Rule engine and LLM combined voice quality inspection method and system
By combining a rule engine with a large language model, the problems of weak semantic understanding and high rule base maintenance costs in speech quality inspection are solved, achieving efficient and accurate speech quality inspection that can adapt to complex scenarios and rapid changes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-11
- Publication Date
- 2026-04-03
AI Technical Summary
Existing speech quality inspection rule engines have weak semantic understanding capabilities, low coverage and high false negative rate for complex and unstructured quality inspection items, high rule base maintenance costs, and machine learning models rely on large-scale labeled data and have limited generalization capabilities.
It adopts a hierarchical architecture that combines a rule engine and a large language model (LLM). Through unified scheduling of agents, the rule engine handles explicit tasks and the LLM handles complex tasks, achieving deep collaboration and improving semantic understanding capabilities and flexibility.
It achieves high-efficiency, high-precision, and low-cost voice quality inspection, can handle complex unstructured quality inspection items, quickly adapt to business changes, reduce the burden of rule maintenance, and improve system flexibility and accuracy.
Smart Images
Figure CN121789652A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of audio quality inspection technology, and in particular to a voice quality inspection method and system that combines a rule engine and an LLM. Background Technology
[0002] In the field of customer service quality inspection, with the continuous increase in interaction volume, the traditional model relying on manual sampling is no longer sufficient to meet actual needs in terms of efficiency, cost, and coverage. The maturity of Automatic Speech Recognition (ASR) technology provides a crucial foundation for efficiently converting massive amounts of call recordings into analyzable text data, driving the development of automated intelligent quality inspection technology. To address the challenges of automated quality inspection, the industry has mainly developed two technical approaches: methods based on predefined rule engines and methods based on data-driven models. While pursuing automation goals, both approaches face significant limitations.
[0003] For example, the existing patent CN111210842A describes a voice quality inspection method, device, terminal, and computer-readable storage medium. The method involves acquiring a voice file to be tested; matching a corresponding voice quality inspection engine to the voice file; processing the voice file using the corresponding voice quality inspection engine to generate initial quality inspection data; processing the voice file according to a preset quality inspection library to generate quality inspection benchmark data; and comparing the initial quality inspection data and the quality inspection benchmark data according to preset comparison rules to output the comparison results. This method utilizes a rule engine for voice quality inspection, but rule engines have weak semantic understanding capabilities, resulting in low coverage and high false negative rates for complex, unstructured quality inspection items. Summary of the Invention
[0004] The purpose of this invention is to address the problems of weak semantic understanding capabilities of existing speech quality inspection rule engines, resulting in low recognition coverage and high false negative rates for complex and unstructured quality inspection items. This invention provides a speech quality inspection method and system that combines a rule engine and an LLM (Local Management Model). It adopts a layered architecture design to combine the rule engine and the LLM, achieving high efficiency, high accuracy, and low cost in speech quality inspection.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A speech quality inspection method combining a rule engine and LLM includes the following steps: Extract the dialogue speech stream from the audio to be inspected and convert the dialogue speech stream into text information; Establish a rule engine and LLM for voice quality inspection to determine the type of voice quality inspection task based on text information; Design an agent-based hierarchical architecture to achieve intelligent and organic collaboration between the rule engine and LLM. For a single voice quality inspection task, the rule engine is used for voice quality inspection. For complex voice quality inspection tasks, LLM is used to identify and extract non-compliant dialogue from text information to obtain quality inspection data.
[0006] The speech quality inspection method combining rule engine and LLM provided by this invention is designed based on an agent architecture to achieve deep organic collaboration between rule engine and LLM. Through unified scheduling of agents, it maximizes complementary advantages, reduces the burden of rule maintenance, and improves overall efficiency and flexibility.
[0007] Preferably, the step of extracting the dialogue speech stream from the audio to be inspected includes: extracting the speaker's speech stream with timestamps from the audio to be inspected; reducing non-human background noise in the speaker's speech stream to obtain a denoised speech stream; using voiceprint recognition technology to separate the denoised speech stream based on the speaker, and extracting the dialogue speech stream containing only the target speaker from the recognition results.
[0008] Preferably, when using a rule engine for voice quality inspection, if the dialogue voice stream is for a single use case, the text information is subjected to voice quality inspection based on the first rule library within the rule engine; if the dialogue voice stream is for multiple use cases, the text information is segmented into words and semantic features are extracted, and the semantic features are subjected to voice quality inspection based on the second rule library within the rule engine.
[0009] Preferably, the LLM is based on a deep learning algorithm and trained on large-scale text data. The LLM is based on the Transformer architecture and includes Transformer units containing multi-head self-attention layers and fully connected layers. The multi-head self-attention layers and fully connected layers are connected by residual connections, and adjacent Transformer units are connected by residual connections. The LLM uses a self-attention mechanism to process the input text and capture the semantic relationships between different parts of the text.
[0010] Preferably, the single voice quality inspection task type includes basic keyword and process compliance inspection tasks, while the complex voice quality inspection task type includes dynamic and semantically intensive tasks.
[0011] Preferably, a quality inspection list is output based on the quality inspection results. The quality inspection list includes violations, violation times, and number of violations.
[0012] As a preferred approach, historical dialogue voice streams are analyzed using machine learning and data analysis methods to identify common user questions and needs. The knowledge base is then optimized based on the identification results, and strategies are adjusted based on real-time data.
[0013] A voice quality inspection system combining a rule engine and LLM, comprising: The system integration layer performs text conversion and intelligent quality inspection on the acquired voice information; The application function layer collects voice information and work order information, and scores customer service quality based on the voice quality inspection results of the system function layer.
[0014] Preferably, the system integration layer includes: a speech transcription engine that receives voice audio uploaded by the application function layer and requests to obtain speech-transcribed text; a data supply platform that pushes data to the interface; and an intelligent agent that performs quality inspection on the voice text sent by the application function layer and returns the parsing results.
[0015] Preferably, the application function layer includes a quality inspection interface conversion engine that uploads voice and audio to the system integration layer. The quality inspection interface conversion engine is connected to a quality inspection application unit. The quality inspection application unit sends voice and text to the intelligent agent interface conversion engine, and the intelligent agent interface conversion engine responds with the parsing results of the system integration layer to the quality inspection application unit.
[0016] Therefore, this invention has the following beneficial effects: By combining the rule engine with LLM, a deep and organic synergy between the rule engine and LLM is achieved: the rule engine focuses on handling explicit, efficient, and highly interpretable tasks, while the LLM takes the lead in handling complex, dynamic, and semantically dense tasks. Through this fusion and synergy mechanism, the rigor of rule logic and the intelligence of LLM semantic understanding are seamlessly combined, fully meeting the multi-dimensional automated quality inspection needs of service specifications, business accuracy, process compliance, and sensitive risks, and achieving high-efficiency, high-precision, and low-cost voice quality inspection. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the steps of the speech quality inspection method combining rule engine and LLM in this invention.
[0018] Figure 2 This is a schematic diagram of the architecture of the voice quality inspection system that combines a rule engine and an LLM in this invention.
[0019] In the diagram: 1. System Integration Layer; 2. Application Function Layer; 3. Speech Transcription Engine; 4. Data Supply Platform; 5. Intelligent Agent; 6. Quality Inspection Interface Conversion Engine; 7. Quality Inspection Application Unit; 8. Intelligent Agent Interface Conversion Engine. Detailed Implementation
[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1: This embodiment provides a speech quality inspection method that combines a rule engine and an LLM, such as... Figure 1As shown, the operation process is as follows: Step 1, extract the dialogue voice stream from the audio to be inspected and convert the dialogue voice stream into text information; Step 2, design a hierarchical architecture based on intelligent agents to achieve intelligent and organic collaboration between the rule engine and LLM. If it is a single voice quality inspection task type, the rule engine is used for voice quality inspection; Step 3, if it is a complex voice quality inspection task type, the LLM is used to identify and extract the illegal dialogue in the text information to obtain the quality inspection data.
[0021] This embodiment provides a speech quality inspection method combining a rule engine and an LLM (Local Level Management Model). It primarily addresses the problems of weak semantic understanding capabilities in speech quality inspection rule engines, low coverage and high false negative rates for complex, unstructured quality inspection items, high costs of data-driven model-based methods, and limited generalization capabilities. By designing an agent-based architecture, it achieves deep and organic collaboration between the rule engine and the LLM. Unified scheduling of the agents maximizes complementary advantages, reduces the burden of rule maintenance, and improves overall efficiency and flexibility.
[0022] The following examples and specific application scenarios further illustrate the technical solution and effects of the present invention. The following examples are explanations of the present invention, but the present invention is not limited to the following examples.
[0023] The speech quality inspection method combining a rule engine and LLM provided in this embodiment is mainly used for customer service quality inspection. Currently, customer service quality inspection primarily relies on predefined rule engines and machine learning models. The core of a predefined rule engine lies in making decisions based on a set of pre-defined, discrete logical rules, but its inherent limitation is its weak semantic understanding ability. The rules themselves lack the ability to handle contextual relationships, implicit intentions, and semantic ambiguity, making it difficult to effectively handle issues that require contextual inference, understanding of implicit intentions, or identification of semantic ambiguity (such as subjective evaluations of service attitude, indirect expressions of evasion, and contextual judgments of irrelevant answers). This results in low coverage and high false negative rates for such complex, unstructured quality inspection items. Furthermore, the dynamic changes and complexity of business rules also lead to high maintenance costs for the rule base, limiting system flexibility and scalability.
[0024] Machine learning typically utilizes pre-trained models (such as BERT) to train specialized predictive models for specific quality inspection tasks (such as service issue classification and emotion recognition). Its main advantage lies in its ability to learn semantic patterns and features from labeled data, thus possessing superior semantic understanding potential compared to pure rules when dealing with similar problems within the training data's coverage. However, its core drawbacks include: model performance heavily reliant on large-scale, high-quality labeled datasets, resulting in significant "cold start" obstacles and training optimization costs; limited generalization ability when facing new problem types or expressions outside the training data distribution; long iteration cycles for model updates to adapt to rapidly evolving business needs; and difficulties in practice in efficiently and transparently co-determining the model's probabilistic output with explicit business rule logic.
[0025] In summary, to address the shortcomings of existing technologies, this embodiment provides a voice quality inspection method combining a rule engine and LLM, meeting the requirements of high efficiency, high accuracy, and low cost for intelligent voice quality inspection of power supply station service telephones. Figure 1 As shown, the voice quality inspection process combining rule engines and LLM specifically includes the following steps: Step (1): Extract the dialogue speech stream from the audio to be inspected and convert the dialogue speech stream into text information.
[0026] Extracting dialogue speech streams from audio to be inspected includes: extracting the speaker speech stream with timestamps from the audio to be inspected; reducing non-human background noise in the speaker speech stream to obtain a denoised speech stream; since in the field of customer service quality inspection, the audio to be inspected generally only includes two speakers, the customer and the service personnel, voiceprint recognition technology can be used to separate the denoised speech stream based on the speaker to obtain denoised speech streams with timestamps classified according to different speakers; and extracting the dialogue speech stream with timestamps containing only the target speakers, including the customer and the service personnel, from the recognition results.
[0027] The timestamp includes the start and end times of the speech stream. Therefore, a denoised speech stream categorized by speaker and including timestamps is represented as: start time, speech stream information, and end time. Following this process, a denoised speech stream containing all target speakers is obtained. This denoised speech stream is then sorted by start time to obtain the dialogue speech stream.
[0028] The speech stream information in the dialogue speech stream is converted into text information using ASR technology. Then, according to the speaker to which the speech stream information belongs, the speaker is added to the text information, resulting in text information with timestamps and speaker information, represented as: speaker, start time, text information, end time.
[0029] The text information, including timestamps and speaker names, is sorted by start time to obtain the dialogue text information. The dialogue text is represented as: Speaker 1: Start Time, Text Message 1, End Time.
[0030] Speaker 2: Start time, text message 2, end time.
[0031] Speaker 1, start time, text information 3, end time.
[0032] Speaker 2: Start time, text information 4, end time.
[0033] Step (2): Establish a rule engine and LLM for voice quality inspection, and determine the voice quality inspection task type based on text information.
[0034] (2.1) Establish a rules engine.
[0035] The core of a rules engine lies in building a structured rules base, including, for example, a keyword matching list, business process node inspection logic, and call duration threshold settings. It uses pattern matching algorithms to scan and trigger judgments on the dialogue text obtained after speech conversion.
[0036] In this embodiment, two rules are constructed for voice quality inspection in different situations.
[0037] (2.2) Establish a large language model (LLM).
[0038] Large Language Model (LLM) is an artificial intelligence model based on deep learning algorithms, trained on large-scale text data. In this embodiment, the LLM is based on the Transformer architecture, processes the input text through a self-attention mechanism, captures the semantic relationships between different parts of the text, understands the meaning of the text, and generates reasonable responses.
[0039] Specifically, the Large Language Model (LLM) includes Transformer units containing multi-head self-attention layers and fully connected layers. The multi-head self-attention layers and fully connected layers use residual connections. There are multiple Transformer units in the LLM, and adjacent Transformer units use residual connections.
[0040] This embodiment also includes using prompt engineering to design and optimize prompt words that interact with the large language model (LLM) to guide the large language model to generate high-quality output that meets expectations.
[0041] The large language model has a built-in risk radar that can automatically detect hidden problems and provide immediate warnings of potential risks. It can combine verbal and nonverbal cues to identify customer dissatisfaction, anxiety, and other emotions.
[0042] Step (3): Design an agent-based hierarchical architecture to achieve intelligent and organic collaboration between the rule engine and LLM.
[0043] To address the issue that rule engines and traditional ML models are often isolated or superficially integrated, lacking a deep collaborative architecture, which prevents the effective fusion of the efficient interpretability of rules and the semantic understanding potential of models, this paper proposes a unified scheduling mechanism for agents to maximize the complementary advantages of rule engines and LLM, thereby improving the efficiency of speech quality inspection.
[0044] Specifically: If it is a single voice quality inspection task type, the rule engine is used for voice quality inspection. The single voice quality inspection task type is a clear, efficient task that requires strong interpretability, such as basic keyword and process compliance checks.
[0045] For complex speech quality inspection tasks, LLM is used to identify and extract non-compliant dialogue from the text information to obtain quality inspection data. Complex speech quality inspection tasks are characterized by their complexity, dynamic nature, and semantic density.
[0046] By deeply integrating the powerful semantic understanding, contextual reasoning, and few-shot learning capabilities of Large Language Models (LLM), the system fundamentally improves the accuracy and coverage of complex and fuzzy quality control items such as service attitude, emotion recognition, and irrelevant answers, overcoming the semantic bottleneck of rule engines. Secondly, leveraging the prompting engineering techniques of LLM and its extensive knowledge base, the system significantly reduces its strong dependence on massive amounts of task-specific labeled data, enhancing its ability to quickly adapt to new scenarios, new expressions, and changes in business rules, thus overcoming the generalization barriers of traditional ML models.
[0047] The following explains the specific process of speech quality inspection performed by the rule engine and the Large Language Model (LLM).
[0048] (3.1) Rule Engine.
[0049] In this embodiment, when using a rule engine for voice quality inspection, the first step is to extract keywords from the dialogue text using a preset knowledge base. The number of keywords in the dialogue text determines whether it belongs to a single use case or multiple use cases. If there is only one keyword in the dialogue text, it is considered a single use case, and voice quality inspection is performed on the text information according to the first rule base within the rule engine. If there are more than one keyword in the dialogue text, it is considered a multiple use case, and the text information is segmented to extract semantic features. Voice quality inspection is then performed on the semantic features using the second rule base within the rule engine.
[0050] The first rule base includes keywords and their corresponding violations, represented as [keyword, violation]. The second rule base contains rule semantic features. By calculating the similarity between the semantic features of the dialogue text and the corresponding rule semantic features in the second rule base, the similarity is compared with a preset threshold to obtain the quality inspection result.
[0051] (2) Large Language Model (LLM).
[0052] In this embodiment, the Large Language Model (LLM) performs speech quality inspection, which includes: text inspection, used to detect whether there are grammatical errors in the dialogue text; and sentiment analysis, used to extract the voices of service personnel and customers from the dialogue text and perform sentiment analysis, mainly analyzing the service attitude of service personnel and the dissatisfaction, anxiety and other emotions of customers.
[0053] In this embodiment, a quality inspection list is output based on the quality inspection results. The quality inspection list includes the violator, the violation, the violation time, the content of the violation dialogue text, and the number of violations.
[0054] Simultaneously, based on machine learning and data analysis methods, historical dialogue voice streams are analyzed to identify common user questions and needs. The knowledge base is optimized based on the identification results, and strategies are adjusted according to real-time data to quickly respond to the latest changes.
[0055] The speech quality inspection method combining rule engine and LLM provided in this embodiment has the following beneficial effects: (1) Integrating the rule engine with LLM to achieve deep and organic synergy between the two: The rule engine focuses on handling clear, efficient, and highly interpretable single voice quality inspection tasks, with high execution efficiency and highly interpretable output results; LLM takes the lead in handling complex, dynamic, and semantically dense voice quality inspection tasks, making up for the shortcomings of the rule engine. The rigor of rule logic is seamlessly combined with the intelligence of LLM semantic understanding to fully meet the multi-dimensional automated quality inspection needs of service specifications, business accuracy, process compliance, and sensitive risks, and achieve high-efficiency, high-precision, and low-cost voice quality inspection.
[0056] (2) Signal purification is performed on the mixed speech of multiple speakers and noisy speech in customer service calls, which reduces the error in the speech processing stage, reduces the misjudgment of the large language model due to inaccurate speech recognition, and improves the accuracy of speech quality inspection tasks. At the same time, combined with multiple technical processing in the text conversion and optimization stage, it provides high-quality text input for subsequent violation detection.
[0057] Example 2: This embodiment, building upon Embodiment 1, provides a speech quality inspection system combining a rule engine and LLM (Local Management Model). It implements the speech quality inspection method combining a rule engine and LLM provided in Embodiment 1, addressing significant challenges faced by traditional quality inspection methods: manual sampling is inefficient, costly, and has limited coverage. Systems based on pure rule engines lack flexibility and accuracy when handling complex semantic scenarios, and the rule base maintenance cost is high. Solutions relying on traditional machine learning models are limited by their strong dependence on large-scale, high-quality labeled data, insufficient model generalization ability, and difficulty in quickly adapting to changes in business rules. Through a fusion and collaboration mechanism, the rigor of rule logic is seamlessly combined with the intelligence of LLM semantic understanding, constructing an efficient, accurate, and scalable intelligent speech quality inspection system. This system comprehensively meets the multi-dimensional automated quality inspection needs of service specifications, business accuracy, process compliance, and sensitive risks, achieving the ultimate goal of high efficiency, high accuracy, and low cost.
[0058] Specifically, a voice quality inspection system that combines a rule engine and LLM, such as... Figure 2 As shown, it includes: a system integration layer 1 and an application function layer 2. The system integration layer is used for text conversion and intelligent quality inspection of the acquired voice information; the application function layer is used for collecting voice information and work order information, and for scoring customer service quality based on the voice quality inspection results of the system function layer. Adopting a layered architecture design, through the collaboration of the system integration layer and the application function layer, it achieves intelligent processing of voice, work order, and other data. It fundamentally solves problems such as low quality inspection efficiency, insufficient semantic understanding, and difficulty in closing the business response loop in power customer service and operation scenarios, covering the entire process of automation and intelligence from "voice and work order collection - text conversion - intelligent quality inspection - result feedback".
[0059] Specifically, the system integration layer includes a speech-to-text engine 3, a data supply platform 4, and an intelligent agent 5. The speech-to-text engine is used to receive the speech audio uploaded by the application function layer and request the speech-to-text transcription; the data supply platform is used to implement interface push; and the intelligent agent is used to perform quality inspection on the speech text sent by the application function layer and return the parsing results.
[0060] The application function layer includes a quality inspection interface conversion engine 6, a quality inspection application unit 7, and an intelligent agent interface conversion engine 8. The speech-to-text engine is connected to the quality inspection interface conversion engine, the quality inspection interface conversion engine is connected to the quality inspection application unit, the quality inspection application unit is connected to the intelligent agent interface conversion engine, and the intelligent agent interface conversion engine is connected to the intelligent agent. The quality inspection interface conversion engine is used to upload speech audio to the system integration layer. The quality inspection application unit is used to send speech text to the intelligent agent interface conversion engine. The intelligent agent interface conversion engine responds with the parsing results from the system integration layer to the quality inspection application unit.
[0061] Based on the above modules, the workflow of the voice quality inspection system combining the rule engine and LLM is as follows: The data supply platform pushes the interface to the quality inspection interface conversion engine, while the quality inspection interface conversion engine collects the voice audio to be inspected and uploads it to the speech-to-text engine. The speech-to-text engine uses ASR technology to convert the voice audio into text information with time information, generates dialogue text, and returns the dialogue text to the quality inspection interface conversion engine. The quality inspection interface conversion engine connects the data in text form to the intelligent quality inspection interface of the quality inspection application unit and sends the dialogue text to the quality inspection application unit. The quality inspection application unit uses it for customer request handling work order quality inspection and sends the dialogue text to the intelligent agent interface conversion engine. The intelligent agent interface conversion engine sends the received voice text to the intelligent agent for quality inspection. The intelligent agent contains the rule engine and the large language model LLM. According to the voice quality inspection task type of the text information, the rule engine and the large language model LLM are organically coordinated to select the corresponding solution for voice quality inspection and return the voice quality inspection result to the intelligent agent interface conversion engine. The intelligent agent interface conversion engine then responds with the voice quality inspection result to the quality inspection application unit.
[0062] In this embodiment, the intelligent agent can also rely on natural language processing (NLP) technology to analyze the dialogue content between customer service and customers, automatically detect the customer's emotional state through sentiment analysis algorithms, and identify and analyze the language features in the dialogue, such as the service personnel's speaking speed and whether the word choice is appropriate, to help enterprises understand customer experience in real time and standardize business processes.
[0063] The intelligent agent can also perform quality inspection on different types of materials such as audio, video, images, and documents. Specifically: when inspecting video, audio information is extracted from the video as the speech stream to be inspected, and then the speech quality inspection method of this application is used for speech quality inspection. When inspecting images and text, text information is extracted from the images and text, and then the speech quality inspection process is performed.
[0064] The voice quality inspection system combining a rule engine and LLM provided in this embodiment has the following beneficial effects: (1) By utilizing intelligent agents with autonomous perception, decision-making and action capabilities, a closed-loop voice quality inspection of "understanding-planning-execution" is achieved through the large voice model LLM, thereby improving the efficiency of voice quality inspection.
[0065] (2) It can realize the quality inspection of all channels and media such as calls, chats and videos. It supports multiple channels such as telephone calls and online chats, and can also detect different types of materials such as audio, video, pictures and documents.
[0066] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications are possible without departing from the technical solutions described in the claims.
Claims
1. A speech quality inspection method combining a rule engine and LLM, characterized in that, include: Extract the dialogue speech stream from the audio to be inspected and convert the dialogue speech stream into text information; Establish a rule engine and LLM for voice quality inspection to determine the type of voice quality inspection task based on text information; Design an agent-based hierarchical architecture to organically coordinate the rule engine and LLM. For a single voice quality inspection task, use the rule engine for voice quality inspection. For complex voice quality inspection tasks, LLM is used to identify and extract non-compliant dialogue from text information to obtain quality inspection data.
2. The speech quality inspection method combining a rule engine and LLM according to claim 1, characterized in that, The process of extracting the dialogue speech stream from the audio to be inspected includes: extracting a speech stream with timestamps from the audio to be inspected; reducing non-human background noise in the speech stream to obtain a denoised speech stream; using voiceprint recognition technology to separate the denoised speech stream based on the speaker, and extracting a dialogue speech stream with timestamps containing only the target speaker from the recognition results.
3. A speech quality inspection method combining a rule engine and LLM according to claim 1 or 2, characterized in that, When using a rule engine for voice quality inspection, if the dialogue voice stream is for a single use case, the text information is subjected to voice quality inspection based on the first rule library within the rule engine; if the dialogue voice stream is for multiple use cases, the text information is segmented and semantic features are extracted, and the semantic features are subjected to voice quality inspection based on the second rule library within the rule engine.
4. The speech quality inspection method combining a rule engine and LLM according to claim 1, characterized in that, The LLM is based on a deep learning algorithm and is trained on large-scale text data. The LLM is based on the Transformer architecture and includes Transformer units containing multi-head self-attention layers and fully connected layers. The multi-head self-attention layers and fully connected layers use residual connections, and adjacent Transformer units use residual connections. The LLM uses a self-attention mechanism to process the input text and capture the semantic relationships between different parts of the text.
5. A speech quality inspection method combining a rule engine and LLM according to claim 1, 2, or 4, characterized in that, The single voice quality inspection task type includes basic keyword and process compliance inspection tasks, while the complex voice quality inspection task type includes dynamic and semantically dense tasks.
6. The speech quality inspection method combining a rule engine and LLM according to claim 1, characterized in that, Output a quality inspection list based on the quality inspection results. The quality inspection list includes violations, violation times, and number of violations.
7. A speech quality inspection method combining a rule engine and LLM according to claim 1, 2, 4, or 6, characterized in that, Based on machine learning and data analysis methods, historical dialogue voice streams are analyzed to identify common user questions and needs. The knowledge base is optimized based on the identification results, and strategies are adjusted based on real-time data.
8. A speech quality inspection system combining a rule engine and an LLM, employing the speech quality inspection method combining a rule engine and an LLM as described in any one of claims 1-7, characterized in that, include: The system integration layer performs text conversion and intelligent quality inspection on the acquired voice information; The application function layer collects voice information and work order information, and scores customer service quality based on the voice quality inspection results of the system function layer.
9. A voice quality inspection system combining a rule engine and LLM according to claim 8, characterized in that, The system integration layer includes: a speech-to-text engine that receives voice audio uploaded by the application function layer and requests the speech-to-text transcription; a data supply platform that pushes data to the interface; and an intelligent agent that performs quality inspection on the voice text sent by the application function layer and returns the parsing results.
10. A voice quality inspection system combining a rule engine and LLM according to claim 8 or 9, characterized in that, The application function layer includes a quality inspection interface conversion engine that uploads voice and audio to the system integration layer. The quality inspection interface conversion engine is connected to a quality inspection application unit. The quality inspection application unit sends voice and text to the intelligent agent interface conversion engine, and the intelligent agent interface conversion engine responds with the parsing results from the system integration layer to the quality inspection application unit.