Customer service voice intelligent quality inspection method

Abstract

The application provides a customer service voice intelligent quality inspection method, comprising: a data storage step: receiving a customer service call recording file and storing it to an original recording library through an input-output layer interface; a voice transcription and semantic enhancement step: calling a customized ASR model to decode audio, generating a preliminary transcription text with a timestamp and performing enhancement processing; a hierarchical causal reasoning step: searching a preset rule knowledge base, matching a final transcription text with a scoring rule, generating a 5-7 layer causal conduction path constraint large language model to deduce a structured quality inspection report; a quality inspection report output step: transmitting a JSON format report to an API interface; a model optimization continuous learning step: fusing quality inspection report with artificial review feedback labeling data, performing field fine-tuning training and DPO reinforcement learning, and updating reasoning model weights to optimize subsequent quality inspection results. Thus, the automatic, efficient quality inspection of full customer service calls and the model dynamic optimization closed loop are realized, and the quality inspection process fragmentation problem is solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and in particular to an intelligent quality inspection method for customer service voice based on hierarchical causal reasoning and continuous learning. Background Technology

[0002] The study of customer service business processes plays a crucial role in enterprise operations and user experience, directly impacting service quality, compliance, and the establishment of customer trust. Through standardized process design, enterprises can ensure service consistency and legality while effectively reducing operational risks. However, the importance of this area is accompanied by complex challenges that urgently require in-depth exploration and solutions.

[0003] Currently, many solutions for analyzing customer service process deviations often remain at the level of surface-level data statistics, lacking in-depth analysis and dynamic tracking of the causes of deviations. This makes it difficult for enterprises to truly understand the root causes of problems in process execution and to take timely and targeted measures, thereby affecting the accuracy of service quality and compliance assessments. Especially in the ever-changing customer service scenarios, the causes of deviations are often hidden and involve multiple coupled factors, making accurate identification extremely difficult.

[0004] Against this backdrop, the core challenges facing the research have become increasingly apparent. The primary issue lies in accurately identifying the differences between standard service procedures and actual execution processes, especially when customer service personnel skip necessary steps. These differences may manifest as missing identity verification or omitted product confirmation steps. Failure to effectively capture these differences makes it difficult to trace the starting point and transmission path of process deviations, leading to unclear attribution of responsibility and a lack of targeted corrective measures.

[0005] Therefore, how to build a mechanism that can accurately identify process differences, trace the starting point of deviation and its transmission path, and improve the accuracy of business compliance scoring through this mechanism has become a key issue that urgently needs to be addressed. Summary of the Invention

[0006] This invention provides an intelligent quality inspection method for customer service voice messages to solve the above-mentioned technical problems and effectively improve the automation level and accuracy of customer service quality management.

[0007] The intelligent quality inspection method for customer service voice recordings of the present invention includes:

[0008] Data storage steps: Receive customer service call recording files and store them in the original recording library through the input / output layer interface;

[0009] Speech transcription and semantic enhancement steps: Decode the audio using a customized ASR model to generate preliminary transcribed text with timestamps, use a large language model to correct typos and optimize speaker separation, and add emotion tags to generate the final transcribed text;

[0010] Hierarchical causal reasoning steps: Retrieve the preset rule knowledge base, match the final transcribed text with the scoring rules, generate a large language model with 5 to 7 layers of causal transmission path constraints, and derive a structured quality inspection report.

[0011] Quality inspection report output steps: Transmit the JSON format report to the API interface;

[0012] The model optimization and continuous learning process involves integrating quality inspection reports with manually reviewed and annotated data, performing domain-specific fine-tuning training and DPO reinforcement learning, and updating the inference model weights to optimize subsequent quality inspection results. This achieves an automated, efficient quality inspection and dynamic model optimization loop for all customer service calls, resolving the fragmented quality inspection process.

[0013] Furthermore, the hierarchical causal reasoning steps include:

[0014] The sub-steps for triggering scoring rules are analyzed: When both the feature steps of the classification text feature keyword detection process and the emotion feature probability threshold verification process are missing, the corresponding causal inference is triggered.

[0015] The dynamic adjustment of deduction weights involves: associating the importance parameters of business scenarios, calculating and updating the deduction values ​​using weighted averages, and sorting the causal transmission paths according to logical hierarchy. This solves the problem of rigid rules, enabling dynamic weight adaptation and priority ranking (weight calculation based on business scenarios), representing an original quantitative design.

[0016] Furthermore, constructing 5 to 7 layers of causal transmission pathways includes:

[0017] Locating the event layer: Extracting semantic summaries of timestamps;

[0018] Matching rule layer: Matches knowledge base entries;

[0019] Description and Explanation Layer: The logic that associates events with rules;

[0020] Calculate deduction levels: Calculate deductions for individual items;

[0021] Analysis of the influence of dimensions: application of classification weights;

[0022] Aggregate total score calculation layer: calculates global deductions;

[0023] Rating and assessment layer: outputs service level. This structurally deciphers the traditional "black box" of quality inspection; the 7-layer hierarchical transmission (from event to rating) enhances traceability, superior to the industry's common 1-3 layer solution.

[0024] Furthermore, the model's continuous optimization learning steps include:

[0025] Verification feedback data volume threshold sub-step: When the number of labeled data reaches 500, trigger training with a domain fine-tuning learning rate of 5e-5, and train for 3 rounds;

[0026] DPO reinforcement learning sub-step: When 1000 data points are reached, DPO reinforcement learning is triggered to generate high-quality and low-quality reports as comparison data pairs to calculate and update the parameters of the preference function. This automatically triggers model updates; the data threshold (500 / 1000 data points) and training parameters (learning rate 5e-5) quantitatively address the high cost of manual iteration.

[0027] Furthermore, the manually reviewed and annotated data includes:

[0028] The standardization of objection labels involves creating a set of objection labels for customer interruptions, sentiment misjudgments, rule matching errors, and event omissions. The addition of a confidence parameter involves attaching a confidence parameter to the objection label set. This parameter's value includes a manually assigned "1" and a probability value output by the model. After being stored in the feedback database, it triggers an incremental training scheduling task. Thus, standardizing objection labels improves error correction accuracy; the confidence parameter (human = 1, model = probability) enables hierarchical management of feedback data, solving the problem of accumulated errors in traditional quality inspection.

[0029] Furthermore, DPO reinforcement learning includes:

[0030] The sub-step for constructing the reward signal involves weighted calculation of the causal transmission path integrity weight, the deduction basis consistency weight, the semantic accuracy weight, and the rule timeliness weight.

[0031] The normalization scoring sub-step involves normalizing the reward scores to the [-1,1] interval, optimizing the preference function parameters, and aggregating the output to update the model weights. This allows for multi-dimensional quantitative evaluation of the model output; weight allocation and score normalization improve the stability of reinforcement learning and address the preference sampling bias inherent in traditional methods.

[0032] Furthermore, a three-level verification process is performed before the structured quality inspection report is output, including:

[0033] Level 1: Detection of rule levels for causal breakage and missing rules;

[0034] Level 2: Verify the contradiction of total deduction score overflow;

[0035] Level 3: Timeliness of query rule knowledge base version;

[0036] If any level of verification fails, the model is triggered to re-infer and an error log is recorded. Thus, the verification mechanism improves the reliability of the report; the three-level verification logic (breakage, overflow, and timeliness) is an original design that solves the problem of unreliable conclusions.

[0037] Furthermore, determining the service level of the rating layer includes: analyzing the impact parameters of individual deductions on service rating downgrades; outputting downgrade thresholds, and mapping the total score range to the final service level of the quality inspection. This further limits the execution of the rating layer, solving the problems of "vague rating standards and uninterpretable results" in traditional quality inspection, and providing precise quantitative basis for customer service quality optimization.

[0038] Furthermore, in the hierarchical causal reasoning step, the lightweight model combination is replaced, and a language model with a large number of 7B parameters is called to generate causal transmission paths. The rule engine is integrated to perform deduction calculations locally, and resource consumption is controlled through single-node deployment to support a daily processing time of ≤5 seconds for 10,000 quality inspections. Therefore, a lightweight, low-cost deployment scheme is designed; the 7B model parameters and the ≤5-second latency quantitative indicators address enterprise resource constraints.

[0039] Furthermore, the speech transcription and semantic enhancement steps include:

[0040] A customized ASR model is invoked for acoustic decoding, outputting timestamped text. This is then cascaded with an LLM semantic enhancement module to correct typos in industry terminology, dynamically adapt the edit distance threshold, separate the customer service voice channel mapping to the left channel identifier, and label and normalize the NRC sentiment probability values. The enhanced text is then transmitted to the hierarchical causal inference step, with the semantic enhancement results fed back to the ASR training set, triggering an incremental fine-tuning closed-loop mechanism. This cross-module closed-loop improves data quality; the cascaded acoustic decoding and semantic enhancement (with dynamically set edit distance thresholds) combined with the feedback loop solves the problem of fragmented transcription.

[0041] The present invention achieves beneficial effects through the following technical solutions:

[0042] To address the issue of unexplainable quality inspection results: By constructing a 5-7 layered causal reasoning module, a structured report containing timestamp event types and causal transmission path fields is forced to be output, making the scoring results verifiable and improving the trust of managers.

[0043] Achieving a closed-loop model optimization: By combining knowledge fine-tuning, manual review and feedback annotation, and DPO reinforcement learning, a self-evolving system of data collection, model optimization, and business verification is formed, reducing manual annotation costs and shortening the iteration cycle.

[0044] Improving speech transcription quality: Customized ASR models are infused with industry terminology and dialect data, and combined with LLM semantic enhancement to achieve typo correction, speaker separation and emotion annotation, significantly improving transcription accuracy in complex scenarios.

[0045] Supports full-scale automated quality inspection: JSON-formatted structured reports support efficient and objective full-scale quality inspection, replacing manual sampling with low coverage, and forming a self-evolving ecosystem through module collaborative optimization, ultimately improving quality inspection efficiency and result accuracy. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the intelligent quality inspection method for customer service voice commands according to the present invention.

[0047] Figure 2 The flowchart illustrates the continuous learning closed-loop process for model optimization in this invention. Detailed Implementation

[0048] While conducting in-depth research in the field of customer service quality inspection, the inventors discovered that although existing technical solutions have improved automation levels by introducing artificial intelligence and speech recognition, they still suffer from four core problems: the causal relationship of quality inspection scores is unexplainable, making it impossible for managers to trace the root cause of deductions; the quality of speech transcription is significantly affected by industry terminology and dialects, reducing the accuracy of analysis; the model optimization mechanism lacks a closed loop and is unable to respond to changes in business rules; and the deviation of processes from their starting points and transmission paths cannot be accurately traced. These problems collectively restrict the efficiency, accuracy, and interpretability of quality inspection. The limitations of existing solutions, such as insufficient interpretability, poor transcription accuracy, and weak model adaptability, make it difficult to systematically solve these problems through simple rule matching or general model optimization. Especially in high compliance scenarios, it is necessary to build a mechanism that can integrate deep causal analysis, customized speech processing, and dynamic model optimization. This requires the inventors to break through the traditional static analysis mode of quality inspection in terms of technical implementation, and upgrade the quality control system from surface data statistics to full-process traceability. The difficulty lies in how to accurately match technical feasibility with business needs in complex and ever-changing customer service scenarios.

[0049] Example 1

[0050] Please refer to Figures 1 to 2 This embodiment provides a method for intelligent quality inspection of customer service voice messages, and the specific steps are as follows:

[0051] 1. Data storage steps

[0052] Customer service call recordings are transferred to the original recording library via an input / output layer interface (such as a RESTful API or message queue). Recording file formats include PCM, WAV, or MP3, with a sampling rate of 16kHz and a bit rate of 16bit. The original recording library uses a distributed storage architecture (such as HDFS or MinIO), supports TB-level data expansion, and is partitioned by date with a retention period of 180 days.

[0053] 2. Speech-to-text transcription and semantic enhancement steps

[0054] Call the customized ASR model (an end-to-end model based on the Conformer architecture) to decode the recorded audio file, generating a preliminary transcript with timestamps, where the timestamp accuracy is at the millisecond level. The preliminary transcript is corrected for typos by a large language model (such as GPT-4 or LLaMA-2), for example, correcting "payment?" to "payment code"; speaker separation is optimized, and the customer service and customer voices are distinguished through voiceprint feature clustering, and NRC emotion tags (such as "angry", "satisfied") are attached to generate the final transcript. The probability values of the emotion tags are normalized to the range [0, 1] through the Sigmoid function.

[0055] Among them, in the step of attaching emotion tags, the "NRC emotion probability value" used refers to the emotion probability distribution calculated based on the NRC Emotion Lexicon released by the National Research Council Canada (NRC for short) and related emotion analysis models. The NRC Emotion Lexicon is a commonly used dictionary resource in the field of emotion analysis. Its core is to associate words with multiple basic emotion categories, defining 8 basic emotions such as joy, sadness, anger, fear, trust, disgust, surprise, and anticipation, as well as 2 polarities of positive and negative. The related emotion analysis model analyzes the text input (i.e., the transcribed dialogue text) based on this dictionary, and outputs the probability estimates of the text on each of the above basic emotions, forming an emotion probability distribution.

[0056] The "normalized output" refers to performing softmax normalization on the various emotion probability values output by the above model, making the sum of the probabilities of all emotions (i.e., the 8 basic emotions defined in the NRC Emotion Lexicon) equal to 1, and selecting the emotion with the highest probability value as the main emotion tag for this segment of the dialogue to attach. This standardization process ensures the objectivity and consistency of emotion judgment, providing quantitative emotion feature inputs for subsequent causal reasoning.

[0057] 3. Hierarchical causal reasoning step

[0058] Retrieve the preset rule knowledge base (such as in JSON or Protobuf format), and match the final transcript with scoring rules (such as "deduct 5 points if the customer interrupts more than 3 times"). The causal conduction path is a hierarchical structure with 5 to 7 layers, including:

[0059] Location event layer: Extract the semantic summary of the timestamp, such as "00:15:23 - 00:15:25 Customer interrupts";

[0060] Match rule layer: Retrieve the knowledge base entries, such as "interruption count rule";

[0061] Description and Explanation Layer: The logic that associates events with rules, such as "interruptions affect service experience";

[0062] Calculate deduction levels: Calculate individual deductions according to the rules;

[0063] Analysis of the impact layer: application business scenario weight (e.g., complaint scenario weight is 1.5);

[0064] Aggregate total score calculation layer: weighted calculation of global deductions;

[0065] Determine the rating level: output the service level (e.g., A level, B level).

[0066] The five layers include: the event location layer, the matching rule layer, the description and explanation layer, the deduction calculation layer, and the judgment and rating layer; the six layers are the analysis dimension influence layer added between the calculation deduction layer and the judgment and rating layer; and the seven layers are the aggregated total score calculation layer added between the analysis dimension influence layer and the judgment and rating layer.

[0067] Specifically, the basis and technical characteristics for setting the range of 5 to 7 layers in constructing causal transmission paths are as follows:

[0068] 1) Setting basis:

[0069] The necessity of a minimum of 5 layers: fewer than 5 layers (such as the common 1-3 layer scheme) will result in the absence of core transmission links (such as logical association or dimensional influence layers), leading to a break in the causal chain and the formation of a reasoning black box.

[0070] The upper limit of 7 layers is critical: Experiments have shown that introducing more than 7 layers introduces redundant conversion layers, which introduces unnecessary noise, significantly increases computational complexity, and has limited accuracy gains.

[0071] 2) Technical rationality:

[0072] Minimal sufficiency: 5-7 layers cover the entire chain of "fact layer → rule matching layer → logical association layer → quantitative deduction layer → dimension influence layer → total score aggregation layer → level determination layer", with each layer having independent and irreplaceable functions.

[0073] Performance balance: Achieving optimal synergy between traceability (results can be traced back to specific events), accuracy (compared to a 3-layer solution), and efficiency.

[0074] 3) Universality:

[0075] The core framework is general: "Event → Rule Mapping → Correlation Analysis → Quantification → Multidimensional Integration → Scoring → Rating" logic is adaptable to multiple fields (financial compliance, e-commerce after-sales service, etc.).

[0076] Flexible hierarchical configuration: rules for each layer can be customized, but layers 5-7 remain fixed as the "minimum sufficient set of layers" and are not affected by business differences.

[0077] Furthermore, this scope solves the problems of "black box" and "inefficiency" of traditional solutions through quantitative verification, and achieves synergistic optimization of traceability, accuracy and efficiency.

[0078] 4. Quality Inspection Report Output Steps

[0079] The structured quality inspection report is transmitted to the target system (such as a CRM or BI platform) in JSON format via an API interface. The report fields include call ID, total score, deduction details, event timestamp, etc.

[0080] 5. Model Optimization and Continuous Learning Steps

[0081] The structured quality inspection report is integrated with the manually reviewed feedback and labeled data. Domain fine-tuning training (learning rate 5e-5, batch size 32) and DPO (Direct Preference Optimization) reinforcement learning are performed to update the inference model weights to optimize subsequent quality inspection results.

[0082] Preferably, the training objective function (fine-tuning loss function) for domain fine-tuning training is:

[0083] ,

[0084] Where xi is a dialogue segment with rule prompts, yi is the corresponding deduction event label, θft is the fine-tuning parameter, N is the number of training samples, the learning rate is set to 5e-5, the training is performed for 3 epochs, and P(|) is the conditional probability symbol.

[0085] When using the DPO algorithm for reinforcement learning, the corresponding objective function (direct preference optimization loss function) is:

[0086] ,

[0087] Where E is the expectation operator, x is the input sample, y+ represents a high-quality report, and y- represents a low-quality report. Let σ be the sample sampling identifier, and let σ be the sigmoid function. For high-quality report scoring functions, Let θ be the scoring function for substandard reports, and θ be the model parameters. The significance of this objective function is to sample "pairs of good and bad reports from the same call" from the manually reviewed quality inspection data. By optimizing the model parameters, the model's score for good reports should be significantly higher than that for substandard reports. Ultimately, the model's output quality inspection results should align with human standards, addressing the problems of "lack of closed-loop model optimization and weak ability to adapt to human preferences" in existing technologies. Preferably, reinforcement learning is triggered every 1000 D_feedback entries collected, with each training session taking approximately 30 minutes.

[0088] Example 2

[0089] Based on Example 1, the hierarchical causal steps in the causal reasoning process further include:

[0090] The sub-steps for analyzing the scoring rule triggering conditions are as follows: the rule is triggered when both the classification text feature keyword detection process (such as using the TF-IDF algorithm to extract keywords such as "complaint" and "rejection") and the emotion feature probability threshold verification process (such as when the keyword probability is ≥0.8) are missing.

[0091] The sub-step for dynamically adjusting the deduction weight is to associate the importance parameters of the business scenarios (e.g., the weight of the complaint scenario = 1.2, and the weight of the ordinary consultation scenario = 0.8) and update the deduction value by weighting.

[0092] Preferably, the weighted calculation formula is: Deduction value = Basic deduction value × Scene weight × Rule level coefficient, where the rule level coefficient lᵣ = 1-7 corresponding to 0.1-0.7.

[0093] Example 3

[0094] Based on Example 1, when the hierarchical structure of the causal transmission path is specifically 6 layers, it includes:

[0095] Location of event layer: Extracting semantic summaries of timestamps using the BERT model;

[0096] Matching rules layer: Uses Elasticsearch to retrieve knowledge base entries;

[0097] Description and Explanation Layer: Generates explanatory text using rule templates;

[0098] Calculate deduction points: Calculate deduction points according to the rules (e.g., deduct 2 points for each interruption);

[0099] Analysis of the impact layer: application business weight (e.g., VIP customer weight = 1.5);

[0100] Determine the rating level: output the service level (e.g., a total score of ≥90 is grade A).

[0101] Example 4

[0102] Based on Example 1, the model optimization continuous learning steps include:

[0103] Verification of feedback data volume threshold sub-step: When the amount of labeled data reaches 500, trigger training with a domain fine-tuning learning rate of 5e-5, and train for 3 rounds;

[0104] DPO reinforcement learning sub-step: When the amount of labeled data reaches 1000, generate high-quality reports (human score ≥ 90) and low-quality reports (human score < 70) as comparison data pairs, and update the preference function parameters.

[0105] Example 5

[0106] Based on Example 1, the manually reviewed and annotated data includes:

[0107] Standardizing objection tags: Create a set of objection tags (e.g., "customer interruption", "emotional misjudgment", "rule mismatch", "event omission").

[0108] The sub-step of attaching confidence parameters: Attach confidence parameters (manually labeled as "1" or model output probability values) to the objection labels, store them in the feedback database, and then trigger the incremental training scheduling task.

[0109] Example 6

[0110] Based on Example 1, DPO reinforcement learning includes:

[0111] The sub-step for constructing the reward signal is as follows: weighted calculation of the causal path integrity weight (e.g., 0.3 or 0.35), the deduction basis consistency weight (e.g., 0.3), the semantic accuracy weight (e.g., 0.2), and the rule timeliness weight (e.g., 0.2).

[0112] Normalization scoring sub-step: Normalize the reward score to the [-1,1] interval, optimize the preference function parameters, and aggregate the output to update the model weights.

[0113] Example 7

[0114] Based on Example 1, a three-level verification is performed before outputting the structured quality inspection report:

[0115] Detect whether there are broken or missing rule levels in the causal transmission path;

[0116] Verify whether the total deduction score overflows (e.g., total score > 100) or contains a logical contradiction;

[0117] Check the version timeliness of the rule knowledge base (e.g., rule update time ≤ 30 days).

[0118] Example 8

[0119] Based on Example 3, the process for determining the service level of the rating layer includes:

[0120] Analyze the parameters that affect the service rating downgrade for each deduction (e.g., downgrade by one level for every 5 points deducted).

[0121] Output the downgrade threshold and associate it with the total score range mapping (e.g., 90-100 is grade A).

[0122] Example 9

[0123] Based on Example 1, the speech transcription and semantic enhancement steps include:

[0124] Call a customized ASR model for acoustic decoding and output text with timestamps;

[0125] Cascaded LLM semantic enhancement modules to correct typos in industry terminology;

[0126] Edit distance threshold dynamic adaptation (e.g., edit distance threshold ≤ 2), separate customer service and customer audio channels and mark left / right audio channels;

[0127] The NRC sentiment probability values ​​were labeled and normalized.

[0128] The enhanced text is transmitted to the hierarchical causal inference step, where the semantic enhancement results are fed back to the ASR training set, triggering an incremental fine-tuning closed-loop mechanism.

[0129] Example 10

[0130] Based on Example 1, the hierarchical causal reasoning steps adopt a lightweight deployment scheme:

[0131] Use a large number of parameters in a 7B language model (such as LLaMA-2-7B) to generate causal transmission paths;

[0132] The integrated rules engine performs localized score deduction calculations.

[0133] By controlling resource consumption through single-node deployment, it supports the processing of 10,000 quality inspections per day with a single processing time of ≤5 seconds.

[0134] Explanation of proper nouns

[0135] RESTful API: A network interface based on the REST architecture that uses HTTP requests to transmit data, used for transmitting audio files.

[0136] Message queues: An asynchronous communication mechanism that decouples producers and consumers and improves system reliability.

[0137] PCM: Pulse Code Modulation, a method for digitizing analog signals to store raw audio data.

[0138] Sampling rate: Number of samples per second, 16kHz covers the range of frequencies audible to the human ear.

[0139] Bit rate: The number of bits transmitted per second; 16 bits ensures high fidelity for voice communication.

[0140] Distributed storage architecture: Data is stored across multiple nodes, supporting high scalability (such as HDFS, MinIO).

[0141] HDFS: Hadoop Distributed File System, suitable for large-scale data storage.

[0142] MinIO: High-performance distributed object storage, compatible with Amazon S3 API.

[0143] ASR Model: Automatic Speech Recognition Model, based on the Conformer architecture to achieve high-precision transcription.

[0144] Conformer architecture: A hybrid architecture combining CNN and Transformer to optimize speech recognition.

[0145] End-to-end model: A model that directly generates output from input, simplifying the processing flow.

[0146] Large language models: Deep learning models trained on massive amounts of text (such as GPT-4 and LLaMA-2).

[0147] Timestamp: An event time stamp with millisecond-level precision.

[0148] Voiceprint feature clustering: distinguishing speakers by voiceprint features.

[0149] NRC Emotion Tags: Emotion classification tags based on an emotion dictionary (such as "anger" or "satisfaction").

[0150] The Sigmoid function is a mathematical function that maps numerical values ​​to the interval [0,1].

[0151] Rule knowledge base: A database (JSON or Protobuf format) that stores scoring rules.

[0152] Causal transmission path: a hierarchical logical structure that describes the causal relationship between events and rules.

[0153] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A customer service voice intelligent quality inspection method, characterized in that, The method includes: Data storage steps: Receive customer service call recording files and store them in the original recording library through the input / output layer interface; Speech transcription and semantic enhancement steps: Decode the audio using a customized ASR model to generate preliminary transcribed text with timestamps, use a large language model to correct typos and optimize speaker separation, and add emotion tags to generate the final transcribed text; The hierarchical causal reasoning process involves: retrieving a pre-defined rule knowledge base, matching the final transcribed text with scoring rules, generating a 7-layer causal transmission path constraint large language model to derive a structured quality inspection report. The 7-layer causal transmission path includes: Event Location Layer: extracting timestamp semantic summaries; Rule Matching Layer: matching knowledge base entries; Description and Explanation Layer: logically associating events with rules; Deduction Calculation Layer: calculating individual deductions; Dimensional Impact Analysis Layer: applying classification weights; Total Score Calculation Layer: calculating global deductions; and Rating Determination Layer: outputting service level. Quality inspection report output steps: Transmit the JSON format report to the API interface; The model optimization and continuous learning steps are as follows: Integrate the quality inspection report with manually reviewed feedback and labeled data, perform domain fine-tuning training and DPO reinforcement learning, and update the inference model weights to optimize subsequent quality inspection results. The manually reviewed feedback and labeled data includes: a standardized objection label sub-step: creating a set of objection labels containing customer interruptions, sentiment misjudgments, rule matching errors, and event omissions; and an additional confidence parameter sub-step: attaching a confidence parameter to the objection labels, with values ​​of either "1" for manually labeled labels or the probability value output by the model. After storing the labeled objection labels in the feedback database, an incremental training scheduling task is triggered.

2. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, The hierarchical causal reasoning steps include: The sub-step for dynamically adjusting the deduction weight is as follows: associate the importance parameters of the business scenario, calculate and update the deduction value by weight, and sort the causal transmission path according to the logical hierarchy.

3. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, The continuous optimization learning steps of the model include: Verification feedback data volume threshold sub-step: When the number of labeled data reaches 500, trigger training with a domain fine-tuning learning rate of 5e-5, and train for 3 rounds; DPO reinforcement learning sub-step: When 1000 data points are reached, DPO reinforcement learning is triggered to generate high-quality reports and low-quality reports as comparison data pairs to calculate and update the parameters of the preference function.

4. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, DPO reinforcement learning includes: The sub-step for constructing the reward signal involves weighted calculation of the causal transmission path integrity weight, the deduction basis consistency weight, the semantic accuracy weight, and the rule timeliness weight. Normalization scoring sub-step: Normalize the reward score to the [-1,1] interval, optimize the preference function parameters, and aggregate the output to update the model weights.

5. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, The structured quality inspection report undergoes three levels of verification before output, including: Level 1: Detection of rule levels for causal breakage and missing rules; Level 2: Verify the contradiction of total deduction score overflow; Level 3: Timeliness of query rule knowledge base version; If any level of verification fails, the model will be re-inferenced and an error log will be logged.

6. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, Determining the service level of the rating layer includes: analyzing the impact parameters of individual deductions on service rating downgrades; outputting downgrade thresholds, and mapping the total score range to the final service level of the quality inspection.

7. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, In the hierarchical causal reasoning process, the lightweight model combination is replaced, and the 7B language model with a large number of parameters is called to generate the causal transmission path. The rule engine is integrated to perform the deduction calculation locally. Resource consumption is controlled through single-node deployment to support the daily processing time of 10,000 quality inspections ≤ 5 seconds.

8. The intelligent quality inspection method for customer service voice according to claim 1, characterized in that, The speech transcription and semantic enhancement steps include: The system calls upon a customized ASR model for acoustic decoding, outputs timestamped text, cascades an LLM semantic enhancement module, corrects typos in industry terminology, dynamically adapts distance thresholds, separates customer service voice channel mapping to the left channel identifier, labels the NRC sentiment probability value for normalized output, and transmits the enhanced text to the hierarchical causal inference step. The semantic enhancement results are fed back to the ASR training set, triggering an incremental fine-tuning closed-loop mechanism.

Images