A live broadcast playback quality inspection method and system

Abstract

The application discloses a live broadcast playback quality inspection method and system, which comprises the following steps: obtaining live broadcast playback audio and video data to be inspected and a preset quality inspection standard; performing speech transcription on the audio in the audio and video data to obtain time-stamped transcription text; calling a large language model based on the quality inspection standard to analyze the transcription text in at least two quality inspection dimensions, obtaining judgment information, score information of each dimension, and evidence text segments and their time stamps associated with the judgment information and / or score information; generating a quality inspection report based on the analysis result, which contains a comprehensive score, dimension score information, problem items, evidence segment time stamps, and rectification suggestions. The above scheme can convert playback quality inspection into a structured process that can be automatically executed, making the quality inspection output have a unified basis and can be located and verified by time stamp, reducing the need for full-time manual viewing and improving the overall coverage capacity.

Description

Technical Field

[0001] This application relates to the field of live broadcast replay quality inspection technology, and in particular to a live broadcast replay quality inspection method and system. Background Technology

[0002] With the rapid development of short videos and live-streaming e-commerce, especially in scenarios such as car sales, used car promotion, and online customer acquisition for brick-and-mortar stores, it has become commonplace for live streamers to introduce brand backgrounds, vehicle information, after-sales policies, and transaction processes to users. Live-streaming content not only directly influences users' purchasing decisions and brand image but is also subject to constraints from advertising compliance, platform rules, and consumer rights protection requirements. Therefore, conducting quality checks and compliance reviews of live-streaming replays (hereinafter referred to as "quality control") to ensure the completeness of key information explanations, standardized expression, and the truthfulness and fulfillment of promises, as well as to promptly identify risky wording and misleading statements, is of significant practical importance.

[0003] In existing technologies, live-stream quality inspection is still primarily conducted manually: inspectors watch the live stream replay and meticulously record items such as brand introduction, vehicle parameters and condition descriptions, warranty and after-sales commitments, price and discount information, prohibited words and inappropriate expressions according to a checklist, and then form quality inspection conclusions and rectification suggestions. To improve efficiency, some solutions use general speech recognition to convert speech into text, and then perform hit judgments based on keywords or rule bases, or filter and alert on sensitive words. However, the above methods still have significant shortcomings: First, manual review requires long-term viewing of replays, often only allowing for sampling inspections, making it difficult to achieve full coverage, and the differences in experience among different reviewers can lead to inconsistent judgment standards; Second, general speech recognition is prone to recognition errors when dealing with automotive terminology, vehicle configurations, accents, and changes in speech speed, thus affecting the reliability of subsequent judgments; Third, keyword / rule-based detection struggles to understand contextual semantics and implicit commitments, making it difficult to identify complex issues such as paraphrasing, suggestive advertising, and contradictions, and the location of specific segments where problems occur is not precise enough, resulting in insufficient traceability of quality inspection conclusions and timeliness of rectification feedback.

[0004] Therefore, in the quality inspection of live replays, the time-consuming inspection process and the difficulty in achieving full coverage, the susceptibility of inspection standards to subjective human influence resulting in insufficient consistency and objective quantification, and the weak ability to identify complex semantics and compliance risks and accurately locate problematic segments have become urgent technical problems that need to be solved. Summary of the Invention

[0005] This application provides a live replay quality inspection method and system, aiming to solve the problems of existing technologies in live replay quality inspection, such as time-consuming inspection process, difficulty in achieving full coverage, quality inspection standards being easily affected by human subjectivity and lacking consistency and objective quantification, and weak ability to identify complex semantics and compliance risks and accurately locate problematic segments.

[0006] Firstly, a method for quality inspection of live stream replays, the method comprising:

[0007] Acquire the audio and video data of the live broadcast replay to be inspected, as well as the preset quality inspection standards;

[0008] The audio in the live playback audio and video data is transcribed into speech to obtain transcribed text with timestamps;

[0009] Based on the quality inspection standards, the large language model is invoked to analyze the transcribed text in at least two quality inspection dimensions to obtain the analysis results of each quality inspection dimension. The analysis results include at least: dimension judgment information, dimension score information, and evidence text fragments associated with the dimension judgment information and / or dimension score information and their timestamps.

[0010] A quality inspection report is generated based on the analysis results of each quality inspection dimension. The quality inspection report includes a comprehensive score, score information for each quality inspection dimension, problem items and their corresponding evidence text fragments and timestamps, and rectification suggestions.

[0011] Optionally, the quality inspection standards in the above scheme include: a set of quality inspection dimensions, a set of inspection points corresponding to each quality inspection dimension, and scoring rules; the quality inspection dimensions include at least one or more of the following dimensions: brand introduction dimension, after-sales service dimension, vehicle condition description dimension, compliance dimension, and professional dimension.

[0012] Optionally, in the above scheme, before performing speech transcription on the audio in the live playback audio and video data, audio preprocessing is performed on the audio. The audio preprocessing includes: noise reduction processing, volume normalization processing, speech rate normalization processing, accent / dialect adaptation processing, pause and tone marking processing, and processing to expand or constrain the recognition vocabulary based on the automotive domain lexicon.

[0013] In the above scheme, optionally, the speech transcription is implemented by a streaming speech recognition model or a non-streaming speech recognition model; the transcribed text with timestamps includes at least the segment start time, the segment end time, and the segment text content.

[0014] Optionally, after obtaining the transcribed text, the above scheme further includes performing text post-processing on the transcribed text. The text post-processing includes: punctuation restoration, interjection filtering, merging of repeated segments, typo correction, sentence segmentation according to semantic boundaries, and splicing processing that retains the segment-level timestamp index.

[0015] Optionally, in the above scheme, the step of calling a large language model to analyze the transcribed text based on the quality inspection standard and obtaining the analysis results of each quality inspection dimension includes: constructing corresponding quality inspection prompt word templates for different quality inspection dimensions, and inputting the transcribed text and the corresponding quality inspection prompt word templates into the large language model to obtain the analysis results of the corresponding quality inspection dimensions; wherein, the quality inspection prompt word template includes at least three of the following fields: quality inspection dimension identifier, inspection point list, scoring rules, output format constraints, example sample, evidence fragment output requirements, and timestamp field constraints.

[0016] Optionally, in the above scheme, the step of calling a large language model to analyze the transcribed text based on the quality inspection standard and obtaining the analysis results of each quality inspection dimension further includes: performing information extraction on the transcribed text and / or the analysis results of each quality inspection dimension to obtain a set of structured elements corresponding to the quality inspection standard; and performing semantic matching between the set of structured elements and the inspection points in the quality inspection standard to determine the covered items and missing items.

[0017] Optionally, the structured element set in the above scheme includes: brand name element, model / configuration element, vehicle year element, mileage element, vehicle condition level element, inspection report element, warranty terms element, maintenance terms element, and discount or commitment element.

[0018] Optionally, in the above scheme, the method further includes: performing consistency verification on the statements at different time positions in the transcribed text, outputting the consistency verification result, and writing the consistency verification result into the quality inspection report.

[0019] Secondly, a live broadcast replay quality inspection system, the system comprising:

[0020] The data acquisition module is used to acquire live playback audio and video data as well as preset quality inspection standards;

[0021] The speech-to-text module is used to perform speech-to-text transcription on the audio in the audio and video data to obtain transcribed text with timestamps;

[0022] The model quality inspection module is used to call a large language model to analyze the transcribed text in at least two quality inspection dimensions based on the quality inspection standards, so as to output the analysis results of each quality inspection dimension. The analysis results include at least dimension judgment information, dimension score information, and evidence text fragments and timestamps.

[0023] The report generation module is used to generate a quality inspection report based on the analysis results. The quality inspection report includes a comprehensive score, score information for each quality inspection dimension, problem items and their corresponding evidence text fragments and timestamps, and rectification suggestions.

[0024] Compared with the prior art, this application has at least the following beneficial effects:

[0025] This application, based on further analysis and research into existing technical problems, recognizes that existing technologies in live replay quality inspection suffer from several shortcomings: the inspection process is time-consuming and difficult to achieve full coverage; quality inspection standards are easily influenced by human subjectivity, resulting in insufficient consistency and objective quantification; and the ability to identify complex semantics and compliance risks, as well as accurately locate problematic segments, is weak. By acquiring live replay audio and video data, the application first transcribes the audio into timestamped text, transforming the subsequent quality inspection target from "audiovisual content that must be watched in its entirety" into "a text sequence that can be directly calculated and analyzed." This logically reduces the reliance on manual segment-by-segment viewing and supports automated processing of the entire replay. Simultaneously, based on preset quality inspection standards, a large language model is invoked to analyze the transcribed text across at least two quality inspection dimensions. Each dimension output includes a dimension judgment, a dimension score, and evidence text fragments associated with the judgment / score along with their timestamps. This means... The quality inspection conclusions are structured results formed across multiple dimensions based on the same set of standards, thus providing a unified basis for judgments and scores obtained from different sessions and by different implementing entities. Furthermore, the large language model analyzes the text and outputs evidence fragments bound to timestamps, enabling quality inspection to not only judge the content and generate problem items at the semantic level, but also to match problem items with specific time positions in the playback, ensuring that the conclusions are traceable, verifiable, and can directly locate the rectification location. Finally, the results from all dimensions are summarized to form a quality inspection report that includes a comprehensive score, problem item evidence and timestamps, and rectification suggestions. This transforms the quality inspection output from scattered subjective records into a quantifiable, locatable, and reusable data report, thus providing a corresponding solution to the pain points of existing technologies such as "time-consuming quality inspection leading to difficulty in full coverage, inconsistent standards leading to insufficient consistency, and weak ability to locate complex semantic problems and problem fragments." Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a live replay quality inspection method provided in one embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] Definitions of terms in this embodiment:

[0029] DeepSeek Large Model: A large language model open-sourced by DeepSeek, possessing powerful text understanding and reasoning capabilities;

[0030] Live replay: Video content automatically saved by the system after the streamer finishes the live broadcast;

[0031] ASR technology: Automatic Speech Recognition technology, used to convert speech into text;

[0032] Quality control dimensions: various standards for evaluating the quality of live stream content, including brand introduction, after-sales service, vehicle condition description, etc.

[0033] Host: A staff member responsible for explaining car sales information during live broadcasts;

[0034] Compliance testing: Checking whether the live stream content complies with platform regulations and industry standards;

[0035] Currently, the platform primarily uses manual spot checks to inspect the quality of livestream content. The traditional inspection process involves quality inspectors reviewing the entire replay after the livestream, meticulously checking whether the streamer has adequately presented key information such as brand details, after-sales policies, and vehicle condition. Based on the company's established script guidelines and content standards, inspectors manually record any issues and generate an inspection report. If the streamer omits key information or makes inaccurate statements, training and corrective measures are required.

[0036] This quality inspection method relies on the professional competence and sense of responsibility of the quality inspectors. It requires the quality inspectors to have knowledge of the automotive industry, be familiar with the platform rules, and be able to watch video content attentively for extended periods of time.

[0037] Traditional manual quality inspection methods for live streaming content have many problems:

[0038] The first problem is the extremely low efficiency of quality inspection. Each live stream typically lasts from 30 minutes to 2 hours, and quality inspectors need to watch each live stream in its entirety to make an evaluation. With the increasing number of streamers on the platform, there may be hundreds of live streams every day. Manual quality inspection simply cannot cover all live stream content, and can only be done through random checks, resulting in a large amount of live stream content not being effectively supervised.

[0039] The second problem is the difficulty in standardizing quality inspection criteria. Different quality inspectors may have different judgments on the same live broadcast; some inspectors are strict, while others are relatively lenient. For subjective standards such as "whether the brand is fully introduced" and "whether the vehicle condition description is detailed," it is difficult to form a unified evaluation standard, which makes it impossible to guarantee the fairness and consistency of the quality inspection results.

[0040] The third problem is the lack of real-time feedback. Since manual quality control can only be conducted after the live stream ends, the streamer only receives feedback before the next stream, making timely adjustments and improvements impossible. This delayed feedback mechanism reduces the actual effectiveness of quality control.

[0041] The fourth issue is the high labor costs. A professional quality inspection team is required, and these inspectors not only need to understand automotive knowledge but also be familiar with the platform's rules. The training period is long, resulting in high labor costs. Furthermore, prolonged video viewing can easily lead to fatigue, affecting the quality of the inspection.

[0042] The fifth problem is the difficulty in quantifying and evaluating. Manual quality inspection can often only provide qualitative evaluations, such as "the introduction is not detailed enough" or "the wording is not standardized enough," lacking objective quantitative indicators, making it difficult to accurately measure the anchor's performance and track the improvement effect.

[0043] To address the aforementioned problems, in one embodiment, such as Figure 1 As shown, a method for quality inspection of live stream replays is provided, including the following steps:

[0044] Acquire the audio and video data of the live broadcast replay to be inspected, as well as the preset quality inspection standards;

[0045] The audio in the live playback audio and video data is transcribed into speech to obtain transcribed text with timestamps;

[0046] Based on the quality inspection standards, the large language model is invoked to analyze the transcribed text in at least two quality inspection dimensions to obtain the analysis results of each quality inspection dimension. The analysis results include at least: dimension judgment information, dimension score information, and evidence text fragments associated with the dimension judgment information and / or dimension score information and their timestamps.

[0047] A quality inspection report is generated based on the analysis results of each quality inspection dimension. The quality inspection report includes a comprehensive score, score information for each quality inspection dimension, problem items and their corresponding evidence text fragments and timestamps, and rectification suggestions.

[0048] In this embodiment, the quality inspection standard includes: a set of quality inspection dimensions, a set of inspection points corresponding to each quality inspection dimension, and scoring rules; the quality inspection dimensions include at least one or more of the following dimensions: brand introduction dimension, after-sales service dimension, vehicle condition description dimension, compliance dimension, and professional dimension.

[0049] In this embodiment, before performing speech transcription on the audio in the live playback audio and video data, audio preprocessing is also performed on the audio. The audio preprocessing includes: noise reduction processing, volume normalization processing, speech rate normalization processing, accent / dialect adaptation processing, pause and tone marking processing, and processing to expand or constrain the recognition vocabulary based on the automotive domain lexicon.

[0050] In this embodiment, the speech transcription is implemented by a streaming speech recognition model or a non-streaming speech recognition model; the transcribed text with timestamps includes at least the segment start time, the segment end time, and the segment text content.

[0051] In this embodiment, after obtaining the transcribed text, the text post-processing is further performed on the transcribed text. The text post-processing includes: punctuation restoration, interjection filtering, merging of repeated segments, typo correction, sentence segmentation according to semantic boundaries, and splicing processing that retains the segment-level timestamp index.

[0052] In this embodiment, the step of calling a large language model to analyze the transcribed text based on the quality inspection standard and obtaining the analysis results for each quality inspection dimension includes: constructing corresponding quality inspection prompt word templates for different quality inspection dimensions, and inputting the transcribed text and the corresponding quality inspection prompt word templates into the large language model to obtain the analysis results for the corresponding quality inspection dimensions; wherein, the quality inspection prompt word template includes at least three of the following fields: quality inspection dimension identifier, inspection point list, scoring rules, output format constraints, example sample, evidence fragment output requirements, and timestamp field constraints.

[0053] In this embodiment, the step of calling a large language model to analyze the transcribed text based on the quality inspection standard and obtaining the analysis results of each quality inspection dimension further includes: performing information extraction on the transcribed text and / or the analysis results of each quality inspection dimension to obtain a set of structured elements corresponding to the quality inspection standard; and performing semantic matching between the set of structured elements and the inspection points in the quality inspection standard to determine the covered items and missing items.

[0054] In this embodiment, the structured element set includes: brand name element, model / configuration element, vehicle year element, mileage element, vehicle condition level element, inspection report element, warranty terms element, maintenance terms element, and discount or commitment element.

[0055] In this embodiment, the method further includes: performing consistency checks on the statements at different time positions in the transcribed text, outputting consistency check results, and writing the consistency check results into the quality inspection report.

[0056] The construction of the professional training dataset in this embodiment includes:

[0057] We collected historical live stream replay data from the platform to construct a labeled dataset of 500 car sales live streams. The dataset covers diverse scenarios with different anchor styles, car brands, and live stream durations. 300 of these are high-quality, manually labeled live stream samples, including standard scripts such as complete brand introductions, detailed after-sales policy explanations, and comprehensive vehicle condition descriptions. The remaining 200 samples contain various issues, including: omitted brand introductions, unclear after-sales policy explanations, incomplete vehicle condition descriptions, and instances of inappropriate statements.

[0058] Establish a multi-dimensional quality inspection and evaluation system for live-streaming car sales. Design standardized quality inspection dimension templates:

[0059] Brand introduction dimension: Check whether the host mentions key information points such as brand name, brand history, brand positioning, and brand advantages, and mark it as "complete introduction", "partial introduction" or "not introduced".

[0060] After-sales service dimension: Check whether the host explains the warranty policy, maintenance service, after-sales commitment, repair network, etc., and mark the coverage of service items and the detail of the explanation.

[0061] Vehicle condition description dimension: Check whether the host introduces information such as the vehicle's year, mileage, vehicle condition level, inspection report, vehicle configuration, and usage history, and ensure the completeness and accuracy of the information.

[0062] Compliance dimension: Check whether the anchor's words contain illegal content such as false advertising, exaggerated promises, or disparaging competitors, and mark the type and severity of the violation.

[0063] Professionalism dimension: Evaluate whether the anchor's expression is clear and fluent, whether the terminology is used accurately, and whether the logic is coherent, and mark the level of professionalism.

[0064] A team of experienced automotive sales trainers and quality inspection experts completes the annotations according to unified professional standards. Each live broadcast generates a structured quality inspection report, including scores for each dimension, timestamps pinpointing specific issues, and improvement suggestions.

[0065] Specialized data preprocessing is performed to address the characteristics of live audio:

[0066] Audio noise reduction processing: Remove background noise, music, sound effects and other interference factors in the live broadcast environment to extract the pure broadcaster's voice signal.

[0067] Speech rate normalization: handles the differences in speech rate among different broadcasters and standardizes the alignment of audio duration with text.

[0068] Dialectal accent adaptation: Establish an accent recognition and standardized mapping model based on the accent characteristics of broadcasters from different regions.

[0069] Pause marking: Mark the pauses, stresses, and interjections of the broadcaster to aid in understanding the key semantic points.

[0070] Domain-specific terminology database: Establish a professional terminology database for the automotive industry, including proprietary terms such as brand names, vehicle models, technical parameters, and configuration names.

[0071] This embodiment of the intelligent quality inspection system based on the DeepSeek large model includes:

[0072] By directly leveraging the powerful semantic understanding and reasoning capabilities of the DeepSeek large language model, and through carefully designed prompt engineering and few-shot learning, the task of live-streaming car quality inspection is implemented.

[0073] Quality Inspection Notification Template Design: Develop standardized notification templates for different quality inspection dimensions. For example:

[0074] Brand introduction detection prompts: Please analyze the following live stream text to determine whether the host introduced a car brand. You need to check if the following information was mentioned: brand name, brand history, brand characteristics, brand advantages, etc. Please provide your judgment result (yes / no), confidence level (0-1), specific content snippets mentioned, and timestamps.

[0075] After-sales service inspection prompts: Please analyze the live stream text to determine whether the host explained after-sales service. Check if it includes: warranty policy, maintenance services, after-sales commitments, service outlets, etc. Please list the mentioned service items and assess the level of detail in the explanations (detailed / general / brief / not mentioned).

[0076] Vehicle Condition Description and Inspection Tips: Please analyze the live stream text and assess the comprehensiveness of the host's description of the vehicle's condition. Key information points to check include: vehicle year, mileage, vehicle condition rating, inspection report, exterior and interior condition, mechanical condition, accident history, etc. Please list the information points already presented and any omissions.

[0077] Few-shot example construction: Embed 3-5 labeled examples in the prompt words to demonstrate the standard quality inspection analysis process and output format, and guide the model to understand the specific requirements of the quality inspection task.

[0078] Design a parallel quality inspection architecture by leveraging the multi-task processing capabilities of the DeepSeek model:

[0079] Dimensional decoupling analysis: The quality inspection task is broken down into independent sub-tasks, and each dimension is analyzed using a specific prompt word template. Dimensions such as brand introduction, after-sales service, vehicle condition description, compliance, and professionalism are processed in parallel without interference from each other.

[0080] Information Extraction Module: Utilizes the model's information extraction capabilities to accurately locate and extract key information from live text. For example, it identifies brand names, extracts warranty period numbers, and locates vehicle condition descriptions to build a structured information graph.

[0081] Semantic matching judgment: The extracted information is semantically matched with predefined quality inspection standards. It is not a simple keyword matching, but based on deep semantic understanding, to determine whether the anchor's statement substantially meets the quality inspection requirements. For example, although the anchor's statement "This brand has a long history" and "This brand was founded in the 1930s" are different expressions, they both meet the requirement of "mentioning the brand's history".

[0082] Logical consistency check: This checks for contradictions in the anchor's statements at different times. For example, if the anchor says "the car is in excellent condition" but later says "it needs repairs," the model will identify this logical inconsistency and mark it as a potential problem.

[0083] Based on the multi-dimensional quality inspection results, a quantitative score and a detailed report are generated:

[0084] Multi-dimensional scoring system: Weights and scoring rules are set for each quality inspection dimension. Brand introduction accounts for 20%, after-sales service for 25%, vehicle condition description for 30%, compliance for 15%, and professionalism for 10%. Each dimension is scored from 0 to 100 points based on its completion rate.

[0085] Overall Quality Inspection Score: The overall quality inspection score is calculated based on the scores and weights of each dimension, providing a direct reflection of the overall quality of the live stream. A score of 60 is set as the passing grade, 80 or above is considered excellent, and scores below 60 require improvement.

[0086] Problem identification: For dimensions that do not meet quality inspection standards, pinpoint the exact time point in the live stream where the problem segment occurred. For example, "The host introduced the brand history from 15 minutes 30 seconds to 16 minutes 20 seconds, but did not mention the brand's advantages; it is recommended to supplement this information."

[0087] Improvement suggestion generation: Leveraging the text generation capabilities of the DeepSeek model, targeted improvement suggestions are provided to broadcasters. This not only points out problems but also offers specific examples of delivery techniques and suggestions for better expression.

[0088] Comparative analysis: The performance of this live stream is compared with the streamer's historical average level and the platform's standards for excellent streamers, generating trend analysis charts to help the streamer understand their room for improvement.

[0089] Establish a special mechanism for compliance testing of live streaming content:

[0090] Illegal Language Identification Library: This library includes sample examples of illegal language such as false advertising, exaggerated promises, disparaging competitors, and misleading consumers. The DeepSeek model is used for semantic similarity matching to identify whether the broadcaster uses similar statements.

[0091] Sensitive word detection: Checks live stream text for prohibited industry terms and legally risky words. It not only detects precise matches but also identifies implicit expressions through semantic understanding.

[0092] Risk Level Assessment: The identified potential violations are assessed for risk level and divided into three levels: high risk (e.g., false advertising), medium risk (e.g., imprecise wording), and low risk (e.g., inappropriate wording).

[0093] Real-time early warning mechanism: Real-time early warnings are triggered for high-risk content, which are given priority in the quality inspection report, and a manual review process can be initiated when necessary.

[0094] The system deployment and optimization in this embodiment include:

[0095] Design an efficient system architecture to support large-scale live streaming quality inspection:

[0096] Task queue management: After the live broadcast ends, the quality inspection tasks are automatically added to the processing queue, and the tasks are scheduled according to the live broadcast duration and priority.

[0097] Parallel processing: Multiple ASR instances and DeepSeek model instances work in parallel to achieve simultaneous quality inspection of multiple live broadcasts, significantly improving processing efficiency.

[0098] Dynamic resource allocation: Computing resources are dynamically allocated based on the live broadcast duration and quality inspection complexity to ensure efficient system operation.

[0099] Results caching mechanism: Establish a results cache for live streams that have undergone quality inspection, supporting fast querying and secondary analysis, and avoiding duplicate calculations.

[0100] Employing multiple technical methods to improve the accuracy of quality inspection:

[0101] Multi-model integration: Multiple models are used to make judgments in parallel for key quality inspection dimensions, and the reliability of the judgment is improved by voting or weighted fusion.

[0102] Confidence assessment: A confidence score is provided for each quality inspection conclusion. Judgments with low confidence can be transferred to the manual review process.

[0103] Feedback and learning mechanism: Collect feedback from quality inspectors on the automatic quality inspection results, and use it regularly to optimize prompt word templates and scoring rules, forming a closed loop of continuous improvement.

[0104] Boundary case handling: For ambiguous boundary cases, establish special handling strategies, such as introducing more detailed sub-dimension judgments and requiring higher levels of evidence support.

[0105] Design a user-friendly interface for displaying quality inspection results:

[0106] Visualized Reports: Using radar charts, bar charts, and other visualization methods to display multi-dimensional scores and intuitively present the anchor's performance.

[0107] Timeline display: Mark the time position of key quality inspection points in the video player, and support one-click jump to the problematic segment.

[0108] Comparative Analysis View: Displays the historical quality inspection data trends of the anchors, comparing them with the platform's average level and the standards for excellent anchors.

[0109] Interactive feedback: Broadcasters and administrators can evaluate and provide feedback on the quality inspection results, helping the system to continuously optimize.

[0110] The above-described embodiments can achieve the following beneficial effects:

[0111] Significantly improved quality inspection efficiency: AI-powered automated quality inspection can process a large number of live stream replays simultaneously. A one-hour live stream can be analyzed and inspected within five minutes, representing a more than 10-fold increase in efficiency compared to manual quality inspection. It can achieve full coverage quality inspection of all live streams on the platform, eliminating reliance on spot checks.

[0112] Completely unified quality inspection standards: Based on a unified AI model and quality inspection rules, all live streams are evaluated according to the same standards, completely eliminating the problem of inconsistent standards caused by differences in personal judgment in manual quality inspection, and ensuring the fairness and comparability of quality inspection results.

[0113] The quality inspection results are objectively quantified: each quality inspection dimension has clear scoring standards and quantitative indicators, generating detailed data reports to facilitate tracking changes in anchor performance, evaluating training effectiveness, and developing improvement plans.

[0114] Precise problem identification: It not only points out the problems of the anchor, but also pinpoints the specific time and content segment in the live broadcast, greatly improving the traceability of the problem and the pertinence of the improvement.

[0115] Highly timely feedback: A quality inspection report can be generated quickly after the live broadcast ends, allowing the streamer to understand their performance and areas for improvement in a timely manner, enabling rapid feedback and continuous optimization.

[0116] Significantly reduced labor costs: Automated quality inspection systems can replace most manual quality inspection work. Quality inspectors only need to handle a small number of boundary cases that AI cannot judge and review high-risk content, reducing labor costs by more than 70%.

[0117] Continuous improvement of anchor capabilities: Through continuous automated quality checks and feedback, anchors are helped to optimize their scripts and expressions, improve their professional level, and ultimately enhance the overall quality of the platform's live streaming content.

[0118] Platform content quality assurance: Comprehensive intelligent quality inspection effectively prevents the spread of illegal content, ensures that all live streams comply with platform regulations and industry standards, reduces platform operational risks, and enhances brand image.

[0119] Data-driven decision-making: The massive amount of accumulated quality inspection data can be used to analyze the effectiveness of anchor training, optimize script standards, and identify industry trends, providing data support for platform operation decisions.

[0120] This embodiment leverages the powerful semantic understanding and reasoning capabilities of the DeepSeek large language model to construct an intelligent live streaming content quality inspection system. The system employs a meticulously designed prompt word engineering approach, decomposing the quality inspection task into multiple independent sub-tasks, with each quality inspection dimension corresponding to a specific prompt word template.

[0121] The brand introduction quality control module uses specific prompts to guide the model in analyzing the live stream text, identifying whether the host mentions key information points such as the brand name, brand history, brand positioning, brand characteristics, and brand advantages. The model not only performs keyword matching but, more importantly, uses deep semantic understanding to determine whether the host's statements substantively introduce the brand. For example, while both "This brand is familiar to everyone" and "This is a German luxury brand founded in 1900" mention the brand, the latter is a valid introduction that meets the quality control standards. The model extracts specific introduction content, locates the corresponding timestamps, assesses the completeness and detail of the introduction, and ultimately provides a judgment of "complete introduction," "partial introduction," or "no introduction," generating a detailed evaluation report.

[0122] The after-sales service quality inspection module checks whether the presenter has explained the warranty policy, maintenance services, after-sales commitments, service outlets, and repair guarantees. The model identifies each after-sales service mentioned by the presenter, extracts specific policy details (such as "3-year / 100,000 km warranty" or "lifetime free maintenance"), and judges the level of detail in the explanation. For cases where the presenter only vaguely mentions "comprehensive after-sales service" without providing specific information, the model will determine that the explanation is insufficient. The final output includes a structured result such as a list of explained service items, missing important items, and a level of detail rating (detailed / average / brief / not mentioned).

[0123] Vehicle Condition Description Quality Inspection Module: This is the most critical quality inspection dimension in used car sales live streams. The model needs to check whether the host has comprehensively introduced key vehicle condition information such as the vehicle's year, mileage, condition level, inspection report results, exterior condition, interior condition, mechanical condition, repair history, and accident records. The model uses information extraction technology to accurately locate and extract each piece of vehicle condition information from the live stream text, establishing a structured representation of the vehicle information. It then compares this representation with standard vehicle condition description requirements, identifying both included and omitted information. For the content presented, the model further evaluates the specificity and accuracy of the description to prevent the host from using vague language to get by.

[0124] The compliance inspection module establishes a mechanism to identify prohibited language in live streams, detecting false advertising, exaggerated promises, misleading consumers, disparaging competitors, and the use of superlative terms. The model not only detects precise violations but, more importantly, identifies implicit violations through semantic understanding. For example, even if the streamer doesn't directly say "best," but says "there's no better choice," the model can recognize this as a disguised absolute statement. The system has a built-in knowledge base containing hundreds of prohibited language patterns, comprehensively identifying various compliance risks through semantic similarity matching and logical reasoning. Identified violations are categorized by risk level: high-risk (e.g., blatant false advertising, legally prohibited words) triggers an alert mechanism; medium-risk (e.g., imprecise expressions, easily misunderstood) are marked as warnings; and low-risk (e.g., inappropriate wording) are included in improvement suggestions.

[0125] Professional Quality Inspection Module: This module assesses the broadcaster's presentation quality and professional level. The model analyzes whether the broadcaster's language is clear and fluent, logically coherent, uses professional terminology accurately, contains grammatical errors and slips of the tongue, and is well-organized. Through natural language processing technology, it identifies grammatical errors, logical contradictions, and disorganized expressions in the text. It also evaluates the broadcaster's professionalism, determining whether standard car sales terminology was used and whether a standard presentation flow was followed.

[0126] All quality inspection modules run in parallel without interfering with each other. Each module independently completes its analysis and outputs structured results. The system then merges the outputs from all modules to generate a comprehensive quality inspection report. The report includes multi-dimensional scores, specific problem location (accurate to the video timestamp), a list of missing content, violation markers, and improvement suggestions. The entire quality inspection process is based on deep semantic understanding, rather than simple keyword matching, accurately determining whether the broadcaster truly meets the quality inspection standards and effectively preventing broadcasters from evading inspection through superficial statements.

[0127] This embodiment establishes a scientific quantitative scoring system, transforming subjective quality inspection judgments into objective data indicators. The system sets clear scoring rules for each quality inspection dimension:

[0128] Brand Introduction Dimension Scoring: Scores are calculated based on the quantity and quality of brand information points mentioned by the presenter. Mentioning the brand name earns 20 points, introducing the brand history earns 30 points, explaining the brand positioning earns 25 points, and outlining the brand's advantages earns 25 points, for a total of 100 points. Each information point receives a corresponding percentage of the score based on the level of detail in the introduction.

[0129] After-sales service scoring: List 8 key after-sales service items (warranty policy, maintenance service, repair guarantee, service outlets, parts supply, after-sales commitment, customer service support, and value-added services). Each item mentioned and explained in detail will receive 12.5 points. Items that are only vaguely mentioned without specific information will receive 50% of the score.

[0130] Vehicle Condition Description Scoring: A list of 10 mandatory vehicle condition information items will be provided (vehicle year, mileage, condition rating, inspection report, exterior condition, interior condition, mechanical condition, repair history, accident record, and vehicle configuration), each worth 10 points. Points will be awarded based on the completeness and accuracy of the information provided.

[0131] Compliance scoring: A deduction system is used, with a base score of 100 points. High-risk violations will result in a deduction of 30-50 points, medium-risk violations 10-20 points, and low-risk issues 5 points. Live streams with serious violations may receive a compliance score of 0.

[0132] Professionalism score: The score is based on four sub-dimensions: fluency of expression (30 points), logical coherence (30 points), accuracy of terminology (20 points), and standardization of speech (20 points).

[0133] The system calculates a comprehensive quality inspection score based on scores across various dimensions and preset weights (brand introduction 20%, after-sales service 25%, vehicle condition description 30%, compliance 15%, professionalism 10%), providing a clear picture of the overall quality of the live stream. Quality inspection levels are also set: 90 points and above is excellent, 80-90 points is good, 60-80 points is satisfactory, and below 60 points is unsatisfactory.

[0134] Based on the quantitative scoring results, the system utilizes the text generation capabilities of the DeepSeek model to automatically generate personalized improvement suggestions. These suggestions are not simply template-based but are generated specifically according to the particular problem.

[0135] For any omissions, the system will clearly indicate, "This live stream did not introduce the brand's history. It is recommended to add a brand background introduction at the beginning. You can refer to the following script: 'Brand XX was founded in XX year and is a leader in the XX field...'"

[0136] For content that is not detailed enough, the system will suggest, "The after-sales warranty policy is too brief. It is recommended to clearly specify the warranty period, mileage limits, coverage, and other specific details."

[0137] For inappropriate expressions, the system will mark the specific offending wording and provide compliant alternative expressions.

[0138] For technical issues, the system will point out specific errors in expression, such as "At 23 minutes and 15 seconds, there is a slip of the tongue, saying 'turbocharger' instead of 'turbocharger'. We suggest paying attention to the accuracy of terminology."

[0139] In one embodiment, a live broadcast replay quality inspection system is provided, comprising:

[0140] The data acquisition module is used to acquire live playback audio and video data as well as preset quality inspection standards;

[0141] The speech-to-text module is used to perform speech-to-text transcription on the audio in the audio and video data to obtain transcribed text with timestamps;

[0142] The model quality inspection module is used to call a large language model to analyze the transcribed text in at least two quality inspection dimensions based on the quality inspection standards, so as to output the analysis results of each quality inspection dimension. The analysis results include at least dimension judgment information, dimension score information, and evidence text fragments and timestamps.

[0143] The report generation module is used to generate a quality inspection report based on the analysis results. The quality inspection report includes a comprehensive score, score information for each quality inspection dimension, problem items and their corresponding evidence text fragments and timestamps, and rectification suggestions.

[0144] The specific implementation details of each module can be found in the above description of the limitations of the live replay quality inspection method, and will not be repeated here.

[0145] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for quality inspection of live broadcast replays, characterized in that, The method includes: Acquire the audio and video data of the live broadcast replay to be inspected, as well as the preset quality inspection standards; The audio in the live playback audio and video data is transcribed into speech to obtain transcribed text with timestamps; Based on the quality inspection standards, the large language model is invoked to analyze the transcribed text in at least two quality inspection dimensions to obtain the analysis results of each quality inspection dimension. The analysis results include at least: dimension judgment information, dimension score information, and evidence text fragments associated with the dimension judgment information and / or dimension score information and their timestamps. A quality inspection report is generated based on the analysis results of each quality inspection dimension. The quality inspection report includes a comprehensive score, score information for each quality inspection dimension, problem items and their corresponding evidence text fragments and timestamps, and rectification suggestions.

2. The method according to claim 1, characterized in that, The quality inspection standards include: a set of quality inspection dimensions, a set of inspection points corresponding to each quality inspection dimension, and scoring rules; the quality inspection dimensions include at least one or more of the following dimensions: brand introduction, after-sales service, vehicle condition description, compliance, and professionalism.

3. The method according to claim 1, characterized in that, Before performing speech-to-speech transcription on the audio in the live playback audio and video data, the audio preprocessing is also performed. The audio preprocessing includes: noise reduction processing, volume normalization processing, speech rate normalization processing, accent / dialect adaptation processing, pause and tone marking processing, and processing to expand or constrain the recognition vocabulary based on the automotive domain lexicon.

4. The method according to claim 1, characterized in that, The speech transcription is implemented by a streaming speech recognition model or a non-streaming speech recognition model; the timestamped transcribed text includes at least the segment start time, the segment end time, and the segment text content.

5. The method according to claim 1, characterized in that, After obtaining the transcribed text, the text post-processing is performed on the transcribed text. The text post-processing includes: punctuation restoration, interjection filtering, merging of repeated segments, typo correction, sentence segmentation according to semantic boundaries, and splicing processing that retains the segment-level timestamp index.

6. The method according to claim 1, characterized in that, The step of analyzing the transcribed text based on the quality inspection standard using a large language model to perform analysis on at least two quality inspection dimensions, and obtaining the analysis results for each quality inspection dimension, includes: constructing corresponding quality inspection prompt word templates for different quality inspection dimensions, and inputting the transcribed text and the corresponding quality inspection prompt word templates into the large language model to obtain the analysis results for the corresponding quality inspection dimensions; wherein, the quality inspection prompt word template contains at least three of the following fields: quality inspection dimension identifier, list of inspection points, scoring rules, output format constraints, example sample, evidence fragment output requirements, and timestamp field constraints.

7. The method according to claim 1, characterized in that, The step of analyzing the transcribed text based on the quality inspection standard by calling a large language model to perform analysis on at least two quality inspection dimensions and obtaining the analysis results of each quality inspection dimension also includes: performing information extraction on the transcribed text and / or the analysis results of each quality inspection dimension to obtain a set of structured elements corresponding to the quality inspection standard; and performing semantic matching between the set of structured elements and the inspection points in the quality inspection standard to determine the covered items and missing items.

8. The method according to claim 7, characterized in that, The structured set of elements includes: brand name element, model / configuration element, vehicle year element, mileage element, vehicle condition level element, inspection report element, warranty terms element, maintenance terms element, and discount or commitment element.

9. The method according to claim 1, characterized in that, The method further includes: performing consistency checks on the statements at different time positions in the transcribed text, outputting consistency check results, and writing the consistency check results into the quality inspection report.

10. A live broadcast replay quality inspection system, characterized in that, The system includes: The data acquisition module is used to acquire live playback audio and video data as well as preset quality inspection standards; The speech-to-text module is used to perform speech-to-text transcription on the audio in the audio and video data to obtain transcribed text with timestamps; The model quality inspection module is used to call a large language model to analyze the transcribed text in at least two quality inspection dimensions based on the quality inspection standards, so as to output the analysis results of each quality inspection dimension. The analysis results include at least dimension judgment information, dimension score information, and evidence text fragments and timestamps. The report generation module is used to generate a quality inspection report based on the analysis results. The quality inspection report includes a comprehensive score, score information for each quality inspection dimension, problem items and their corresponding evidence text fragments and timestamps, and rectification suggestions.

Images