z The system matches the data against the corresponding threshold range to determine the appropriate response strategy and generate interactive response information. The specific matching method is as follows: If the joint confidence level is C z If the value is ≥0.85, the corresponding response strategy is to directly output the answer; If 0.6 ≤ joint confidence level C z If the value is less than 0.85, the corresponding response strategy is to trigger follow-up questions for clarification. If the joint confidence level is C z If the value is less than 0.6, the corresponding response strategy is to transfer the request to a human operator or refuse to answer.

6. The virtual digital outbound intelligent customer service system based on generative AI according to claim 1, characterized in that, The intelligent customer service comprehensive judgment module generates normal or abnormal status information in the following specific way: The system analyzes the results of responses without knowledge anchors, the proportion of semantically ambiguous responses, and the correlation rate of frequently asked complaints. If any one of these groups is abnormal, it indicates that the overall state of the intelligent customer service is abnormal, and status adjustment information is generated. Conversely, if all three are normal, it indicates that the overall state of the intelligent customer service is normal, and status normal information is generated.

7. The virtual digital outbound intelligent customer service system based on generative AI according to claim 1, characterized in that, The specific method by which the interactive transmission analysis module obtains the combined data packets is as follows: The interactive response information is obtained and split into independent data packets, which are labeled as n, where n = 1, 2, ..., m, and m represents the number of data packets. The characteristics of the corresponding data packets are obtained, the byte size of the data packets is obtained, and the byte tail value is obtained. At the same time, byte segments are randomly generated with double the byte tail value as the standard byte number. The generated byte segments are added to the data packet characteristics to obtain the combined data packets.

8. The virtual digital outbound intelligent customer service system based on generative AI according to claim 1, characterized in that, The specific method by which the interactive transmission analysis module generates interactive transmission information is as follows: All combined data packets are obtained, and h groups are randomly selected to bundle them to obtain bundled data packets. This process is repeated to obtain all bundled data packets. The obtained bundled data packets are sent out of order, and periodic interval transmission is performed using a random delay method. Interactive transmission information is generated.

Images