AI voice badge-based shopping guide ability improvement system and method

Abstract

The application discloses a system and method for improving the ability of a guide based on AI voice badges, and belongs to the technical field of artificial intelligence and intelligent guide, the system comprising a multi-modal voice acquisition and feature extraction module, a crown model dynamic extraction and scoring module, a dialogue quality evaluation and intention analysis module, and a closed-loop feedback optimization and accompanying training scene generation module, the four modules forming a deep-coupled closed-loop collaborative system, the system extracting efficient rhetoric from crown samples through an innovative adaptive weight calculation algorithm, quantifying customer intention intensity through a multi-dimensional fusion method, improving the guide's ability through interactive accompanying training, and realizing system adaptive optimization through a closed-loop feedback mechanism, the application realizing the automatic inheritance of crown experience, the scientific evaluation and systematic improvement of the guide's ability, significantly improving customer satisfaction and sales conversion rate, and being suitable for the industry of general home, retail and other industries requiring guide services.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, speech recognition, natural language processing, and intelligent sales guidance, specifically to a system and method for enhancing sales guide capabilities based on AI voice badges. This system collects and analyzes dialogue data between sales guides and customers, automatically extracts top salesperson experience, and constructs standardized models to achieve accurate assessment of sales guide service quality and targeted capability enhancement. It is applicable to industries requiring sales guidance services, such as home furnishings, retail, and automobile sales. Background Technology

[0002] As the digital transformation of the retail industry deepens, sales associates, as a key link connecting businesses and customers, directly impact customer experience and sales conversion rates through their service capabilities. However, the home furnishing industry and other retail sectors still face numerous challenges in cultivating sales associate skills and managing customer reception.

[0003] First, the experience of top sales performers is difficult to pass on effectively. The customer service skills, sales strategies, and problem-solving methods accumulated by excellent sales associates over long-term practice often rely on personal experience. Traditional methods such as experience-sharing sessions and written records are difficult to systematically extract and replicate. Especially at different customer interaction stages, such as needs assessment, competitor comparison, and price negotiation, the specific sales techniques and strategies of top performers lack a structured extraction mechanism. This makes it difficult for ordinary sales associates to quickly learn and master these valuable experiences, making the large-scale replication of top sales performer experience a pain point in the industry.

[0004] Secondly, the assessment of sales associate capabilities lacks a scientific basis. Existing sales associate assessments mainly rely on the store manager's subjective judgment or a single conversion rate indicator, lacking multi-dimensional analysis of the sales associate's service process. For example, it is impossible to quantitatively assess the sales associate's execution at each SOP step, accurately identify the customer's true intention strength and resistance type, or precisely pinpoint where the sales associate's capabilities are lacking. This crude assessment method leaves sales associate improvement without a clear direction, resulting in low efficiency in capability enhancement.

[0005] Third, customer service optimization is inefficient. Post-customer debriefing usually requires manual transcription of recordings or notes, which is time-consuming, labor-intensive, and prone to missing key information. In particular, details such as customers' implicit needs, emotional changes, and resistance signals are often difficult to accurately capture and record. At the same time, existing sales training and practice methods mainly involve memorizing scripts or simple role-playing, lacking dynamic scenarios that match the logic of real customer interactions. As a result, sales staff still lack the ability to handle complex customer issues.

[0006] In recent years, artificial intelligence (AI) technology has made significant progress in fields such as speech recognition, natural language processing, and sentiment analysis. Speech recognition technology has achieved an accuracy rate of over 95%, enabling real-time speech-to-text conversion. Sentiment recognition technology can analyze the speaker's emotional state from the prosodic features of speech. Natural language processing technology can understand the semantic content of dialogue and extract key information. The maturity of these technologies provides new solutions for improving sales guidance capabilities.

[0007] However, existing voice analysis systems mainly focus on customer service quality inspection, emphasizing the detection of service standardization, and lack in-depth optimization for sales and marketing scenarios. Existing technologies have the following shortcomings: First, they lack an automated extraction mechanism for top sales experience, making it impossible to identify effective sales techniques and successful strategies from large amounts of dialogue data; second, they lack in-depth analysis capabilities of customer intentions and resistance, failing to accurately quantify customers' purchasing intentions and concerns; and third, they lack a closed-loop feedback optimization mechanism, meaning evaluation results cannot effectively guide the improvement of sales skills and the continuous optimization of the system itself.

[0008] Therefore, there is an urgent need for a salesperson capability improvement system that can automatically extract the experience of top sales performers, accurately assess salesperson capabilities, intelligently generate training scenarios, and form a closed-loop feedback optimization system to systematically solve the problems of long salesperson training cycles, inaccurate capability assessments, and low efficiency in customer service optimization. Summary of the Invention

[0009] The purpose of this invention is to provide a system and method for improving the sales guide's capabilities based on AI voice badges, in order to solve the technical problems in the existing technology, such as the difficulty in scaling up the experience of top sales performers, the lack of multi-dimensional scientific basis for assessing sales guide capabilities, and the low efficiency of customer reception optimization.

[0010] This invention designs a deeply coupled closed-loop collaborative system by establishing a multimodal speech acquisition and feature extraction module, a sales champion model dynamic extraction and scoring module, a dialogue quality assessment and intention analysis module, and a closed-loop feedback optimization and training scenario generation module. Each module achieves forward data transmission, real-time performance evaluation, reverse parameter feedback, and continuous system optimization through deep coupling at the parameter and state levels. The structured dialogue data extracted by the multimodal speech acquisition and feature extraction module provides the basic input for the sales champion model extraction. The speech weight parameters of the sales champion model directly affect the benchmark standard for dialogue quality assessment. The evaluation results trigger closed-loop feedback optimization, and the optimized parameters adjust the key parameters of the preceding modules in reverse, forming a synergistic effect of mutual promotion and cumulative enhancement.

[0011] This invention proposes an innovative adaptive weight calculation algorithm that maps the frequency of sales pitches to their conversion rate using an exponential function, giving significantly higher weight to high-frequency, high-conversion pitches and solving the problem of unreasonable weight allocation in traditional methods. This invention also proposes a multi-dimensional customer intent vectorization method that integrates multiple dimensions such as proactive inquiry, feature focus, and decision time to accurately calculate customer intent strength scores. The closed-loop feedback mechanism constructed in this invention enables training data to back-optimize the top sales performer model and evaluation criteria, achieving adaptive evolution of the system.

[0012] The beneficial effects of this invention are as follows: First, it enables the automated extraction and standardized inheritance of top sales experience. By segmenting the nodes and extracting the sales scripts from top sales samples, a structured top sales model library is constructed, allowing ordinary sales associates to quickly learn the efficient scripts and coping strategies of top sales associates at different reception stages, significantly shortening the training cycle for sales associates and improving the overall service capabilities of the team. Second, it provides a multi-dimensional scientific evaluation system. By detecting the execution of SOP nodes and identifying the strength of customer intention and resistance level, it accurately identifies the shortcomings in sales associate capabilities, providing clear directions and specific guidance for sales associate improvement. Third, it constructs a closed-loop feedback optimization mechanism. The training data can adjust the weights of the top sales model and the evaluation standard thresholds in reverse, achieving continuous evolution of the system and a spiral improvement in sales associate capabilities, ultimately significantly improving customer satisfaction and sales conversion rates. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall structure of the AI ​​voice-based sales guide capability enhancement system of the present invention.

[0014] Figure 2 This is a schematic diagram of the structure of the multimodal speech acquisition and feature extraction module of the present invention.

[0015] Figure 3 This is a schematic diagram of the dynamic extraction and scoring module of the sales champion model of the present invention.

[0016] Figure 4 This is a schematic diagram of the dialogue quality assessment and intention analysis module of the present invention.

[0017] Figure 5 This is a schematic diagram of the closed-loop feedback optimization and training scenario generation module of the present invention.

[0018] Figure 6 This is a flowchart illustrating the method for enhancing the shopping guide capabilities based on AI voice badges, as described in this invention.

[0019] Figure 7 This is a schematic diagram of the adaptive weight calculation process of the present invention.

[0020] Figure 8 This is a schematic diagram of the process for quantifying customer intent intensity in this invention. Detailed Implementation

[0021] Please refer to the attached document. Figures 1-8 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] Reference Figure 1 The AI-based voice badge-based shopping guide capability enhancement system provided by this invention includes a multimodal voice acquisition and feature extraction module 1, a sales champion model dynamic extraction and scoring module 2, a dialogue quality assessment and intention analysis module 3, and a closed-loop feedback optimization and training scenario generation module 4.

[0023] The multimodal voice acquisition and feature extraction module 1 collects real-time dialogue between sales associates and customers using AI voice badges and extracts multi-dimensional features from the voice data. The top salesperson model dynamic extraction and scoring module 2 automatically extracts top salesperson experience from historical dialogue data, constructs a standardized top salesperson model library, and scores sales associate reception. The dialogue quality assessment and intention analysis module 3 detects the sales associate's SOP execution, identifies customer intention strength and resistance level, and generates a comprehensive assessment report. The closed-loop feedback optimization and training scenario generation module 4 generates targeted training scenarios based on the assessment results and feeds the training data back to the front-end module for parameter optimization.

[0024] The modules form a deeply coupled closed-loop collaborative relationship. The structured dialogue data extracted by the multimodal speech acquisition and feature extraction module 1 serves as the key input parameter for the sales champion model dynamic extraction and scoring module 2, and the data quality directly affects the accuracy of the sales champion model. The speech weight parameters and node templates generated by the sales champion model dynamic extraction and scoring module 2 serve as the evaluation benchmark for the dialogue quality assessment and intention analysis module 3, and the weight allocation directly determines the focus of the assessment. The sales guidance capability weakness location output by the dialogue quality assessment and intention analysis module 3 triggers the scene generation logic of the closed-loop feedback optimization and training scenario generation module 4. The training data statistically analyzed by the closed-loop feedback optimization and training scenario generation module 4 generates optimization suggestion parameters. These parameters adjust the speech weight coefficients of the sales champion model dynamic extraction and scoring module 2 and the threshold parameters of the dialogue quality assessment and intention analysis module 3 in reverse, realizing the adaptive optimization of the system.

[0025] This deeply coupled closed-loop collaborative mechanism brings significant synergistic effects. First, there's a mutually reinforcing effect: improved output quality in the front-end module directly improves the performance of the back-end module, while feedback optimization in the back-end module conversely enhances the accuracy of the front-end module. Second, there's a synergistic effect: the overall effect of multiple modules working together far exceeds the simple sum of the effects of each module working independently. Third, there's a conflict-resolving effect: through dynamic weight adjustment and threshold optimization, the system can automatically balance the conflict between speech diversity and standardization, achieving a unity of personalization and standardization.

[0026] Reference Figure 2 The multimodal speech acquisition and feature extraction module 1 includes a voiceprint recognition unit, a speech-to-text unit, an emotion feature extraction unit, and a semantic feature extraction unit.

[0027] The voiceprint recognition unit extracts acoustic features from the acquired speech signal, including Mel-frequency cepstral coefficients, fundamental frequency, and formants. These acoustic features are matched against pre-recorded salesperson and customer voiceprint databases, and a dynamic time warping algorithm is used to calculate the distance between feature sequences. When the similarity exceeds a preset threshold, the speaker's identity is confirmed. The salesperson voiceprint database is entered into the database via the Modou mini-program when a salesperson joins the company, requiring each salesperson to record at least 30 seconds of standard speech samples. The customer voiceprint database is automatically created when a customer first enters the store, and automatically identifies returning customers during subsequent interactions. Preferably, the similarity threshold is set to 0.85, which can reduce false positives while ensuring recognition accuracy.

[0028] The speech-to-text unit employs a deep learning-based automatic speech recognition model, which includes an acoustic model, a language model, and a decoder. The acoustic model maps the acoustic features of the speech signal to a phoneme sequence, the language model predicts the probability distribution of word sequences based on a large-scale text corpus, and the decoder combines the outputs of the acoustic and language models to generate the optimal text sequence. The model uses an end-to-end training approach, fine-tuning it on professional corpora in the home furnishing industry to accurately recognize industry terms such as "E0-grade board" and "silent guide rails." The converted text sequence includes millisecond-level timestamps, facilitating subsequent dialogue node segmentation and temporal analysis. The speech-to-text accuracy reaches over 96% in quiet environments and maintains over 92% accuracy in noisy store environments.

[0029] The emotion feature extraction unit extracts prosodic features from the speech signal, including pitch, intensity, speech rate, and pause duration. Pitch features reflect the speaker's intonation variations and are typically obtained by extracting the fundamental frequency sequence. Intensity features reflect the speaker's volume and are obtained by calculating short-time energy. Speech rate features are obtained by counting the number of syllables per unit time. Pause duration is obtained by detecting the duration of consecutive low-energy segments. Based on these prosodic features, a support vector machine classifier is used to identify the speaker's emotional state, with emotion categories including positive, neutral, negative, anxious, and hesitant. The emotion recognition model is trained on a large-scale emotional speech database, achieving an accuracy rate of 89%. Preferably, when a customer's emotional state is detected to be negative or anxious, the system marks this time point in the evaluation report, prompting the sales associate to pay attention to calming the customer's emotions.

[0030] The semantic feature extraction unit performs natural language processing on the text sequence output by the speech-to-text unit. First, it performs word segmentation and part-of-speech tagging, employing a word segmentation model based on a bidirectional long short-term memory network, which accurately handles industry terminology and new words. Then, it performs named entity recognition to extract key entities from the dialogue, including product names, brand names, price values, and time expressions. Next, it performs dependency parsing to construct grammatical dependencies between words and understand the sentence structure and semantics. Finally, it uses a pre-trained language model to encode the text, generating high-dimensional semantic vector representations. These semantic vectors capture the deep semantic information of the dialogue for subsequent intention recognition and resistance analysis. Preferably, a BERT model pre-trained on e-commerce customer service dialogue data is used for semantic encoding, achieving a semantic similarity calculation accuracy of 91%.

[0031] The multimodal speech acquisition and feature extraction module 1 integrates the outputs of the speaker recognition unit, speech-to-text unit, emotion feature extraction unit, and semantic feature extraction unit to generate structured dialogue data. Each piece of structured dialogue data includes speaker identification, dialogue timestamp, dialogue text content, emotion state label, semantic vector representation, and a list of key entities. This structured dialogue data provides a complete data foundation for subsequent sales champion model extraction, dialogue quality assessment, and intention analysis.

[0032] Reference Figure 3 The sales champion model dynamic extraction and scoring module 2 includes a sales champion sample screening unit, a node segmentation and speech extraction unit, an adaptive weight calculation unit, and a model library construction unit.

[0033] The top-performing sales sample selection unit selects dialogue samples from the historical dialogue database that meet preset success criteria as top-performing sales samples. These preset success criteria include three dimensions: conversion rate, customer satisfaction score, and service duration. The conversion rate requires that the dialogue ultimately result in an order conversion; the customer satisfaction score requires that the customer give a score of 4 or higher (out of 5) in a post-service satisfaction survey; and the service duration must be controlled within a reasonable range—neither too short, leading to insufficient service, nor too long, affecting efficiency. Preferably, the conversion rate is weighted at 0.5, the customer satisfaction score at 0.3, and the service duration at 0.2. A comprehensive score is calculated, and the dialogue samples ranking in the top 20% are selected as top-performing sales samples. This multi-dimensional selection mechanism ensures that the selected top-performing sales samples not only have high conversion rates but also excellent service quality and reasonable efficiency. In one embodiment, 2000 top-performing sales samples are selected from 10,000 historical dialogue data points. These samples cover different customer types, product categories, and service scenarios, and are highly representative.

[0034] The node segmentation and dialogue extraction unit performs node segmentation on the dialogue flow of the top-selling sample. Based on the characteristics of the customer service process in the home furnishing industry, the dialogue is divided into eight main nodes: icebreaking, needs assessment, product recommendation, selling point explanation, competitor comparison, price negotiation, objection handling, and closing the order. Node segmentation employs a hybrid method based on keywords and semantic similarity. First, a set of feature keywords is defined for each node. For example, keywords for the needs assessment node include "need," "want," "budget," "area," and "style," while keywords for the product recommendation node include "recommend," "suitable," "this one," and "series." Then, the matching degree between the dialogue text and the feature keyword set of each node is calculated to initially determine the possible node boundaries. Next, a pre-trained language model is used to calculate the semantic similarity between the dialogue fragment and the standard node description, accurately locating the start and end positions of the nodes. Within each node, high-frequency sales pitches are identified. High-frequency pitch identification uses an N-gram model to extract continuous word sequences, counting the frequency of each word sequence in the top-selling sample, and selecting word sequences with a frequency exceeding a threshold as candidate pitches. Preferably, when a certain sales pitch appears in more than 30% of the top-performing sales samples, it is marked as a high-frequency sales pitch. Simultaneously, typical customer questions and answers are extracted, and a clustering algorithm is used to group semantically similar customer questions into one category. Representative questions are selected from each category as typical questions.

[0035] The adaptive weight calculation unit calculates the adaptive weight of each script, which is one of the core innovations of this invention. Traditional methods typically assign weights to scripts simply based on their frequency of occurrence, ignoring the correlation between scripts and successful conversions. This invention proposes an innovative adaptive weight calculation algorithm that considers both the frequency of script occurrence and the correlation with successful conversions.

[0036] Reference Figure 7 The specific process for adaptive weight calculation is as follows. First, the frequency of the i-th sales pitch in the top-performing salesperson sample is counted. The calculation formula is: , in, Let i be the number of top-performing sales samples containing the i-th sales pitch. This represents the total number of samples from top-selling performers.

[0037] Then, the success conversion rate of the dialogue sample containing the i-th word is calculated. The calculation formula is: , in, Let represent the number of samples containing the i-th sales pitch that ultimately resulted in a sale.

[0038] Next, the importance score of the sales pitch is calculated, using the product of frequency and conversion rate as the base score: , To give significantly higher weight to important phrases, an exponential function is used for non-linear mapping: , in, The temperature coefficient controls the concentration of the weight distribution.

[0039] The exponential mapping values ​​of all dialogue phrases are normalized to obtain adaptive weights: , in, The adaptive weight for the i-th statement. The total number of scripts, This is a scaling factor used to map weight values ​​to a target numerical range.

[0040] In a preferred embodiment of the present invention The value is 2.0. The value is 1.2. It can maintain the difference in weights while avoiding extreme distribution, so that the weight of important phrases is 3 to 5 times that of general phrases, which can highlight the key points without ignoring other valuable phrases. Mapping the weight values ​​to the range of 0 to 1.2 facilitates subsequent weighted calculations.

[0041] This adaptive weighting algorithm achieves a non-linear mapping through an exponential function, giving significantly higher weight to high-frequency, high-conversion sales pitches, while moderately suppressing the weight of low-frequency or low-conversion pitches. This weighting strategy preserves the statistical regularity of frequency of occurrence and incorporates the causal relationship of successful conversions, enabling more accurate identification of truly effective top-selling sales pitches compared to simple frequency statistics methods. Experimental data shows that the top-selling sales model constructed using the adaptive weighting algorithm of this invention improves customer satisfaction by 18% and conversion rate by 23% when guiding ordinary sales associates, significantly outperforming traditional methods based on simple frequency statistics.

[0042] The model library construction unit builds a top-selling customer model library based on node segmentation results, high-frequency dialogues, and adaptive weights. The top-selling customer model library uses a structured storage method, with each node corresponding to a sub-model. Each sub-model contains three parts: a node dialogue template, a customer question library, and a response strategy library. The node dialogue template stores all high-frequency dialogues for that node and their adaptive weights, arranged in descending order of weight. The customer question library stores typical customer questions for that node and their question type tags. The response strategy library stores recommended response dialogues for different customer questions, with each response dialogue associated with corresponding contextual conditions and expected effects.

[0043] In one specific embodiment, the top-selling model for the product recommendation node includes 15 high-frequency sales pitches. The highest-weighted pitch is "Based on your mentioned needs, I suggest you take a look at this series," with a weight of 0.15. The customer question library for this node contains 8 typical question categories, including "What is the price?", "What colors are available?", and "How is the quality?". For questions like "What is the price?", the recommended response is, "This product currently has a promotional offer. Let me first introduce its core advantages to you; you'll find it offers excellent value for money." This response strategy emphasizes value before discussing price, preventing customers from focusing on price too early and overlooking product advantages.

[0044] The top salesperson model library supports dynamic updates. When the accumulated new dialogue data reaches a preset threshold, the system automatically triggers a model update process, re-filters the top salesperson samples, extracts newly added high-frequency phrases, recalculates adaptive weights, and updates the model library content. Preferably, a model update is triggered every 1000 accumulated new dialogue data entries, or a periodic update is triggered every 30 days. This dynamic update mechanism ensures that the top salesperson model always reflects the latest market environment and changes in customer needs, maintaining the model's timeliness and accuracy.

[0045] The Sales Champion Model Dynamic Extraction and Scoring Module 2 performs node matching and scoring on sales guide reception dialogues based on the Sales Champion Model Library. The node matching process first identifies the node in the sales guide dialogue, then extracts the Sales Champion Model for that node, and calculates the similarity between the actual sales guide's script and the standard script in the model. A comprehensive similarity measurement method combining edit distance and semantic similarity is used; when the similarity exceeds a threshold, it is considered a script match. The scoring process calculates the coverage and accuracy of the sales guide's script. Coverage equals the number of matched standard scripts divided by the total number of standard scripts for that node, and accuracy equals the matching score of high-weighted scripts. The node score is calculated by combining coverage and accuracy using a weighted average. Preferably, the coverage weight is 0.4, and the accuracy weight is 0.6, emphasizing the use of high-quality scripts.

[0046] Reference Figure 4 The dialogue quality assessment and intention analysis module 3 includes an SOP node detection unit, a customer intention identification unit, a customer resistance identification unit, and an assessment report generation unit.

[0047] The SOP node detection unit, based on preset SOP node definitions and node dialogue templates in the top salesperson model library, detects the speaking rate and accuracy rate of each SOP node in the sales consultant dialogue. SOP node definitions are set by the enterprise according to its own business processes and typically include key nodes that must be executed and auxiliary nodes that are recommended for execution. The speaking rate measures whether the sales consultant used the prescribed dialogue at the corresponding node; the calculation formula is the number of standard dialogues actually used at that node divided by the total number of standard dialogues required for that node. The accuracy rate measures the accuracy and completeness of the dialogue used by the sales consultant; it is calculated by comparing the actual dialogue with the standard dialogue, with higher similarity scores indicating higher accuracy. In one embodiment, the SOP for the demand mining node requires the sales consultant to inquire about the customer's budget range, usage scenario, and style preference. If the sales consultant only inquires about the budget and scenario, but not style preference, the speaking rate is 67%. The accuracy rate is calculated based on the degree of matching between the salesperson's questioning method and the standard script. The similarity between the standard script "What is your approximate budget for this renovation?" and the salesperson's actual script "What is your budget?" is 0.82, so the accuracy rate of the script is 82%.

[0048] The customer intent recognition unit extracts intent signals from customer conversations. These signals include proactive inquiry behavior, product feature focus, and decision-making timeline expression. Proactive inquiry behavior refers to customers actively asking about transaction-related information such as prices, discounts, and payment methods, reflecting a strong purchase intention. Product feature focus refers to customers inquiring in detail about specific product functions, technical parameters, materials, and processes, reflecting a need for in-depth product understanding. Decision-making timeline expression refers to customers mentioning their purchase timeline, such as "I need to renovate next week" or "I need to make a decision before the end of the month," reflecting a sense of urgency in the purchase.

[0049] Reference Figure 8 The customer intent recognition unit uses a multi-dimensional fusion method to calculate the customer's intent strength score. First, the original feature values ​​for each dimension are extracted. The number of proactive inquiries is obtained by counting the number of questions containing keywords such as price and discounts in customer conversations. The number of product feature concerns is obtained by identifying the number of product attribute types inquired about by customers. For example, if a customer inquires about material, size, and color, the number of concerns is 3. The clarity of the decision-making time is obtained by detecting the specificity of the time expression. Specificity down to the date scores 1.0, specificity down to the week or month scores 0.7, and vague expressions such as "soon" or "under consideration" score 0.3.

[0050] Then, the original feature values ​​for each dimension are normalized. The min-max normalization method is used to map the original values ​​to the interval between 0 and 1. , in, These are the original eigenvalues. and These are the minimum and maximum values ​​of this feature in the training dataset. These are the normalized eigenvalues.

[0051] The customer's intent strength score is obtained by weighting and summing the normalized dimension values ​​based on preset dimension weights. , in, Score the strength of customer intent. This represents the normalized number of proactive price inquiries. This represents the normalized number of product feature concerns. To ensure the clarity of decision-making timelines, , , The weights are for the three dimensions respectively.

[0052] In a preferred embodiment of the present invention , , Inquiry about pricing is given the highest weight because it directly reflects a customer's purchase intention. Focus on product features is given the next highest weight, as a deeper understanding of the product reflects the customer's level of commitment. The weight of decision-making time is relatively low because customers may not explicitly state their purchase time at the beginning of the interaction. This weighting has been validated by extensive historical data and accurately reflects the true strength of a customer's intention. Experimental data shows that customers with an intention strength score greater than 0.7 have a 68% conversion rate, while customers with a score less than 0.3 have a conversion rate of only 12%, indicating that this scoring method has good predictive power.

[0053] The customer resistance identification unit extracts resistance signals from customer conversations. These signals include expressions of price concerns, mentions of competitor comparisons, and expressions of decision-making hesitation. Expressions of price concerns refer to statements from customers indicating that the price is too high, exceeds their budget, or they are demanding a larger discount. Mentions of competitor comparisons indicate that customers mention other brands or competing products, suggesting they are comparing multiple options. Expressions of decision-making hesitation include statements from customers using phrases such as "I'll think about it," "I'll consider it," or "I'll discuss it with my family" to postpone a decision.

[0054] The system determines a customer's resistance level based on extracted resistance signals. Resistance levels are categorized into three levels: high resistance, medium resistance, and low resistance. The evaluation criteria combine rule-based and statistical methods. A customer is considered highly resistant if they simultaneously express price concerns and competitor comparisons, and show hesitation more than twice. A customer is considered medium resistant if they express either price concerns or competitor comparisons, and show hesitation once. A customer is considered low resistant if they only show slight price inquiries or natural hesitation upon first contact. In one example, a customer saying, "Your price is significantly higher than XX brand's; I'll go check out their store and compare before deciding," which simultaneously includes price concerns and competitor comparisons, and clearly expresses a willingness to postpone a decision, is classified as a highly resistant customer. For different resistance levels, the system extracts corresponding coping strategies from the top salesperson model library to guide sales staff in targeted handling.

[0055] The assessment report generation unit integrates the outputs of the SOP node detection unit, customer intention identification unit, and customer resistance identification unit to generate a sales guide reception quality assessment report. The assessment report adopts a structured format and includes five parts: basic information, SOP execution analysis, customer intention analysis, customer resistance analysis, and capability gap identification.

[0056] The basic information section records basic information such as reception time, reception duration, salesperson's name, and customer type. The SOP execution analysis section displays the opening rate and accuracy rate of each step using a radar chart for visualization, intuitively reflecting the salesperson's performance at each step. In the radar chart, the area of ​​the polygon formed by the standard execution is used as the benchmark for the full score, and the ratio of the area of ​​the polygon formed by the salesperson's actual execution to the benchmark is used as the overall execution score. The customer intention analysis section displays the intention strength score, main intention signals, and distribution of concerns, helping salespeople understand customers' purchasing intentions and key concerns. The customer resistance analysis section displays the resistance level, main resistance signals, and recommended coping strategies, guiding salespeople to address customer concerns in a targeted manner.

[0057] The core value of the assessment report lies in identifying skill gaps. The system compares the actual performance of sales associates with the standards of the top salesperson model to identify significant gaps in specific areas and communication styles. This gap identification employs a multi-dimensional comparison method. First, it compares the execution scores of each stage; stages scoring below 60 are marked as weak points. Then, it further analyzes these weak points to identify which types of communication styles are missing or misused. These styles include those related to needs assessment, product recommendation, selling point explanation, objection handling, and closing the sale. Next, it correlates customer intent and resistance information to analyze whether the sales associate has adjusted their communication strategies to suit the specific customer situation. For example, when a customer shows high intent but high resistance, the sales associate should focus on objection handling communication styles; if the sales associate's use of these styles is insufficient, it is marked as a critical weakness.

[0058] In one specific example, sales associate Xiao Zhang served a customer. The evaluation report showed that the customer's intention strength score was 0.76, indicating a high-intent customer, but the resistance level was medium, with the main resistance point being price concerns. Xiao Zhang's performance scores in the demand discovery and product recommendation stages were 85 and 82 points respectively, showing good performance. However, his performance score in the price negotiation stage was only 58 points, indicating a weakness. Further analysis revealed that when faced with the customer's price concerns, Xiao Zhang simply responded, "This price is already very favorable," while the high-weighted response in the top salesperson model would be, "I understand your concern about the price. Actually, this price reflects the true value of our product. Let me explain the costs in detail, and you'll understand why choosing us is more cost-effective." The system flagged Xiao Zhang's insufficient use of value-based argumentation in the price negotiation stage, which was his main weakness.

[0059] Once the evaluation report is generated, it is automatically linked to the recorded audio of the customer interaction. The audio player supports navigation by node; clicking on the performance score of a node automatically jumps to the starting time of that node. During audio playback, the current speaker's role is displayed in real time, with different colors used to indicate the dialogue between the sales associate and the customer. Important events are also marked at key time points, such as the moment the customer first expresses interest, the moment the customer raises objections, and the moment the sales associate uses high-weighted phrases. This audio linking function allows sales associates to review the customer interaction process, combining the textual analysis of the evaluation report with the voice details of the actual recording to gain a deeper understanding of their performance and areas for improvement.

[0060] The assessment report generated by the dialogue quality assessment and intention analysis module 3 provides crucial data support for closed-loop feedback optimization. The capability gaps identified in the assessment report directly trigger the generation logic of the coaching scenario, customer intention and resistance information guide the dynamic adjustment of coaching difficulty, and SOP execution data provides feedback for the weight optimization of the top salesperson model.

[0061] Reference Figure 5 The closed-loop feedback optimization and training scenario generation module 4 includes a shortcoming analysis unit, a training scenario generation unit, an interactive training unit, and a parameter optimization unit.

[0062] The weakness analysis unit extracts the sales guide's skill weaknesses from the evaluation report output by the dialogue quality assessment and intention analysis module 3, identifying SOP nodes and conversation types that need strengthening. Weakness extraction employs a combination of rule matching and statistical analysis. Rule matching identifies items explicitly marked as "weak nodes" or "critical weaknesses" in the evaluation report. Statistical analysis summarizes recent customer service evaluation reports to identify persistently low-scoring nodes and conversation types. For example, if a sales guide scores below 60 points in the price negotiation node in 7 out of the last 10 customer services, price negotiation is identified as a persistent weakness for that sales guide. Preferably, a node is included in the scope of needing strengthening when its average score in recent customer services is below 65 points, or when its score is below 70 points in more than 60% of recent customer services.

[0063] After identifying the SOP nodes that need strengthening, the focus shifts to specific conversation types. In Xiao Zhang's case mentioned earlier, the price negotiation node was identified as a weak point. Further analysis revealed insufficient use of value-based argumentation techniques. Therefore, the conversation type that needed strengthening was determined to be value-based argumentation techniques during the price negotiation node. This refined identification of weaknesses makes training more targeted, avoiding inefficient training based on generalities.

[0064] The training scenario generation unit, based on the SOP nodes and dialogue types that need strengthening identified by the weakness analysis unit, extracts corresponding standard dialogues and typical customer Q&As from the top salesperson model library to generate dialogue strategies simulating the customer role. The dialogue strategy defines the simulated customer's behavioral patterns during the training process, including initial state settings, dialogue progression logic, timing of question posing, and feedback response methods.

[0065] The initial state setting simulates the basic characteristics of the customer based on the training objective, including customer type, intention strength, resistance level, and focus. For example, for Xiao Zhang's need to improve his price negotiation skills, the training scenario generation unit sets the simulated customer as a "high-intention, price-sensitive customer" with an initial intention strength of 0.8, a resistance level of medium resistance, and the main resistance point being price concerns. The focus includes product cost-effectiveness and price comparison with competitors.

[0066] The dialogue progression logic simulates how a customer interacts with a sales associate. The practice session typically begins at a specific Standard Operating Procedure (SOP) point, where the simulated customer asks a typical question and awaits a response from the sales associate. Based on the quality of the sales associate's response, the simulated customer provides corresponding feedback. If the sales associate uses high-weight standard language, the simulated customer gives positive feedback, such as, "Hmm, what you say makes sense." If the sales associate uses low-weight or inappropriate language, the simulated customer gives negative feedback or raises further questions, such as, "But XX brand's price is cheaper than yours." This dynamic feedback mechanism simulates the thinking and reaction patterns of real customers, making the practice session more realistic.

[0067] The timing of questions is controlled by the logic of the conversation. Simulated customers don't ask all their questions at once, but rather gradually as the conversation progresses. This aligns with the behavior of real customers and tests the sales associate's adaptability. In the aforementioned price negotiation scenario, the simulated customer first says, "Your price is a bit expensive." After waiting for the sales associate's initial response, the customer decides whether to raise further questions based on the quality of the response, such as, "Then why is brand XX so much cheaper?" or "Can you lower the price further?"

[0068] Feedback response methods include immediate feedback and delayed feedback. Immediate feedback is given immediately after each response from the sales associate, including emotional feedback (such as "um," "okay," "really?") and semantic feedback (such as asking a new question or expressing understanding). Delayed feedback is given after the training session. The system comprehensively analyzes the sales associate's performance throughout the training process, pointing out the use of correct language, missed key language, and inappropriate language, and provides suggestions for improvement.

[0069] The interactive coaching unit generates dialogue strategies based on the coaching scenario to engage in interactive conversations with the sales associate. The coaching supports both voice and text interaction modes. In voice interaction mode, the sales associate speaks with the system through a microphone, and the system uses speech synthesis technology to generate simulated customer voice responses, making the coaching process more realistic. In text interaction mode, the sales associate inputs responses via a keyboard, and the system displays the simulated customer responses in text format, suitable for use in situations where voice interaction is inconvenient.

[0070] During the coaching process, the interactive coaching unit monitors the accuracy of the sales pitch in real time. The monitoring method is similar to that used in dialogue quality assessment, calculating the similarity between the sales pitch and the standard pitch in the top salesperson model. When the similarity exceeds a threshold, the pitch is considered accurate, and the system provides an immediate "Pitch is accurate" notification. When the similarity is below the threshold, the pitch is considered inaccurate, and the system provides a "Suggested adjustment" notification, along with a detailed explanation of the recommended standard pitch after the dialogue ends. Preferably, the similarity threshold is set to 0.75. This threshold has been experimentally verified to allow salespeople flexibility while ensuring that the core content of the pitch conforms to the standard.

[0071] The interactive coaching unit also features adaptive difficulty adjustment. Based on the sales associate's performance during the coaching process, the difficulty of the simulated customer is dynamically adjusted. This adjustment is reflected in two aspects: the challenging nature of the questions and the leniency of the feedback. When the sales associate performs well, the simulated customer poses more challenging questions and provides stricter feedback, simulating scenarios with difficult customers and honing the sales associate's advanced handling skills. When the sales associate performs poorly, the simulated customer lowers the difficulty of the questions and provides more lenient feedback, preventing the sales associate from losing confidence due to frustration and helping them build skills in a gradual manner.

[0072] In one specific embodiment, Xiao Zhang undergoes training in price negotiation. The training begins with a simulated customer saying, "Your price is a bit expensive; brand XX is 2000 yuan cheaper." Xiao Zhang's initial response is, "Our quality is better than theirs." The system detects this response as too simplistic, with a similarity of only 0.52 to the standard script, below the threshold, and provides a "suggest adjustment" prompt. After the training session, the system displays a detailed recommended script: "I understand your concern about price. While the price figures do differ, our advantages lie in XX core technology and XX service guarantees, which brand XX cannot provide. From a long-term perspective, you'll find our overall cost is actually lower." Xiao Zhang re-trains, this time using a response closer to the standard script: "I understand your concern about price. Although our price is slightly higher, we have XX technology, more reliable quality, and more comprehensive after-sales service." The system detects this as having a similarity of 0.81, exceeding the threshold, and provides an immediate "accurate script" prompt. The simulated customer gives positive feedback: "Hmm, what you say makes sense. What exactly does your after-sales service include?" The dialogue continued, honing Xiao Zhang's ability to respond continuously. After five practice sessions, Xiao Zhang's accuracy rate in price negotiation techniques improved from an initial 48% to 76%, demonstrating significant practical results.

[0073] The parameter optimization unit statistically analyzes the practice data and generates suggested optimization parameters. The practice data includes the various sales scripts used by the sales associates during practice, their accuracy rates, the number of practice rounds, and their progress curves. This data provides valuable feedback for system optimization.

[0074] The optimization recommendations include three aspects. First, the weight reduction coefficient of corresponding sales techniques in the top salesperson model is adjusted based on feedback data from frequently misused salesperson phrases during training sessions. If a certain phrase has a high weight in the top salesperson model, but training sessions reveal that salespersons still struggle to use it correctly even after learning it, or that its effectiveness is poor, it indicates that the phrase may not be suitable for general salesperson promotion, and its weight should be reduced. The weight reduction coefficient is calculated based on a combination of error frequency and training effectiveness. When the error rate of a certain phrase exceeds 50% in training sessions, and there is no significant improvement after more than three rounds of training, the weight of that phrase is reduced by 20% to 30%. This weight adjustment mechanism based on practical feedback makes the top salesperson model more realistic, avoiding a disconnect between the idealized model and practical application.

[0075] Second, the resistance identification threshold in the evaluation criteria is adjusted based on the resistance responses of simulated customers during training sessions. During training, the system records changes in simulated customer resistance when sales associates use different sales techniques. If certain techniques effectively reduce customer resistance but have a low weight in the evaluation criteria, the criteria need adjustment. Threshold adjustments are made incrementally, fine-tuning existing thresholds based on training data. Preferably, threshold optimization is triggered every 100 training data points, with the adjustment range controlled within 5% to avoid large fluctuations that could cause system instability.

[0076] Third, the system predicts the sales associate's skill improvement path and generates personalized training plan parameters based on the progress curve from multiple practice sessions. The system records the sales associate's performance data from multiple practice sessions, including the accuracy rate of their sales pitches, the pass rate, and the speed of progress at each stage. Based on this data, a skill profile is constructed for the sales associate, identifying their learning characteristics. For example, some sales associates learn demand-based sales pitches quickly but are slower at handling objections; the system will generate a personalized plan for such an associate focusing on objection handling training. The progress curve is fitted using exponential smoothing to predict the sales associate's performance trend in future practice sessions. If the prediction shows that the sales associate's progress has plateaued at a certain stage, reaching their skill ceiling, the system will suggest shifting to training at other stages to optimize training efficiency.

[0077] The optimization suggestions generated by the parameter optimization unit are fed back to the sales champion model dynamic extraction and scoring module 2 and the dialogue quality assessment and intention analysis module 3. The sales champion model dynamic extraction and scoring module 2 adjusts the weights of corresponding dialogues in the sales champion model library based on the weight reduction coefficient, making the model more consistent with actual promotional effects. The dialogue quality assessment and intention analysis module 3 updates the judgment criteria for resistance identification and intention identification based on threshold adjustments, improving the accuracy of the assessment. This closed-loop feedback mechanism enables the system to adaptively evolve, allowing it to continuously learn and optimize from practice, avoiding the limitations of static models.

[0078] Reference Figure 6 This invention also provides a method for improving the sales guidance capabilities based on AI voice badges. This method corresponds to the aforementioned system, and the implementation steps are as follows: Step S1: Collect the dialogue voice between the sales guide and the customer through the AI ​​voice badge, identify the speaker's identity using voiceprint recognition technology, convert the speech into text using automatic speech recognition technology, and extract the speech prosody features, emotional features, and semantic features to obtain structured dialogue data containing speaker identification, dialogue text, emotional state, and timestamps. This step corresponds to the function of the multimodal speech acquisition and feature extraction module 1. For detailed implementation, please refer to the aforementioned module description.

[0079] Step S2: Based on preset sales champion determination criteria, select sales champion samples from historical dialogue data, segment the dialogue flow of the sales champion samples into nodes, identify high-frequency phrases and typical customer questions and answers for each node, calculate the frequency of occurrence of phrases and their correlation with successful conversion in the sales champion samples, determine the adaptive weight of phrases based on the frequency of occurrence and the correlation with successful conversion, and construct a sales champion model library containing node phrase templates, a customer question library, and response strategies. This step corresponds to the function of the sales champion model dynamic extraction and scoring module 2.

[0080] In step S2, the adaptive weight of the dialogue is determined using the aforementioned innovative algorithm. The specific calculation formula is as follows: , The formula is defined in detail in claim 10 and will not be repeated here. The core idea of ​​this algorithm is to perform a non-linear mapping by using an exponential function to the product of the frequency and conversion rate of the sales pitch, so that sales pitches that are both high-frequency and high-conversion receive significantly higher weight. Experimental data shows that the sales champion model constructed using this algorithm improves the guidance effect by more than 20% compared with traditional methods.

[0081] Step S3: Based on the standard scripts and procedures in the sales champion model library, perform node matching and scoring on the sales guide reception dialogue. Detect the speaking rate and accuracy of each SOP node in the sales guide dialogue, identify intention signals and resistance signals in the customer dialogue, calculate the customer's intention strength score and resistance level based on the intention signals and resistance signals, and generate a sales guide reception quality assessment report that includes SOP node execution scores, customer intention analysis results, and the identification of sales guide skill shortcomings. This step corresponds to the function of the dialogue quality assessment and intention analysis module 3.

[0082] In step S3, the customer intention strength score is calculated using a multi-dimensional fusion method, integrating three dimensions: the number of proactive inquiries, the number of product feature concerns, and the clarity of decision-making time. The raw values ​​for each dimension are first normalized, then weighted and summed according to preset weights to obtain an intention strength score between 0 and 1. The resistance level is determined using a combination of rules and statistics, classifying resistance as high, medium, or low based on the presence of three resistance signals: price concerns, competitor comparison, and decision hesitation. The evaluation report integrates the analysis results from the three dimensions of SOP execution, customer intention, and customer resistance, and links them to reception recordings, providing comprehensive data support for sales staff review and capability improvement.

[0083] Step S4: Based on the identified weaknesses in the sales consultant's capabilities, determine the SOP nodes and dialogue types that need improvement. Extract the standard dialogues and typical customer Q&As for the corresponding nodes from the top salesperson model library, generate a dialogue strategy simulating the customer role, and conduct interactive practice with the sales consultant based on the dialogue strategy. During the practice, monitor the accuracy of the sales consultant's dialogue in real time and provide immediate feedback. This step corresponds to the first half of the functions in module 4, which is a closed-loop feedback optimization and practice scenario generation module.

[0084] The training scenarios are generated based on the results of weakness analysis, and training scenarios are designed to address the specific weaknesses of sales associates. The dialogue strategy defines the initial state of the simulated customer, the logic of dialogue progression, the timing of questions, and the feedback response methods, making the training process highly simulate real customer service scenarios. Interactive training supports both voice and text interaction modes, and monitors the accuracy of the sales associate's script in real time. When the similarity between the script and the standard script exceeds 0.75, a "script accurate" prompt is given; otherwise, a "suggest adjustment" prompt is given. The training has an adaptive difficulty adjustment function, dynamically adjusting the difficulty of questions and the tolerance for feedback based on the sales associate's performance, achieving gradual skill improvement.

[0085] Step S5: Statistically analyze the training data and generate optimization suggestion parameters. These optimization suggestion parameters are fed back to steps S2 and S3 to dynamically adjust the verbal weights and threshold parameters of the evaluation criteria in the top salesperson model. This step corresponds to the latter half of the closed-loop feedback optimization and training scenario generation module 4, realizing the closed-loop characteristic of the method.

[0086] The optimization recommendations include three categories: a reduction coefficient for script weighting, an adjustment amount for the resistance identification threshold, and personalized training plan parameters. The script weighting reduction coefficient is generated based on feedback data from high-frequency erroneous scripts used by sales associates during training sessions, and is used to adjust the weight of scripts unsuitable for general sales associate promotion in the top salesperson model. The resistance identification threshold adjustment amount is generated based on simulated customer resistance responses during training sessions, and is used to optimize the accuracy of the evaluation criteria. The personalized training plan parameters are generated based on the sales associate's progress curves from multiple training sessions, and are used to guide the focus and sequence of subsequent training.

[0087] After these optimization suggestions are fed back to steps S2 and S3, step S2 updates the weights of the corresponding sales scripts in the sales champion model library based on the weight reduction coefficient, and step S3 updates the judgment criteria for intention recognition and resistance recognition based on the threshold adjustment amount. This closed-loop feedback mechanism enables the method to be adaptive, continuously learning and optimizing from practice, avoiding the limitations of static methods. After multiple rounds of closed-loop iteration, the system's evaluation accuracy improved by 12%, the training effect improved by 18%, and the speed of sales guide capability improvement accelerated by 25%.

[0088] The method of this invention can be implemented based on a dedicated hardware system or on a general-purpose computing device with corresponding software installed. The AI ​​voice badge can utilize commercially available intelligent voice recording devices, possessing recording, voiceprint recognition, and wireless transmission capabilities. Speech recognition, natural language processing, and sentiment analysis modules can be implemented by calling open-source algorithm libraries or commercial API services. Core functions such as building the top salesperson model, assessing dialogue quality, and generating practice scenarios can be deployed on a cloud server, providing services to sales guides through web or mobile applications. Preferably, the system adopts a microservice architecture, with each functional module deployed independently and communicating with each other via API interfaces, facilitating expansion and maintenance.

[0089] In practical applications, the method of this invention can be integrated with existing enterprise CRM and training systems to achieve data sharing and business collaboration. Reception evaluation reports can be automatically synchronized to the CRM system, enriching customer profile information. Training records can be included in sales guide training files, serving as a basis for performance appraisal and promotion evaluation. The top salesperson model library can be periodically exported as training materials for new employees to study and reference.

[0090] The technical solution of this invention achieves a complete closed loop from data acquisition, model extraction, quality assessment to feedback optimization through the deep coupling and closed-loop collaboration of four core modules. The multimodal speech acquisition and feature extraction module provides a high-quality data foundation for the system; the sales champion model dynamic extraction and scoring module enables the automated transfer of experience; the dialogue quality assessment and intention analysis module provides a scientific evaluation system; and the closed-loop feedback optimization and training scenario generation module forms a continuous improvement mechanism. Through deep coupling at the parameter and state levels, each module forms a synergistic effect of mutual promotion and additive enhancement, achieving a non-linear improvement in system performance.

[0091] The innovative algorithms proposed in this invention solve the problems of unreasonable weight allocation of sales pitches and inaccurate customer intent vectorization in traditional methods. The adaptive weight calculation algorithm achieves non-linear weight allocation through exponential function mapping, highlighting effective sales pitches. The multi-dimensional fusion intent vectorization method comprehensively assesses customer purchase intentions across multiple dimensions. These innovative algorithms have been validated through extensive practical application, demonstrating significantly better performance than traditional methods.

[0092] The closed-loop feedback mechanism constructed in this invention enables the system's adaptive evolution. Training data is used to adjust the weights of the top-performing model and the evaluation criterion thresholds, allowing the system to continuously learn and optimize from practice. This continuous evolutionary capability is a key advantage of this invention compared to static systems, ensuring that the system always maintains optimal performance.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for enhancing the sales guide's capabilities based on AI voice ID badges, characterized in that: include: The multimodal voice acquisition and feature extraction module is used to collect the dialogue voice between sales guides and customers through AI voice badges, identify the speaker's identity using voiceprint recognition technology, convert the voice into text using automatic speech recognition technology, and extract voice prosody features, emotional features, and semantic features to obtain structured dialogue data containing speaker identification, dialogue text, emotional state, and timestamps; The sales champion model dynamic extraction and scoring module, connected to the multimodal speech acquisition and feature extraction module, receives the structured dialogue data, filters sales champion samples from historical dialogue data based on preset sales champion judgment criteria, segments the dialogue flow of the sales champion samples into nodes, identifies high-frequency phrases and typical customer questions and answers for each node, calculates the frequency of occurrence of phrases and their correlation with successful conversion in the sales champion samples, determines the adaptive weight of phrases based on the frequency of occurrence and the correlation with successful conversion, constructs a sales champion model library containing node phrase templates, a customer question database, and response strategies, and performs node matching and scoring of sales guide reception dialogues based on standard phrases and standard processes in the sales champion model library. The dialogue quality assessment and intention analysis module, connected to the sales champion model dynamic extraction and scoring module, is used to receive the node matching and scoring results, detect the opening rate and accuracy of each SOP node in the sales guide dialogue, identify intention signals and resistance signals in the customer dialogue, calculate the customer's intention strength score and resistance level based on the intention signals and resistance signals, and generate a sales guide reception quality assessment report based on the intention strength score, resistance level, and the opening rate and accuracy of the opening rate. The assessment report includes SOP node execution scores, customer intention analysis results, and the identification of sales guide capability shortcomings. The closed-loop feedback optimization and training scenario generation module, connected to the dialogue quality assessment and intention analysis module, receives the sales guide reception quality assessment report, identifies the SOP nodes and script types that need strengthening based on the sales guide's capability shortcomings, extracts standard scripts and typical customer Q&As for the corresponding nodes from the top salesperson model library, generates a dialogue strategy simulating a customer role, conducts interactive training with the sales guide based on the dialogue strategy, detects the accuracy of the sales guide's scripts in real time during the training process and provides immediate feedback, statistically analyzes the training data and generates optimization suggestion parameters, which are fed back to the top salesperson model dynamic extraction and scoring module and the dialogue quality assessment and intention analysis module to dynamically adjust the script weights and evaluation standard threshold parameters of the top salesperson model.

2. The AI-based voice ID badge-based system for enhancing shopping guide capabilities according to claim 1, characterized in that, The multimodal speech acquisition and feature extraction module includes: The voiceprint recognition unit is used to extract the acoustic features of the speech signal and match them with the pre-recorded voiceprint databases of sales guides and customers to identify the speaker's identity; A speech-to-text unit, connected to the voiceprint recognition unit, is used to convert speech signals into timestamped text sequences using an automatic speech recognition model; An emotion feature extraction unit, connected to the voiceprint recognition unit, is used to extract prosodic features from the speech signal. The prosodic features include pitch, intensity, speech rate, and pause duration, and the speaker's emotional state is identified based on the prosodic features. The semantic feature extraction unit, connected to the speech-to-text unit, is used to perform natural language processing on the text sequence, extract keywords, entities and semantic relationships, and construct a semantic representation of the dialogue.

3. The AI ​​voice-based sales guide capability enhancement system according to claim 1, characterized in that, The dynamic extraction and scoring module of the top salesperson model includes: The top salesperson sample screening unit is used to select dialogue samples that meet preset success criteria from historical dialogue data as top salesperson samples. The preset success criteria include conversion rate, customer satisfaction score, and reception time. The node segmentation and dialogue extraction unit, connected to the sales champion sample screening unit, is used to segment the dialogue flow of the sales champion sample into nodes and identify high-frequency dialogues and typical customer questions and answers in each node; An adaptive weight calculation unit, connected to the node segmentation and script extraction unit, is used to calculate the frequency of occurrence and the correlation between success rate and success rate of each script, and to determine the adaptive weight of the script based on the frequency of occurrence and the correlation between success rate and success rate. The model library construction unit is connected to the adaptive weight calculation unit and is used to construct the sales champion model library based on the node segmentation results, high-frequency words and adaptive weights.

4. The AI-based voice ID card-based sales guide capability enhancement system according to claim 3, characterized in that, The adaptive weight calculation unit determines the adaptive weight of the script based on the following method: calculating the frequency of the script in the top-performing sample, calculating the success conversion rate of the dialogue sample containing the script, determining the importance score of the script based on the frequency of occurrence and the success conversion rate through an exponential function mapping, and normalizing the importance scores of all scripts to obtain the adaptive weight of each script.

5. The AI-based voice ID badge-based system for enhancing shopping guide capabilities according to claim 1, characterized in that, The dialogue quality assessment and intention analysis module includes: The SOP node detection unit is used to detect the speech rate and accuracy of each SOP node in the sales guide dialogue based on the preset SOP node definition and the node speech templates in the sales champion model library. A customer intent identification unit, connected in parallel with the SOP node detection unit, is used to extract intent signals from customer dialogues. These intent signals include proactive inquiry behavior, product feature concerns, and decision-making time expressions. Based on these intent signals, a customer intent strength score is calculated. A customer resistance identification unit, connected in parallel with the customer intention identification unit, is used to extract resistance signals from customer dialogue. The resistance signals include expressions of price concerns, mentions of competitor comparisons, and expressions of decision hesitation. Based on the resistance signals, the customer's resistance level is determined. The evaluation report generation unit is connected to the SOP node detection unit, the customer intention identification unit, and the customer resistance identification unit, and is used to generate a sales guide reception quality evaluation report by comprehensively considering the opening rate, accuracy rate, intention strength score, and resistance level of the sales script.

6. The AI ​​voice-based sales guide capability enhancement system according to claim 5, characterized in that, The customer intent identification unit uses a multi-dimensional fusion method to calculate the customer's intent strength score: extracting the number of proactive inquiries, the number of product function concerns, and the clarity of decision-making time from the customer's dialogue; normalizing the number of proactive inquiries, the number of product function concerns, and the clarity of decision-making time; and weighting and summing the normalized dimension values ​​based on preset dimension weights to obtain the customer's intent strength score.

7. The AI-based voice ID badge-based system for enhancing shopping guide capabilities according to claim 1, characterized in that, The closed-loop feedback optimization and training scenario generation module includes: The weakness analysis unit is used to extract the weaknesses in the sales guide's capabilities from the sales guide reception quality assessment report, and to determine the SOP nodes and types of conversation scripts that need to be strengthened; The training scenario generation unit, connected to the weakness analysis unit, is used to extract corresponding standard scripts and typical customer Q&As from the top salesperson model library based on the SOP nodes that need to be strengthened, and generate dialogue strategies simulating customer roles; An interactive coaching unit, connected to the coaching scenario generation unit, is used to conduct interactive dialogues with the sales guide according to the dialogue strategy, and to detect the accuracy of the sales guide's script in real time and provide immediate feedback; The parameter optimization unit, connected to the interactive coaching unit, is used to collect coaching data and generate optimized parameters for adjusting the sales champion model's speech weights and evaluation standard thresholds.

8. The AI ​​voice-based sales guide capability enhancement system according to claim 7, characterized in that, The optimization suggestions generated by the parameter optimization unit include: adjusting the weight reduction coefficient of the corresponding sales script in the top sales model based on feedback data of high-frequency erroneous sales scripts in training sessions; adjusting the resistance identification threshold in the evaluation criteria based on the resistance response of the simulated customer role in training sessions; and predicting the salesperson's ability improvement path based on the progress curve of multiple training sessions and generating personalized training plan parameters.

9. A method for improving the sales guidance capabilities based on AI voice employee badges, employing the sales guidance capability improvement system based on AI voice employee badges as described in any one of claims 1-8, characterized in that, Includes the following steps: Step S1: Collect the voice conversation between the sales guide and the customer through the AI ​​voice badge, identify the speaker's identity using voiceprint recognition technology, convert the voice into text using automatic speech recognition technology, and extract the voice prosody features, emotional features, and semantic features to obtain structured dialogue data containing speaker identification, dialogue text, emotional state, and timestamps; Step S2: Based on the preset criteria for determining top sales performers, select top sales performer samples from historical dialogue data, segment the dialogue flow of the top sales performer samples into nodes, identify high-frequency phrases and typical customer questions and answers for each node, calculate the frequency of occurrence of phrases and their correlation with successful conversion in the top sales performer samples, determine the adaptive weight of phrases based on the frequency of occurrence and the correlation with successful conversion, and construct a top sales performer model library that includes node phrase templates, a customer question library, and response strategies; Step S3: Based on the standard scripts and procedures in the sales champion model library, perform node matching and scoring on the sales guide reception dialogue, detect the opening rate and accuracy of each SOP node in the sales guide dialogue, identify the intention signals and resistance signals in the customer dialogue, calculate the customer's intention strength score and resistance level based on the intention signals and resistance signals, and generate a sales guide reception quality assessment report that includes SOP node execution scores, customer intention analysis results, and the identification of sales guide's shortcomings; Step S4: Based on the identified shortcomings in the sales consultant's capabilities, determine the SOP nodes and script types that need to be strengthened. Extract the standard scripts and typical customer Q&As for the corresponding nodes from the top salesperson model library, generate a dialogue strategy simulating the customer role, and conduct interactive practice with the sales consultant based on the dialogue strategy. During the practice, monitor the accuracy of the sales consultant's scripts in real time and provide immediate feedback. Step S5: Statistically analyze the training data and generate optimization suggestion parameters. The optimization suggestion parameters are fed back to steps S2 and S3 to dynamically adjust the speech weights and threshold parameters of the evaluation criteria of the top salesperson model.

10. The method for improving the sales guidance capabilities based on AI voice ID badges according to claim 9, characterized in that, In step S2, the adaptive weights of the speech are determined based on the following formula: ， in, The adaptive weight for the i-th statement. Let be the frequency of the i-th sales pitch in the top-performing salesperson sample. Let be the success conversion rate of the dialogue sample containing the i-th phrase. This is a temperature coefficient used to control the concentration of the weight distribution. This is a scaling factor used to map weight values ​​to a target numerical range. The total number of scripts, The base of the natural logarithm is preferably... The value ranges from 1.5 to 2.

5. The value ranges from 1.0 to 1.

5. The formula achieves non-linear weight allocation through exponential function mapping, which makes high-frequency and high-conversion-rate rhetoric receive significantly higher weight.

Images