AI voice customer service platform

Abstract

The invention provides an AI voice customer service platform method, which belongs to the technical field of voice processing, and specifically comprises the following steps: a voice acquisition and preprocessing module adopts an advanced noise reduction and endpoint detection algorithm; the speech recognition module applies a leading-edge deep learning architecture to improve the recognition precision; the natural language processing module deeply analyzes semantics and intentions; the knowledge base management module has an intelligent retrieval and updating mechanism; the dialogue management module accurately controls multiple rounds of dialogue processes and manual switching decisions; and the answer generation module realizes intelligent screening and fusion of answers. Through cooperative operation of all the modules, the customer service efficiency and the user satisfaction degree are greatly improved, the existing technical problems are solved, remarkable innovation and practicability are achieved, and the accuracy of customer service answering processing is improved.

Description

Technical Field

[0001] This invention belongs to the field of voice processing technology, and in particular relates to an AI voice customer service platform. Background Technology

[0002] In today's digital age, the importance of customer service to businesses is self-evident. With the development of artificial intelligence technology, AI voice customer service systems are widely used in various enterprises to improve customer service efficiency and reduce labor costs. Similar technical solutions are presented in invention patent applications CN112509574B "An Investment Consulting Service System Based on Big Data" and CN109413286A "An Intelligent Customer Service Voice Response System and Method".

[0003] In CN109413286A, "An Intelligent Customer Service Voice Response System and Method," the voice interaction service platform transmits text to an intelligent interactive AI. The intelligent interactive AI performs semantic analysis and sends the text to a knowledge base for matching. The result is then returned to the intelligent interactive AI, which sends the received text to a speech synthesis module. The speech synthesis module converts the text into a response speech and returns it to the intelligent voice interaction service platform via the intelligent interactive AI. Finally, the intelligent voice interaction service platform returns the response to the user, improving the accuracy and reliability of the interaction processing. However, the above technical solution has the following technical problems: In the process of generating answers, popularity or similarity is often used to retrieve answers from the knowledge base. Specifically, for different question types, if there are answers in the knowledge base that are highly similar to multiple users' questions, but at the same time their adoption popularity is low, it means that they no longer match the users' needs. Therefore, if the knowledge base update plan cannot be determined based on the interference of similar answers with other answers under different question types, and answers with high similarity but low adoption popularity are not deleted in a timely and effective manner, it may affect the efficiency of answer filtering and processing to a certain extent.

[0004] To address the aforementioned technical issues, this application specifically provides an AI voice customer service platform. Summary of the Invention

[0005] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, this application provides an AI voice customer service platform, which includes: The system includes a speech recognition module, a natural language processing module, a knowledge base management module, an answer generation module, a human customer service transfer module, and a speech synthesis module. The speech recognition module is responsible for converting the user's pre-processed speech signal into text, and the natural language processing module is responsible for transmitting the intent and text features identified by the text to the knowledge base management module and the human customer service transfer module through the semantic analysis model. The knowledge base management module is responsible for combining a hybrid retrieval strategy of keyword matching and similarity calculation, using text features to filter relevant answers, determining the recommendation processing strategy for relevant answers based on the distribution of relevant answers in different user questions, the number of relevant answers for user questions, and usage, and determining the deletion processing method for relevant answers based on the recommendation processing strategy and the interference in different user questions. The answer generation module is responsible for filtering answers based on the similarity calculation results of recommended related answers, popularity ranking, and recommendation processing strategies. The speech synthesis module uses the filtered answers and uses speech synthesis technology to generate corresponding audio signals, and uses the audio signals for automatic response processing. The human customer service transfer module is responsible for transferring users to available human customer service agents.

[0006] The beneficial effects of this invention are as follows: Improved speech recognition accuracy: Advanced speech acquisition and preprocessing technologies, along with cutting-edge speech recognition models, effectively reduce the impact of environmental noise, accents, and other factors on recognition, significantly improving the accuracy of speech recognition and making information transmission more accurate.

[0007] Deepening semantic understanding: The application of the BERT model-based multi-task learning architecture in the natural language processing module can deeply understand the user's complex semantics and implicit intentions, accurately grasp user needs, and provide more targeted answers.

[0008] Ensuring the timeliness of the knowledge base: The intelligent retrieval and update mechanism enables the knowledge base to keep up with industry developments and business changes in real time, providing users with the latest knowledge and answers in a timely manner, thereby enhancing the system's usability and adaptability.

[0009] Improved efficiency of human customer service transfer: The optimized dialogue management module and efficient human customer service transfer module have greatly simplified the transfer process, shortened the waiting time, increased the transfer success rate, and improved overall service efficiency and user satisfaction.

[0010] By utilizing recommendation processing strategies and considering interference in different user questions, a method for deleting related answers is determined. This avoids the technical problem of low efficiency in answer filtering caused by answers being related in multiple question types, especially for question types with a large number of related answers. Targeted deletion of related answers reduces storage space requirements and improves answer filtering efficiency without affecting the accuracy of answering user questions.

[0011] In this invention, answers are selected based on the similarity calculation results of recommended related answers, popularity ranking, and recommendation processing strategies, which ensures the accuracy of related answer recommendation processing. While considering similarity, it also achieves the selection of related answers with high popularity, i.e., high usage frequency, thereby improving the accuracy of answer recommendation processing.

[0012] Furthermore, it also includes a user interaction interface module, which is responsible for receiving voice or text information input by users through various terminal devices.

[0013] Furthermore, if the user interaction interface module receives voice information, it directly passes it to the voice recognition module; if the user interaction interface module receives text information, it skips the voice recognition module and directly enters the natural language processing module.

[0014] Furthermore, it also includes a voice acquisition and preprocessing module, which is responsible for acquiring user voice signals using the device's built-in microphone or external audio input device, converting analog signals into digital signals, and preprocessing the digital signals.

[0015] It is understood that the preprocessing includes noise reduction, endpoint detection, and normalization.

[0016] Furthermore, the user's preprocessed speech signal is converted into text, specifically including: Deep learning algorithms are used to convert the preprocessed speech signal into a phoneme sequence; By combining a language model, the phoneme sequence is decoded and converted into text form.

[0017] Furthermore, the speech model employs an N-Gram model or a neural network-based language model.

[0018] Furthermore, the semantic analysis model is constructed using a BERT pre-trained model.

[0019] Furthermore, the text identifies intent and text features, specifically including: In the lexical analysis stage, a model combining a bidirectional long short-term memory network and a conditional random field is used. BiLSTM can capture the contextual information of words in the text, while CRF uses global information in the text to optimize the word segmentation and part-of-speech tagging results to obtain the lexical analysis results. Using a graph neural network (GNN)-based approach, each word in the text is treated as a node. By constructing a syntactic relationship graph between words, the deep grammatical structure of sentences is mined, and the subject, verb, object, attributive, adverbial, and complement components of the sentences are analyzed to generate a syntactic tree. Based on the BERT pre-trained model, text features are extracted using a syntax tree. The extracted text features are then mapped to different intent categories through a fully connected layer, thus obtaining the intent category mapped to the text features.

[0020] Furthermore, the semantic similarity and popularity of the related answers are used to determine the selected answers, specifically including: First, semantic similarity is used to initially rank related answers. Then, for related answers with similar semantic similarity, they are ranked a second time based on popularity. The answer with the highest ranking is selected as the final answer.

[0021] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0023] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart of an AI voice customer service platform; Figure 2 It is a flowchart for identifying intent and text features from text; Figure 3 This is a flowchart for filtering relevant answers using text features; Figure 4 It is a flowchart that uses the semantic similarity and popularity of related answers to determine the selection of answers. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0026] Example 1 like Figure 1 As shown, this application provides an AI voice customer service platform, specifically including: The system includes a speech recognition module, a natural language processing module, a knowledge base management module, an answer generation module, a human customer service transfer module, and a speech synthesis module. The speech recognition module is responsible for converting the user's pre-processed speech signal into text; Specifically, this includes: Voice acquisition: using the device's built-in microphone or an external audio input device to acquire the user's voice signal and convert the analog signal into a digital signal; Preprocessing: Noise reduction: Employs signal filtering algorithms to remove environmental noise, current noise, and other interference, improving speech clarity. Endpoint detection: Identifies the start and end points of speech, removing invalid silence segments and reducing the amount of data for subsequent processing. Normalization: Normalizes the amplitude of the speech signal to ensure that the speech signals input by different users are within the same magnitude range, facilitating subsequent recognition. The processed speech data is then transmitted to the speech recognition module. Advanced noise reduction algorithms, such as wavelet transform-based adaptive filtering algorithms, are employed to effectively remove interference signals such as environmental noise and current noise. At the same time, a dual-threshold energy detection method is used for endpoint detection to accurately identify the start and end positions of speech. The speech signal amplitude is normalized to provide high-quality data for subsequent speech recognition.

[0027] The speech recognition module includes: Acoustic model processing: Deep learning algorithms (such as models based on the Transformer architecture) are used to convert the preprocessed speech signal into a phoneme sequence. The acoustic model is trained on a large amount of speech data to learn the correspondence between acoustic features and phonemes in the speech signal.

[0028] Language model processing: The phoneme sequence is decoded using a language model (such as an N-Gram model or a neural network-based language model) and converted into text. The language model considers the probability of contextual associations between words, correcting and optimizing the output of the acoustic model to improve recognition accuracy. The recognized text data is then passed to the natural language processing module.

[0029] The natural language processing module is responsible for transmitting the intent and text features identified from the text to the knowledge base management module and the human customer service transfer module through the semantic analysis model. Specifically, this includes: Lexical analysis: segmenting the input text into words, dividing continuous sentences into individual words, and labeling the part of speech of each word (such as noun, verb, adjective, etc.).

[0030] Syntactic analysis: Analyzes the grammatical structure of a sentence, identifies its subject, verb, object, modifier, complement, etc., and constructs a syntactic tree to help understand the overall structure and semantic relationships of the sentence.

[0031] Semantic understanding and intent recognition: Through semantic analysis models (such as deep learning-based semantic representation models), the deeper meaning of text is understood, and combined with a pre-trained intent classification model, the user's intent is determined, such as querying information, handling business, or making complaints and suggestions. The identified intent and related text features are then passed to the knowledge base management module and the dialogue management module.

[0032] The Natural Language Processing (NLP) module leverages a multi-task learning architecture based on the BERT model to perform comprehensive and in-depth analysis of the text after speech recognition. In the lexical analysis stage, a model combining a Bidirectional Long Short-Term Memory (BiLSTM) network and a Conditional Random Field (CRF) is employed. BiLSTM captures the contextual information of words in the text, effectively addressing the difficulty of traditional Recurrent Neural Networks (RNNs) in handling long-distance dependencies; CRF utilizes global information within the text to optimize word segmentation and part-of-speech tagging results, significantly improving the accuracy of lexical analysis.

[0033] Syntactic analysis employs a graph neural network (GNN)-based approach, treating each word in the text as a node. By constructing a syntactic relationship graph between words and leveraging the powerful graph structure learning capabilities of GNNs, the deep grammatical structure of sentences is mined, and components such as subject, verb, object, attributive, adverbial, and complement are analyzed to generate an accurate syntactic tree.

[0034] Semantic understanding and intent recognition are core functions of the natural language processing module. Leveraging the powerful semantic representation capabilities of the BERT pre-trained model, semantic features of text are extracted. BERT, through a bidirectional Transformer architecture, is pre-trained on large-scale unsupervised text to learn general language knowledge. In practical applications, for specific tasks in the customer service domain, a multi-task learning approach is adopted, jointly training tasks such as intent classification and semantic matching. In the intent classification task, fully connected layers map the text features extracted by BERT to different intent categories; in the semantic matching task, the similarity between the user's question and a pre-defined semantic template is calculated to accurately determine the user's intent.

[0035] The knowledge base management module is responsible for combining a hybrid retrieval strategy of keyword matching and similarity calculation, using text features to filter relevant answers, determining the recommendation processing strategy for relevant answers based on the distribution of relevant answers in different user questions, the number of relevant answers for user questions, and usage, and determining the deletion processing method for relevant answers based on the recommendation processing strategy and the interference in different user questions. Knowledge storage: Storing various types of business knowledge, frequently asked questions (FAQs), product information, etc., in a structured or unstructured manner. Data can be stored in relational databases (such as MySQL), non-relational databases (such as MongoDB), or knowledge graphs.

[0036] Knowledge Retrieval: Based on the user intent and text features transmitted by the natural language processing module, a search is performed in the knowledge base. Methods such as keyword matching and semantic similarity calculation are used to find answers and knowledge related to the user's question. If matching knowledge is found, it is passed to the answer generation module; if not found, feedback is sent to the dialogue management module, which may trigger further processing, such as guiding the user to provide additional information or transferring them to human customer service.

[0037] The knowledge base management module employs a hybrid retrieval strategy combining keyword matching and cosine similarity calculation to quickly and accurately find answers relevant to user questions within a vast amount of knowledge entries. Keyword matching is the first step in the retrieval process. After segmenting the user question into words, inverted index technology is used to quickly locate knowledge entries containing relevant keywords within the knowledge base. The inverted index uses keywords as its index, recording the position of each keyword in various knowledge entries, significantly improving retrieval efficiency.

[0038] For candidate knowledge items selected through keyword matching, cosine similarity calculation is used for precise matching. The text of the user question and candidate knowledge items is represented as vectors, typically using a bag-of-words model or a deep learning-based word embedding model (such as Word2Vec or GloVe) to convert the text into vector form. Cosine similarity measures the degree of similarity between two vectors by calculating the cosine of the angle between them. The closer the cosine value is to 1, the higher the similarity between the two texts. By setting an appropriate cosine similarity threshold, knowledge items that highly match the user question are selected, providing accurate knowledge support for the answer generation module.

[0039] The answer generation module is responsible for filtering answers based on the similarity calculation results of recommended related answers, popularity ranking, and recommendation processing strategies; Answer filtering and organization: From multiple possible answers obtained from the knowledge base management module, the most suitable answer is selected based on certain strategies (such as similarity priority, popularity sorting, etc.). The answer is then formatted and polished to make it clearer, more natural, and in line with the user's language habits.

[0040] Answer fusion and expansion: If necessary, multiple related answers can be merged, or the answers can be appropriately expanded based on the context of the user's question to supplement necessary information and ensure the completeness and accuracy of the response. The generated text answer is then passed to the speech synthesis module (if a voice response is required) or the user interaction interface module (if a text response is required).

[0041] The answer generation module filters answers based on similarity priority and popularity to ensure that the most suitable and useful response is provided to the user. Regarding similarity priority, the semantic similarity between the user's question and knowledge entries in the knowledge base is first calculated. Based on the cosine similarity calculation method described earlier, the retrieved knowledge entries are sorted from highest to lowest similarity. The knowledge entry with the highest similarity is considered the answer that best matches the user's question.

[0042] Besides similarity, the popularity of answers is also an important criterion for selection. Popularity is measured by the historical usage frequency of knowledge items. During system operation, the number of times each knowledge item has been used to answer user questions is recorded. When multiple knowledge items have similarities, the knowledge item with the higher historical usage frequency is prioritized. This is because knowledge items with higher popularity are more likely to be answers to questions that users are generally concerned about, and therefore have higher reference value.

[0043] In the actual screening process, if recommendation strategies are not considered, knowledge items are first initially sorted according to similarity. Then, for items with the same similarity, they are sorted again according to popularity. In this way, by comprehensively considering both similarity and popularity, the answers that best meet the user's needs are selected from the knowledge base. The answers are then properly organized and optimized, such as adjusting the format and adding information, and finally, high-quality answers are provided to the user.

[0044] The speech synthesis module uses the filtered answers and generates corresponding audio signals using speech synthesis technology, and then uses the audio signals for automatic response processing. Text analysis: Perform prosodic analysis on the input text to determine the stress, intonation, pauses, and other information of each word, making the synthesized speech more natural and fluent.

[0045] Speech Synthesis: Using speech synthesis technology (such as a deep learning-based TTS model), corresponding audio signals are generated based on text content and prosodic information. During the synthesis process, the model learns the features of a large amount of human speech data, simulating the timbre, pitch, etc. of human voice. The generated audio stream is then passed to the user interaction interface module for playback by the user.

[0046] The human customer service transfer module is responsible for transferring users to available human customer service agents.

[0047] Transfer preparation: When the dialogue management module triggers a transfer to human customer service, it collects the user's basic information (such as user ID, business type), dialogue history, current problem description, etc., and organizes this information into a format that makes it easy for human customer service to quickly understand the situation.

[0048] Transfer Execution: Users are transferred to available human customer service agents via internal communication systems (such as instant messaging tools or customer service platform transfer functions). Simultaneously, the user's relevant information is displayed on the human customer service interface to help the agent quickly initiate the conversation.

[0049] Service Recording and Feedback: During interactions between human customer service representatives and users, the conversation content and processing results are recorded. After the service is completed, the human customer service representative's processing results are fed back to the system. The system can then use this feedback to update its knowledge base, optimize its intent recognition model, and improve the quality of subsequent automated services.

[0050] Furthermore, it also includes a user interaction interface module, which is responsible for receiving voice or text information input by users through various terminal devices.

[0051] Furthermore, if the user interaction interface module receives voice information, it directly passes it to the voice recognition module; if the user interaction interface module receives text information, it skips the voice recognition module and directly enters the natural language processing module.

[0052] Furthermore, it also includes a voice acquisition and preprocessing module, which is responsible for acquiring user voice signals using the device's built-in microphone or external audio input device, converting analog signals into digital signals, and preprocessing the digital signals.

[0053] The user interaction interface module includes: Data Input: Receives voice or text information input by users through various terminal devices (such as web pages and mobile apps). If it is voice, it is directly transmitted to the voice acquisition and preprocessing module; if it is text, it skips the voice-related modules and directly enters the natural language processing module.

[0054] Data Output: The final generated voice or text response is presented to the user. For voice responses, the audio stream output from the speech synthesis module is received and played; for text responses, it is displayed directly on the user interface.

[0055] It is understood that the preprocessing includes noise reduction, endpoint detection, and normalization.

[0056] Furthermore, the user's preprocessed speech signal is converted into text, specifically including: Deep learning algorithms are used to convert the preprocessed speech signal into a phoneme sequence; By combining a language model, the phoneme sequence is decoded and converted into text form.

[0057] Furthermore, the speech model employs an N-Gram model or a neural network-based language model.

[0058] Furthermore, the semantic analysis model is constructed using a BERT pre-trained model.

[0059] Furthermore, such as Figure 2 As shown, the text identifies intent and text features, specifically including: In the lexical analysis stage, a model combining a bidirectional long short-term memory network and a conditional random field is used. BiLSTM can capture the contextual information of words in the text, while CRF uses global information in the text to optimize the word segmentation and part-of-speech tagging results to obtain the lexical analysis results. Using a graph neural network (GNN)-based approach, each word in the text is treated as a node. By constructing a syntactic relationship graph between words, the deep grammatical structure of sentences is mined, and the subject, verb, object, attributive, adverbial, and complement components of the sentences are analyzed to generate a syntactic tree. Based on the BERT pre-trained model, text features are extracted using a syntax tree. The extracted text features are then mapped to different intent categories through a fully connected layer, thus obtaining the intent category mapped to the text features.

[0060] It is understandable that the intent is the intent type with the highest number of matching text features.

[0061] Furthermore, such as Figure 3 As shown, relevant answers are filtered using text features, specifically including: By using text features as keywords and employing inverted indexing technology, knowledge entries containing relevant keywords can be quickly located in the knowledge base. The knowledge entries selected through keyword matching are represented as vectors from the text of the user's question and the candidate knowledge entries. Typically, a bag-of-words model or a word embedding model based on deep learning is used to convert the text into vector form. Cosine similarity is calculated by taking the cosine of the angle between two vectors as the similarity. Knowledge items with a similarity greater than a preset similarity threshold will be used as relevant answers.

[0062] Furthermore, the preset similarity threshold is determined based on the number of knowledge items, wherein the more knowledge items there are, the larger the preset similarity threshold is, and in one possible embodiment it is set to 0.9.

[0063] Furthermore, the knowledge entries are stored in a structured or unstructured manner.

[0064] Furthermore, the knowledge entries are stored in relational databases, non-relational databases, or knowledge graphs.

[0065] Furthermore, the method for determining the recommendation strategy for the relevant answers is as follows: Based on the distribution of relevant answers across different user questions, determine the user question type to which the relevant answers belong, and use that as the question type for answering; Based on the number of relevant answers to the question type and the usage of those answers, determine the types of interference in the question-answering process; The recommended processing strategy for the relevant answers is determined based on the type of interference.

[0066] It should be noted that the question type is determined based on the historical answers to user questions, specifically classifying different user questions into different question types.

[0067] It is understood that the interference type is determined based on the number of relevant answers and the ranking of the frequency of use of the relevant answers in the question type. Specifically, when the number of relevant answers meets the requirements, the interference type is determined to be a type II interference type. When the number of relevant answers does not meet the requirements, if the number of relevant answers with a higher ranking of the frequency of use of the relevant answers in the question type does not meet the requirements, the interference type is determined to be a type I interference type. In other cases, it belongs to type II interference type.

[0068] When the number of relevant answers is less than 4, the interference type in the question type is determined to be a type II interference. When the number of relevant answers is not less than 4, if the number of relevant answers that are higher than the number of times the relevant answer is used in the question type is greater than 3, the interference type is determined to be a type I interference. In other cases, it belongs to type II interference.

[0069] It should be noted that when the number of interference types of the relevant answers does not meet the requirements, and in a possible embodiment there are more than 5, the relevant answers will be recommended in different question types. When there are multiple relevant answers that need to be recommended in a question type, the relevant answer with the lowest ranking result in the question type will be recommended.

[0070] When the number of interference types in a relevant answer is no more than 5, if there is an interference type, the relevant answer recommendation process is carried out in the question type of the interference type.

[0071] Optionally, the method for determining the recommendation strategy for the relevant answers is as follows: Based on the distribution of relevant answers across different user questions, determine the user question type to which the relevant answers belong, and use that as the question type for answering; Based on the number of relevant answers to the question type and the usage of those relevant answers, a recommended processing strategy for the relevant answers is determined.

[0072] It is understandable that if the number of related answers with a higher ranking than the number of times the related answer is used in the question type does not meet the requirements in different question types, then recommendation processing will be performed in all question types. In other cases, the related answers to be recommended will be determined based on the similarity calculation results and popularity ranking.

[0073] Furthermore, the method for determining the deletion process for the relevant answers is as follows: Based on the recommendation processing strategy for the relevant answers, determine the number of question types for which the relevant answers are recommended for processing; Based on the interference patterns in different user questions, determine the type of interference in different question types. Based on the number of related answers used as recommended question types and the types of interference in different question types, a method for deleting related answers is determined.

[0074] It is understandable that when the number of related answers as recommended question types and the number of interference types in different question types both exceed a threshold, the related answer will be deleted from the knowledge base management module if, after being recommended in different question types, the number of times an answer does not meet the user's needs is high. Similarly, when the number of related answers as recommended question types and the number of interference types in different question types both exceed a threshold, the related answer will be deleted from the knowledge base management module only if, after being recommended in different question types, the number of question types where the answer does not meet the user's needs is high.

[0075] In one possible embodiment, when the number of related answers as recommended question types is greater than 8 or the number of one type of interference question types is greater than 6, if the number of times the related answers are not the user's needs after being recommended in different question types is more than 20, then the related answers will be deleted in the knowledge base management module. In one possible embodiment, if the number of answer types for which the relevant answers do not meet the user's needs exceeds 75% of the number of answer types for which the relevant answers are recommended in different question types, then the relevant answers will be deleted from the knowledge base management module.

[0076] Furthermore, the method for filtering answers is as follows: When the user's question belongs to the type of question for which the relevant answer is recommended, the recommendation processing strategy is directly used to determine the selected answer. When the user's question does not belong to the question type for which the related answers are recommended, the semantic similarity and popularity of the related answers are used to determine the selected answer.

[0077] It should be noted that when the matching of keywords between the user question and the answer question type meets the requirements, the user question is determined to belong to the answer question type. In one possible embodiment, when there are more than 4 matching keywords, the user question is determined to belong to the answer question type.

[0078] It is understandable that, such as Figure 4 As shown, the selection of answers is based on the semantic similarity and popularity of the related answers, specifically including: First, semantic similarity is used to initially rank related answers. Then, for related answers with similar semantic similarity, they are ranked a second time based on popularity. The answer with the highest ranking is selected as the final answer.

[0079] It should be noted that the popularity is determined based on the number of times the relevant answer is used.

[0080] Understandably, when the absolute value of the deviation between the semantic similarity of related answers is less than 1%, they are determined to be related answers with similar semantic similarity.

[0081] Specifically, speech synthesis technology is used to generate corresponding audio signals, including: Perform prosodic analysis on the answer text to determine the prosodic information containing stress, intonation, and pauses for each word; Using speech synthesis technology, corresponding audio signals are generated based on the text content and prosodic information of the answer.

[0082] In one possible embodiment, speech synthesis technology is used to generate the corresponding audio signal, specifically including: Using speech synthesis technology, corresponding audio signals are generated based on the text content of the answer.

[0083] Furthermore, it also includes a dialogue management module, which is responsible for tracking dialogue status, handling multi-round dialogues, and making decisions on transferring to human customer service.

[0084] Dialogue Status Tracking: Maintains the dialogue status between users and the customer service system, recording previous user questions, system responses, and the current dialogue progress. Based on the dialogue status and new user input, it determines whether further follow-up questions or clarifications are needed, or whether a direct response should be generated.

[0085] Multi-turn dialogue processing: For user questions that require multiple rounds of interaction to resolve, the module coordinates the information transfer between modules. For example, when a user's question requires additional information, the dialogue management module generates a guiding question, which is then fed back to the user through the user interaction interface module. The module then receives new input from the user and passes it again to the natural language processing module for processing.

[0086] Human Customer Service Transfer Decision: When the system determines that it cannot accurately answer a user's question (e.g., no matching answer in the knowledge base, unclear intent recognition, etc.), the dialogue management module decides whether to transfer the user to a human customer service representative. If transferred, the user's relevant information (including dialogue history, question description, etc.) is passed to the human customer service transfer module, and the subsequent interaction process between human customer service and the system is coordinated.

[0087] Specifically, the dialogue state tracking maintains the dialogue state through a finite state machine, recording the user's previous questions, the system's responses, and the current dialogue progress.

[0088] Understandably, the multi-turn dialogue processing determines that when a user's question requires additional information, the dialogue management module generates a guiding question, which is fed back to the user through the user interaction interface module. Then, it receives new input information from the user and passes it to the natural language processing module for processing.

[0089] Specifically, the human customer service transfer decision: When the system determines that it cannot accurately answer the user's question, the dialogue management module decides to transfer the user to human customer service, passing the user's question and response data to the human customer service transfer module.

[0090] Specifically, if there is no matching answer in the knowledge base management module or the number of times a question has been answered exceeds a preset threshold, the system will determine that it cannot accurately answer the user's question.

[0091] Module implementation details Voice Acquisition and Preprocessing Module: In practical applications, a high-sensitivity microphone is selected as the voice acquisition device to ensure clear acquisition of user voice signals. An adaptive filtering algorithm based on wavelet transform continuously adjusts filter parameters to adapt to noise characteristics under different environments, effectively reducing noise. A dual-threshold energy detection method sets two thresholds (high and low) based on the energy changes of the voice signal to accurately determine the start and end positions of the speech and remove silent segments. Normalization processing unifies the amplitude of the voice signal to the range [-1, 1], facilitating subsequent processing.

[0092] Speech recognition module: Collects a large amount of speech data containing different accents, speaking speeds, and intonations to train an end-to-end speech recognition model based on the Transformer architecture. During training, the model parameters are continuously adjusted to optimize the performance of the acoustic and language models, enabling them to accurately convert speech into text. For example, it achieves high recognition accuracy for common languages ​​such as Mandarin, Cantonese, and English.

[0093] The Natural Language Processing (NLP) module pre-trains a BERT-based multi-task learning architecture using a large-scale text corpus, then fine-tunes it for specific business domains and user needs. In lexical analysis, word segmentation tools are used to accurately segment the text, and part-of-speech tagging is employed to label the part of speech of each word. Syntactic analysis involves constructing a parser to analyze sentence structure. The semantic understanding and intent recognition model accurately determines user intent based on text features and pre-trained knowledge, such as accurately identifying a user's "I want to check recent activities" as a query intent.

[0094] The knowledge base management module stores business knowledge, frequently asked questions, and other information in a relational database called MySQL, creating an efficient data index. During knowledge retrieval, it first quickly filters potentially relevant knowledge items through keyword matching, then uses cosine similarity calculation for further precise matching, improving retrieval efficiency and accuracy. For new business knowledge and user-reported questions, automated scripts regularly update the knowledge base to ensure its timeliness.

[0095] The dialogue management module constructs a finite state machine by defining a series of dialogue states and state transition rules. For example, the initial state is waiting for user input. Upon receiving a user's question, the system transitions to the appropriate state based on the question type and the identified intent, such as a query state or a business processing state. In multi-turn dialogues, the system continuously adjusts the dialogue state based on the user's responses and the system's judgment, guiding the dialogue smoothly. When the system determines that the problem cannot be resolved, it triggers a human customer service transfer decision.

[0096] Answer generation module: When filtering answers, the modules first sort the answers retrieved from the knowledge base according to their matching degree from highest to lowest, prioritizing the answer with the highest matching degree. If multiple answers with similar matching degrees exist, further filtering is performed based on the answer's popularity (i.e., historical usage frequency). The selected answers are then formatted, such as standardizing punctuation and adjusting paragraph formatting. Necessary information is supplemented based on the context of the user's question, such as specific parameters of the relevant product and detailed time and location of the event.

[0097] Human Customer Service Transfer Module: This module connects human customer service representatives to the system via instant messaging. When a transfer is triggered, the system sends the user's detailed information, including user ID, service type, conversation history, and current problem description, to the human customer service representative's interface. During the interaction with the user, the human customer service representative records the conversation content and processing results through the system. The system then updates its knowledge base and optimizes its models based on these records.

[0098] The speech synthesis module utilizes the WaveNet model to train the speech synthesis system, learning features from a large amount of human speech data, including timbre, pitch, and prosody. During speech synthesis, it generates natural and fluent speech based on text content and prosodic analysis results. For example, for texts with different emotional tones, such as inquiries and complaints, it can synthesize speech with corresponding emotional inclinations, improving the user experience.

[0099] System application scenarios In the e-commerce sector: During the shopping process, users can use the AI ​​voice customer service platform to inquire about product information, learn about promotional activities, and check logistics and delivery progress. The system can answer user questions quickly and accurately, improving shopping efficiency. When encountering complex issues, such as disputes over returns and exchanges, users can be promptly transferred to human customer service to ensure that the problem is properly resolved.

[0100] In the financial industry: Customers can use voice commands to inquire about account balances, transaction details, and wealth management product information. The system rigorously verifies user identity to ensure information security. For complex transactions involving risk assessment and investment advice, the system can provide professional answers or transfer users to dedicated customer service personnel based on an accurate understanding of their intent.

[0101] Telecommunications operators: Users can use voice commands to change service plans, check phone bills, and report faults. The system provides personalized service recommendations based on user history and behavior. When handling fault reports, it can quickly locate the problem and provide corresponding solutions or arrange for repair personnel to provide on-site service.

[0102] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0103] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0104] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. An AI voice customer service platform, characterized in that, Specifically, it includes: The speech recognition module is responsible for converting the user's pre-processed speech signal into text, while the natural language processing module is responsible for passing the intent and text features identified by the text to the knowledge base management module and the human customer service transfer module through the semantic analysis model. The knowledge base management module is responsible for combining a hybrid retrieval strategy of keyword matching and similarity calculation, using text features to filter relevant answers, determining the recommendation processing strategy for the relevant answers based on the distribution of relevant answers in different user questions, the number of relevant answers for user questions, and usage, and determining the deletion processing method for the relevant answers based on the recommendation processing strategy and the interference in different user questions. The answer generation module is responsible for filtering answers based on the similarity calculation results of recommended related answers, popularity ranking, and recommendation processing strategies. The speech synthesis module uses the filtered answers and uses speech synthesis technology to generate corresponding audio signals, and uses the audio signals for automatic response processing. The human customer service transfer module is responsible for transferring users to available human customer service seats.

2. The AI ​​voice customer service platform method as described in claim 1, characterized in that, It also includes a user interaction interface module, which is responsible for receiving voice or text information input by users through various terminal devices.

3. The AI ​​voice customer service platform method as described in claim 2, characterized in that, If the user interaction interface module receives voice information, it directly passes it to the voice recognition module; if the user interaction interface module receives text information, it skips the voice recognition module and directly enters the natural language processing module.

4. The AI ​​voice customer service platform method as described in claim 1, characterized in that, It also includes a voice acquisition and preprocessing module, which is responsible for acquiring user voice signals using the device's built-in microphone or external audio input device, converting analog signals into digital signals, and preprocessing the digital signals.

5. The AI ​​voice customer service platform method as described in claim 1, characterized in that, The preprocessing includes noise reduction, endpoint detection, and normalization.

6. The AI ​​voice customer service platform method as described in claim 1, characterized in that, Converting the user's pre-processed speech signal into text specifically includes: Deep learning algorithms are used to convert the preprocessed speech signal into a phoneme sequence; By combining a language model, the phoneme sequence is decoded and converted into text form.

7. The AI ​​voice customer service platform method as described in claim 1, characterized in that, The text identifies intent and text features, specifically including: In the lexical analysis stage, a model combining a bidirectional long short-term memory network and a conditional random field is used. BiLSTM can capture the contextual information of words in the text, while CRF uses global information in the text to optimize the word segmentation and part-of-speech tagging results to obtain the lexical analysis results. Using a graph neural network (GNN)-based approach, each word in the text is treated as a node. By constructing a syntactic relationship graph between words, the deep grammatical structure of sentences is mined, and the subject, verb, object, attributive, adverbial, and complement components of the sentences are analyzed to generate a syntactic tree. Based on the BERT pre-trained model, text features are extracted using a syntax tree. The extracted text features are then mapped to different intent categories through a fully connected layer, thus obtaining the intent category mapped to the text features.

8. The AI ​​voice customer service platform method as described in claim 1, characterized in that, Using text features to filter relevant answers specifically includes: By using text features as keywords and employing inverted indexing technology, knowledge entries containing relevant keywords can be quickly located in the knowledge base. The knowledge entries selected through keyword matching are represented as vectors from the text of the user's question and the candidate knowledge entries. Typically, a bag-of-words model or a word embedding model based on deep learning is used to convert the text into vector form. Cosine similarity is calculated by taking the cosine of the angle between two vectors as the similarity. Knowledge items with a similarity greater than a preset similarity threshold will be used as relevant answers.

9. The AI ​​voice customer service platform method as described in claim 1, characterized in that, The selection of answers is based on the semantic similarity and popularity of the related answers, specifically including: First, semantic similarity is used to initially rank related answers. Then, for related answers with similar semantic similarity, they are ranked a second time based on popularity. The answer with the highest ranking is selected as the final answer.

10. The AI ​​voice customer service platform method as described in claim 1, characterized in that, The corresponding audio signal is generated using speech synthesis technology, specifically including: Perform prosodic analysis on the answer text to determine the prosodic information containing stress, intonation, and pauses for each word; Using speech synthesis technology, corresponding audio signals are generated based on the text content and prosodic information of the answer.

Images