An ai psychological interview response quality management method and system

Abstract

The application discloses an AI psychological interview response quality management method and system, and relates to the technical field of artificial intelligence and digital mental health. The method first analyzes the semantic features of the user's voice through a rule path of zero floating point operation to obtain a first quality score; when the score is in a review interval, a lightweight neural network model is started to analyze the voice energy envelope features to obtain a second quality score; and a single-bit quality flag bit is output after the fusion of the two. According to the flag bit and the cumulative number, the system decides whether to generate a tactful reminder technique through a progressive gate, and automatically triggers a zero-blame closed-loop feedback for high-frequency low-quality response users at the end of the conversation, generates a gentle closing statement and a management list. The application realizes low-power-consumption, high-precision real-time judgment of response quality on the edge terminal and low-intrusion intervention, effectively balances the interview data quality and user experience, and forms a complete quality management closed loop.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and digital mental health technology, and particularly relates to an AI psychological interview response quality management method and system. BACKGROUND

[0002] Early identification and intervention of mental health problems are crucial for maintaining individual well-being and social stability. In recent years, artificial intelligence (AI) psychological interview systems based on voice dialogue have gradually become the digital gateway for large-scale mental health screening in schools, communities, enterprises and online platforms, due to their significant advantages in improving screening coverage, reducing professional labor costs, and protecting user privacy. Through standardized interview processes, these systems collect users' voice responses and conduct preliminary analysis and risk assessment based on content and acoustic features, showing good application prospects.

[0003] However, in the actual deployment process, a long-standing and unresolved technical problem has seriously hindered its effectiveness and user experience: how to accurately identify users' "low-quality responses" in real time on resource-constrained edge terminals (such as self-service counseling terminals, embedded devices, and smartphones), and on this basis, conduct gentle intervention without offending or criticizing to guide users to provide more effective information, while ensuring the completion rate of the interview.

[0004] So-called "low-quality responses" are a complex, multi-dimensional behavioral representation with typical characteristics in acoustic and semantic aspects, mainly including:

[0005] Semantic sparsity: the response text is too short, the content word density is low, and contains a large number of meaningless filler words (such as "um", "er", "randomly", "fairly", etc.).

[0006] Acoustic disorder: the fluctuation of the voice energy envelope (i.e. the curve of the energy of the voice signal changing over time) is chaotic, lacking the rhythm and stress rules of normal conversation.

[0007] Currently, the industry's solutions to such problems mainly exist in the following technical routes, and all have obvious limitations:

[0008] Cloud-based large model solution: This type of solution uploads the complete user voice data to a cloud server and uses large deep learning models (such as BERT, Wav2Vec, etc.) for multi-modal (text + audio) sentiment calculation or concentration analysis. Although this approach may achieve high recognition accuracy, its inherent high latency, high network bandwidth consumption, and potential risks to user privacy make it difficult to meet the real-time and data security requirements of psychological interviews. More importantly, the continuous running of floating-point large model inference consumes a huge amount of computing power (usually more than 100 MFLOPS), which is completely impossible to implement economically and feasibly on edge terminals with limited computing power and power consumption.

[0009] Single-modal rule-based solution based on fixed threshold: To meet the real-time requirements of terminal-side, existing lightweight solutions often use simple rule engines, such as setting fixed keyword lists or sentence length thresholds in the text modality, or setting a single energy threshold in the audio modality. This "one-size-fits-all" discrimination completely ignores individual differences in users (for example, naturally introverted users have less baseline speech), which can easily misjudge "less speech" as "low quality", resulting in high false negative rates. In addition, relying solely on a single modality (such as only looking at text keywords) cannot effectively capture "low quality" of the type "not much speech but not focused" (which is characterized by the disorderliness of acoustic energy), resulting in high false negative risk.

[0010] Lack of intelligent and gradient intervention mechanism: Existing technologies lack intelligent and gradient intervention mechanisms when intervening in users who are judged to be "low quality". Usually, pop-up warnings, indicator light flashes, or mechanical voice interrogations (such as "please answer seriously") are used. This lack of empathy and accusatory interaction method can easily cause users to feel resentful and resistant, leading to a sharp increase in user churn rate and a significant decrease in interview completion rate. More importantly, existing technologies lack a cumulative number-based, progressive intelligent reminder strategy, and cannot achieve a gentle intervention closed loop of "understanding-encouragement-guidance".

[0011] Feedback link missing and management closed loop broken: After the conversation ends, existing methods either have no feedback or only harshly prompt "this result is invalid". This not only fails to effectively utilize the large amount of low-quality data generated, but also fails to effectively recall users who did not fully cooperate due to various reasons (such as emotional resistance, environmental interference). The lack of a zero-blame closed loop feedback management mechanism for potential risk users has led to the industry's pain points of low secondary visit rate and high overall screening leakage risk.

[0012] In summary, the prior art has a significant technical gap in implementing a complete low-quality response identification and intervention closed loop with "low power consumption, low misjudgment, low invasion, and high reflux" on the edge terminal. Therefore, there is an urgent need in the field for an innovative technical solution that can achieve accurate, individualized, and low-invasion whole-process quality management on the terminal side without relying on the cloud. SUMMARY

[0013] To solve the technical problems in the background art, the present application proposes an AI psychological interview response quality management method and system.

[0014] The AI psychological interview response quality management method proposed by the present application comprises the following steps:

[0015] S1, obtain the user's input voice data, perform endpoint detection and word segmentation processing on the voice data to obtain the word sequence, frame number and number of fillers of the current sentence, calculate the semantic features of the current sentence and obtain the first quality score based on the rule path of zero floating point operation according to the word sequence, frame number and number of fillers; the zero floating point operation is an operation that completely uses integers;

[0016] S2, when the first quality score is in a preset review interval, a trained lightweight neural network model is started to analyze the energy envelope features of the voice data to obtain a second quality score; the energy envelope features refer to the features of the energy of the voice signal changing with time; the preset review interval is a score range between a preset lower limit and an upper limit threshold;

[0017] S3, fuse the first quality score and the second quality score to obtain a final quality score, and output a single-bit flag bit representing the current response quality level;

[0018] S4, obtain the single-bit flag bit, when the single-bit flag bit indicates a low-quality response, query the number of low-quality response accumulations stored in the memory, if the number of low-quality response accumulations is less than a first threshold, generate and play a tactful reminder script, and update the storage after adding one to the number of low-quality response accumulations; if the number of low-quality response accumulations is greater than or equal to the first threshold, record the current low-quality response event in the background.

[0019] Preferably, before step S1, it further comprises:

[0020] S0, in the session initialization phase, obtain the user's response voice to the preset welcome word, calculate the average number of words and the average substantive word density of the user in the baseline period after endpoint detection and word segmentation processing; the rule path based on zero floating point operation calculates the semantic features of the current sentence and obtains the first quality score according to the word sequence, frame number and number of fillers, specifically including:

[0021] The average number of words and average content word density of users during the baseline period are used as personal baselines;

[0022] Sentence length feature markers are obtained by comparing the current sentence's word count and frame count with the individual baseline;

[0023] The filler word ratio is calculated based on the number of filler words in the current sentence and the total number of words.

[0024] The number of content words is identified and their density is calculated based on the word sequence of the current sentence, and the content word sparsity marker is obtained by comparing it with the personal baseline.

[0025] The first quality score is calculated using a weighted summation formula based on sentence length feature markers, filler word ratio, and content word sparsity markers.

[0026] Preferably, the lightweight neural network model is a one-dimensional convolutional neural network model quantized to INT8 precision. The structure of the one-dimensional convolutional neural network model is as follows: an input layer for receiving a 64-point energy envelope sequence, a first convolutional layer with 8 convolutional kernels of size 8 and a stride of 2, a second convolutional layer with 16 convolutional kernels of size 4 and a stride of 2, a global max pooling layer, and a fully connected layer.

[0027] Preferably, the training process of the lightweight neural network model specifically includes:

[0028] Obtain a training dataset, which contains multiple voice samples from real psychological interview scenarios, each voice sample being labeled with a corresponding response quality label;

[0029] Construct an initial one-dimensional convolutional neural network, whose structure includes an input layer, a first convolutional layer, a second convolutional layer, a global max pooling layer, and a fully connected layer;

[0030] Based on the initial one-dimensional convolutional neural network, a quantization-aware training method is used to train the initial one-dimensional convolutional neural network. By simulating the INT8 quantization process in the forward inference of the training, the network weights are adapted to low-precision calculations.

[0031] The simulated quantized network is iteratively optimized using the training dataset to minimize the loss between the prediction results and the response quality label;

[0032] After training, the network weights are fixed to INT8 precision, and a corresponding Sigmoid lookup table is generated, ultimately yielding the lightweight neural network model.

[0033] Preferably, the Sigmoid lookup table is a read-only memory with a capacity of 256 bytes, and the Sigmoid lookup table has 256 bytes of data pre-programmed inside;

[0034] The byte-type data is configured in the following way: the standard Sigmoid function is uniformly sampled in a predetermined input range, and the function output value is mapped to an integer range of 0 to 255, thereby establishing a non-linear mapping relationship from the 16-bit integer pre-stage output to the 8-bit probability value.

[0035] Preferably, the process of fusing the first quality score and the second quality score to obtain the final quality score is as follows:

[0036] Final quality score = ×First Quality Score+ × (Second quality score / 255);

[0037] in, The weights associated with the first quality score; Weights associated with the second quality score; .

[0038] Preferably, the step of outputting a single-bit flag representing the current response quality level specifically includes:

[0039] The final quality score is compared with a preset final decision threshold;

[0040] If the final quality score is greater than or equal to the final decision threshold, a binary '1' representing a low-quality response is output as a single-bit flag.

[0041] If the final quality score is less than the final decision threshold, a binary '0' representing a high-quality response is output as a single-bit flag.

[0042] Preferably, the step of generating the euphemistic reminder script specifically includes:

[0043] Construct a context vector, which includes sentence length feature markers, filler word ratio, content word sparsity markers, second quality score, first quality score, final quality score, cumulative number of low-quality responses, current session round, and the user's response text from the last three rounds.

[0044] The context vector is combined with a preset prompt template and input into a preset large language model to obtain the text output by the large language model. The text output by the large language model is used as the euphemistic reminder. The prompt template is used to limit the large language model to generate an empathetic response that does not contain negative words and has a limited word count.

[0045] Preferably, it further includes:

[0046] S5. At the end of the session, obtain the cumulative number of low-quality responses. If the cumulative number of low-quality responses is greater than or equal to the second threshold, generate a low-quality response trajectory table that records detailed data of all low-quality responses in this session. Generate and play a gentle closing statement based on the low-quality response trajectory table. At the same time, generate a list containing user identifiers and the cumulative number of low-quality responses and push it to the management terminal.

[0047] This invention proposes an AI-based psychological interview response quality management system, comprising:

[0048] The response detection module is used to acquire user-input voice data, perform endpoint detection and word segmentation on the voice data, obtain the word sequence, frame number and number of filler words of the current sentence, calculate the semantic features of the current sentence based on the rule path of zero floating-point operation, and obtain the first quality score. The zero floating-point operation is an operation that uses only integers.

[0049] The response detection module is also used to start a trained lightweight neural network model to analyze the energy envelope features of the speech data to obtain a second quality score when the first quality score is in a preset review interval. The energy envelope features refer to the features of speech signal energy changing over time. The preset review interval is a score range between a preset lower limit and an upper limit threshold.

[0050] The response detection module is also used to fuse the first quality score and the second quality score to obtain the final quality score, and output a single-bit flag that represents the current response quality level.

[0051] The response reminder module is used to obtain the single-bit flag. When the single-bit flag indicates a low-quality response, it queries the cumulative number of low-quality responses stored in memory. If the cumulative number of low-quality responses is less than a first threshold, it generates and plays a polite reminder message, and increments the cumulative number of low-quality responses by one and updates the storage. If the cumulative number of low-quality responses is greater than or equal to the first threshold, it records the low-quality response event in the background.

[0052] The AI-based psychological interview response quality management method and system proposed in this invention firstly improves detection capabilities by constructing a dual-path cascaded decision architecture of "zero-floating-point rule-driven + INT8 micro-model supplementation," achieving high-precision, low-power real-time analysis of user response quality at the edge terminal. This mechanism effectively distinguishes between "low-quality responses" and "naturally quiet responses" using dynamic rules based on individual baselines, significantly reducing the false positive rate. Furthermore, it leverages a lightweight CNN network to accurately capture the disordered features of acoustic energy envelopes, compensating for the blind spots of pure semantic rules, thereby systematically improving the accuracy and robustness of screening. Secondly, in terms of interactive experience, it innovatively introduces a counter-based progressive gate intervention strategy, transforming harsh, immediate warnings into data-driven, empathetic, and subtle reminders. This ensures the intervention language is "low-intrusive" and "highly acceptable," significantly alleviating user psychological pressure and effectively guaranteeing interview completion rate and data quality. Finally, in terms of system performance, a closed-loop feedback and management mechanism with "zero blame" as its core concept was established. By automatically generating explanatory summaries and lists of potential risk users, a gentle recall and precise second interview were achieved for those who were missed in the screening. This fundamentally solved the drawback of "the process ends when the interview ends" in the existing technology, and formed a closed-loop quality assurance system of "screening-identification-intervention-recall". Ultimately, it achieved a comprehensive performance improvement with low power consumption, low false positives, low intrusion and high backflow. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the workflow of an AI-based psychological interview response quality management method proposed in this invention;

[0054] Figure 2 This is a schematic diagram of the processing flow of one implementation method of the AI ​​psychological interview response quality management method proposed in this invention;

[0055] Figure 3 This is a schematic diagram of a scoring curve for one implementation of an AI-based psychological interview response quality management method proposed in this invention.

[0056] Figure 4 This is a schematic diagram of one implementation of an AI-based psychological interview response quality management system proposed in this invention. Detailed Implementation

[0057] Reference Figures 1-4 The present invention proposes an AI-based psychological interview response quality management method, comprising the following steps:

[0058] S0. During the session initialization phase, the user's response voice to the preset welcome message is obtained. After endpoint detection and word segmentation, the average number of words and average content word density of the user during the baseline period are calculated.

[0059] S1. Obtain the user-input voice data, perform endpoint detection and word segmentation on the voice data, obtain the word sequence, frame number and number of filler words of the current sentence, calculate the semantic features of the current sentence based on the rule path of zero floating-point operation, and obtain the first quality score. Zero floating-point operation is an operation that uses only integers.

[0060] In this embodiment, based on the rule path of zero-floating-point arithmetic, the semantic features of the current sentence are calculated according to the word sequence, frame number, and number of filler words to obtain the first quality score, specifically including:

[0061] The average number of words and average content word density of users during the baseline period are used as personal baselines;

[0062] Sentence length feature markers are obtained by comparing the current sentence's word count and frame count with the individual baseline;

[0063] The filler word ratio is calculated based on the number of filler words in the current sentence and the total number of words.

[0064] The number of content words is identified and their density is calculated based on the word sequence of the current sentence, and a content word sparsity marker is obtained by comparing it with the individual baseline.

[0065] The first quality score is calculated using a weighted summation formula based on sentence length feature markers, filler word ratio, and content word sparsity markers.

[0066] It should be noted that zero-floating-point arithmetic refers to a computational model that, after processing, does not involve any decimal data operations and completes semantic feature extraction and scoring calculations solely through integer calculations. For example, the rule engine mentioned above uses integers to determine short sentences, count the number of low-quality words, and perform integer ratio conversions, thus avoiding the computational power consumption caused by floating-point arithmetic and adapting to lightweight devices for fast processing.

[0067] Specifically, after the system powers on, it continuously picks up audio at 16kHz. After a 640ms sliding window buffer with a 10ms frame forward, the front-end VAD provides a sentence beginning / end interrupt signal. Then it enters the "zero-floating-point" rule engine, with the pipeline as follows:

[0068] First, the system plays a welcome message, such as "Please briefly introduce yourself." Then the user speaks, and the system calculates the individual baseline within 30 seconds. The individual baseline includes the average number of words. Average frame rate Then, for each sentence from the user, the word segmenter outputs the total number of words W and the total number of frames F in the current sentence.

[0069] Next, determine the relative shortness of the sentence using the following formula:

[0070] short_flag=(W≤ ×0.4)&&(F≤120)?1:0;

[0071] The above judgment formula means that a sentence is considered "too short" only when it is 60% shorter than the user's usual sentence and its duration is less than 1.2 seconds. The short_flag value represents the short sentence judgment value. A short_flag of 1 indicates that the sentence is too short, and a short_flag of 0 indicates that the sentence is not too short, in order to prevent misjudgment based on the user's naturally short speech.

[0072] 12 high-frequency low-quality words are pre-stored {um, uh, just, okay, I don't know, whatever, so-so, nothing, see, all right, it doesn't matter, just like that}. The word segmentation is performed by looking up the table at the same time. Each time a high-frequency low-quality word is hit, filter_cnt is increased by 1, where filter_cnt is the number of low-quality words in the current sentence.

[0073] Specifically, the formula for calculating the proportion of filler words is as follows:

[0074] filler_ratio=filler_cnt / W;

[0075] Here, filler_cnt represents the number of low-quality words in the current sentence; W represents the total number of words in the current sentence; and filler_ratio represents the content of low-quality words in the current sentence, i.e., the filler ratio.

[0076] In this embodiment, the number of content words is identified and their density is calculated based on the word sequence of the current sentence. A sparsity marker for content words is obtained by comparing this density with a personal baseline. Specifically, this includes:

[0077] First, each word in the local 200kB static dictionary is marked with a "meaning" flag (noun / verb / adjective); content_cnt represents the number of meaning words hit;

[0078] density0 = User's usual density = Content word density in user replies within 30 seconds of the system's welcome message = content_cnt0 / W0;

[0079] density = density of the current sentence = density of content words in the current sentence = content_cnt / W;

[0080] The process of determining relative sparsity is as follows:

[0081] rel_density_flag=(density×100≤density0×60)?1:0;

[0082] Among them, rel_density_flag is the density judgment value of the current sentence, density0 is the average content word density, and density is the content word density of the current sentence. It is considered "space" only when the content word density of the current sentence is 40% lower than the density usually encountered by the user.

[0083] In this embodiment, the formula for calculating the first quality score is as follows:

[0084] S=(20×short_flag+45×filler_ratio+35×rel_density_flag) / 100;

[0085] Where S is the first quality score; short_flag is the short sentence judgment value; filler_ratio is the filler word ratio; and rel_density_flag is the current sentence density judgment value.

[0086] S2. When the first quality score is within the preset review interval, a trained lightweight neural network model is started to analyze the energy envelope features of the speech data to obtain the second quality score. The energy envelope features refer to the characteristics of the speech signal energy changing over time. The preset review interval is the score range between the preset lower limit and the upper limit threshold.

[0087] Specifically, the first quality score output is between 0 and 1, and the threshold for the first quality score is θ_rule=0.70; a first quality score ≥0.70 is judged as "low quality".

[0088] It should be noted that the energy envelope feature refers to the characteristic of the energy of a speech signal changing over time. That is, the speech is divided into fixed time frames (such as 10ms), the energy value of each frame is calculated (the greater the energy, the louder the speech in that frame), and these frame energies are arranged in chronological order to form a continuous energy change sequence (such as a 64-point sequence). This can reflect the strength fluctuations of the speech and is a key acoustic feature for judging whether the speech is clear and the expression is coherent.

[0089] In this embodiment, the lightweight neural network model is a one-dimensional convolutional neural network model quantized to INT8 precision. The structure of the one-dimensional convolutional neural network model is as follows: an input layer for receiving a 64-point energy envelope sequence, a first convolutional layer with 8 convolutional kernels of size 8 and a stride of 2, a second convolutional layer with 16 convolutional kernels of size 4 and a stride of 2, a global max pooling layer, and a fully connected layer.

[0090] Specifically, the 64-point energy envelope sequence is as follows:

[0091] ;

[0092] Where E is the 64-point energy envelope obtained by shifting the 640ms speech by 10ms, with a single byte value of 0-255. The larger the value, the louder the frame. The envelope point number is [0, 63], with a total of 64 points.

[0093] The first convolutional layer is specifically as follows:

[0094]

[0095] Where, in the formula The height index of the output feature map; This represents the local offset of the convolution kernel on the input channel; Enter the channel number; The energy envelope of 64 points is located at the window position. The input sample value; and These are the first convolutional layer and their respective... Input channel, first The weights of the kernel position and the bias shared by the output channel are continuously updated during the training phase via gradient descent. The linear convolution result is then hard-saturated and cropped to the [0,6] interval to form the final output feature; For the first convolutional layer at the 1st The output channel, the first The activation output is obtained from each spatial location.

[0096] The second convolutional layer is specifically as follows:

[0097]

[0098] Where, in the formula The height index of the output feature map; For the convolution kernel in Local offset in the height dimension; Number the output channel; It is the output feature map of the first convolutional layer at all 8 input channels and height positions. The 8-dimensional column vector at that location (i.e., the 8-dimensional column vector at that location) (The values ​​of the 8 channels at the same spatial location are retrieved at once). and These are the second convolutional layer, number 1 and 2 respectively. Output channels, kernel offset The weights at the second convolutional layer The shared bias of the output channels is continuously updated during the training phase through gradient descent. The linear convolution result is then hard-saturated and cropped to the [0,6] interval to form the final output feature; For the second convolutional layer in the 1st... The output channel, the first The activation output is obtained from each spatial location.

[0099] The global max pooling layer is as follows:

[0100] ;

[0101] Where, in the formula The height index of the output feature map of the second convolutional layer; This is the output channel number for the second convolutional layer. For the second convolutional layer in the 1st... Channel, First Activation output at height position; The output vector is 16-dimensional, obtained by... Perform global max pooling on the height dimension, and obtain the maximum value at all height positions of the channel.

[0102] The fully connected layer is specifically:

[0103]

[0104] Where, in the formula Number the output channels of the second convolutional layer; and These are the fully connected layers. Output channel weights and shared bias; This is the output vector of the global max pooling layer; This is the output value of the fully connected layer; This means shifting the Z logic right by 8 bits to obtain an index value between 0 and 255; This represents a lookup table (constant array) of length 256. The final lookup table output value is based on... The high 8 bits are immediately read as an 8-bit precision approximation.

[0105] In this embodiment, the training process of the lightweight neural network model specifically includes:

[0106] Obtain the training dataset, which contains multiple voice samples from real psychological interview scenarios, each voice sample is labeled with a corresponding response quality label;

[0107] Construct an initial one-dimensional convolutional neural network, whose structure includes an input layer, a first convolutional layer, a second convolutional layer, a global max pooling layer, and a fully connected layer;

[0108] Based on the initial one-dimensional convolutional neural network, a quantization-aware training method is used to train the initial one-dimensional convolutional neural network. By simulating the INT8 quantization process in the forward inference of the training, the network weights are adapted to low-precision calculations.

[0109] The simulated quantized network was iteratively optimized using the training dataset to minimize the loss between the prediction results and the response quality label.

[0110] After training, the network weights are fixed to INT8 precision, and a corresponding Sigmoid lookup table is generated, ultimately resulting in a lightweight neural network model.

[0111] In this embodiment, the Sigmoid lookup table is a read-only memory with a capacity of 256 bytes, and 256 bytes of data are pre-programmed into the Sigmoid lookup table.

[0112] Byte-type data is configured by uniformly sampling the standard Sigmoid function over a predetermined input range and mapping the function output value to an integer range of 0 to 255, thereby establishing a non-linear mapping relationship from the 16-bit integer pre-stage output to the 8-bit probability value.

[0113] Specifically, considering that "energy levels fluctuate wildly and irregularly" is often accompanied by casual responses, for example:

[0114] A serious answer: The volume is steady, with emphasis and pauses, like a regular heartbeat on an electrocardiogram.

[0115] Low-quality response: Saying whatever comes to mind, sometimes leaning in to shout, sometimes muttering quietly, the energy curve jumps around randomly.

[0116] Therefore, the system determines a "rule edge region" if and only if the first quality score S ∈ the preset review interval [0.4, 0.7] and short_flag = 0. It then immediately activates a lightweight neural network model to quantify the typical low-quality feature of "disordered energy envelope fluctuations," achieving the goal of "if the rules are unclear, the machine will listen again." Ultimately, it still needs to be calculated along with sentence length and filler words, and conclusions won't be drawn solely from fluctuations.

[0117] S3. Combine the first quality score and the second quality score to obtain the final quality score, and output a single-bit flag that represents the current response quality level.

[0118] In this embodiment, the final quality score is obtained by fusing the first quality score and the second quality score, specifically as follows:

[0119] Final quality score = ×First Quality Score+ × (Second quality score / 255);

[0120] in, The weights associated with the first quality score; Weights associated with the second quality score; .

[0121] Specifically, The value is 0.6. If the value is 0.4, the final quality score = 0.6 × first quality score + 0.4 × (second quality score / 255); if the final quality score is ≥ 0.55, it is judged as "low quality" and output level 1; otherwise, level 0.

[0122] The lightweight neural network model in this embodiment fills in the disordered features of the energy envelope in the "rule edge region", complementing the hard rules. This avoids falsely identifying users with low talk volume and accurately identifies low-quality answers that are "talkative but distracted", providing a highly feasible guarantee of focus for self-help psychological counseling.

[0123] In this embodiment, a single-bit flag representing the current response quality level is output, specifically including:

[0124] The final quality score is compared with a preset final decision threshold;

[0125] If the final quality score is greater than or equal to the final decision threshold, a binary '1' representing a low-quality response is output as a single-bit flag.

[0126] If the final quality score is less than the final decision threshold, a binary '0' representing a high-quality response is output as a single-bit flag.

[0127] S4. Obtain the single-bit flag. When the single-bit flag indicates a low-quality response, query the cumulative number of low-quality responses stored in memory. If the cumulative number of low-quality responses is less than the first threshold, generate and play a polite reminder message, and increment the cumulative number of low-quality responses by one and update the storage. If the cumulative number of low-quality responses is greater than or equal to the first threshold, record this low-quality response event in the background.

[0128] In this embodiment, the steps for generating a polite reminder message specifically include:

[0129] Construct a context vector, which includes sentence length feature markers, filler word ratio, content word sparsity markers, second quality score, first quality score, final quality score, cumulative number of low-quality responses, current session round, and the user's response text from the last three rounds.

[0130] The context vector is combined with a preset prompt template and input into a preset large language model to obtain the text output by the large language model. The text output by the large language model is used as a polite reminder. The prompt template is used to limit the large language model to generate an empathetic response that does not contain negative words and has a word limit.

[0131] In this embodiment, it also includes:

[0132] S5. At the end of the session, obtain the cumulative number of low-quality responses. If the cumulative number of low-quality responses is greater than or equal to the second threshold, generate a low-quality response trajectory table that records detailed data of all low-quality responses in this session. Generate and play a gentle closing statement based on the low-quality response trajectory table. At the same time, generate a list containing user identifiers and the cumulative number of low-quality responses and push it to the management terminal.

[0133] Example 1:

[0134] Taking the example of "a user completing a 15-round dialogue in front of an AI psychological interview terminal and being identified by the system as having 4 low-quality responses", this application fully demonstrates how it completes the entire process of "low power consumption detection → low intrusion reminder → zero-blame closed-loop feedback".

[0135] Step 1: System Initialization

[0136] Upon powering on the S1.1 terminal, a 200kB semantic dictionary, 12 low-quality word hash tables, 14kB CNN weights, and a 256B Sigmoid lookup table are loaded.

[0137] The S1.2 voice front-end is set to 16kHz sampling, 10ms frame shift, and 640ms sliding window; the number of reminders counter_f is initialized to 0; the gear cache flag is initialized to 0.

[0138] Step 2: Individual Baseline Acquisition (Round 0):

[0139] The S2.1 terminal plays a welcome message: "Please briefly introduce yourself."

[0140] S2.230s sound pickup, word segmenter outputs total word count. Total number of frames 1. Count the number of content words (content_cnt0), and calculate density0 = content_cnt0 / W0;

[0141] S2.3 Personal baselines are written to SRAM for subsequent rule calls.

[0142] Step 3: Real-time detection round by round (rounds 1 through 15):

[0143] Execute steps 3.1-3.4 for each round of user voice input:

[0144] S3.1 Voice input → VAD sentence segmentation → word segmentation → obtain W, F, filler_cnt, content_cnt;

[0145] S3.2 Rule Path (Zero Floating Point):

[0146] short_flag=(W≤W0×0.4)&&(F≤120)?1:0;

[0147] filler_ratio=filler_cnt / W;

[0148] rel_density_flag=(content_cnt / W≤density0×0.6)?1:0;

[0149] S=(20×short_flag+45×filler_ratio+35×rel_density_flag) / 100;

[0150] S3.3 If S≥0.70 → directly determine "gear 1" and jump to S3.5;

[0151] If S∈[0.40,0.70] and short_flag=0, start the lightweight neural network model:

[0152] a) 640ms energy envelope with 64 points → INT8 quantization → CNN forward pass → obtain the second quality score ∈ [0,255];

[0153] b) Final quality score = 0.6 × first quality score + 0.4 × (second quality score / 255);

[0154] c) If the final quality score is ≥0.55, it is judged as "gear 1"; otherwise, it is judged as "gear 0"; otherwise, it is judged as "gear 0".

[0155] S3.4 Gear 0: This round is over, waiting for the next sentence;

[0156] S3.5 gear 1: Enter the intervention process (step 4).

[0157] Step 4: Progressive Intervention

[0158] S4.1 Check the number of reminders counter_f:

[0159] If counter_f < 4, pack an 8-dimensional context vector. The 8-dimensional context vector includes:

[0160] a) (Short sentence judgment value, low quality word density, content word density judgment value, second quality score, first quality score, final quality score, number of reminders, which round of dialogue this is in the entire psychological interview process, and the original text of the previous 3 rounds of dialogue).

[0161] b) Feed the 8-dimensional vector into the "reminder coach" prompts used by the large model, so that the large model returns a conversational response of ≤28 characters;

[0162] In this embodiment, the low-quality responses, i.e., the corresponding prompts and replies, are shown in Table 1:

[0163] Table 1. Comparison Table of Low-Quality Response Alerts:

[0164]

[0165] c) The terminal plays the response; the number of reminders is incremented by 1, i.e., counter_f++;

[0166] d) Backend database entry: Record the current round of trajectory (round number, text, various indicators, final quality score).

[0167] If counter_f≥4:

[0168] a) Backend database entry only: Records the current round's trajectory (round number, text, various metrics, final quality score);

[0169] b) Do not play any audio messages to avoid excessive reminders.

[0170] Step 5: Session ends, triggering feedback (end of round 15):

[0171] S5.1 When counter_f=4, the system detects that the "gentle ending" condition of counter_f≥4 is met, and stores the user information in the "low-quality response" user table for the interview administrator to conduct secondary screening interviews. The statistics of the number of low-quality responses are shown in Table 2:

[0172] Table 2. Statistics on the Number of Low-Quality Responses:

[0173]

[0174] S5.2 generates a "Low-Quality Trajectory" table and a line snapshot in the background, and writes it to the log along with the counter_f value. The log in this example is shown in Table 3:

[0175] Table 3. Low-quality response log:

[0176]

[0177] like Figure 3As shown in the figure, the black curve represents the first quality score. The coordinates of the four black dots marked on the curve are (1, 1), (9, 0.45), (10, 1), and (13, 0.45). The first quality score for rounds 1 and 10 is 1, which is immediately judged as "low quality response". The first quality score for rounds 9 and 13 is 0.45, which is between 0.4 and 0.7, and short_flag=0, which is within the fuzzy limit of "low quality response". Therefore, the final quality score is calculated. The red dot (9, 0.62) in the figure represents the final quality score of 0.62 for round 9, and (13, 0.62) represents the final quality score of 0.62 for round 13. Both are greater than 0.55, so they are judged as "low quality response".

[0178] In the feedback phase, S5.3 feeds the "low-quality trajectory" table from S5.2 into the feedback prompts used by the language model, so that the model outputs a mild closing statement of ≤80 characters.

[0179] S5.4 terminal playback example declaration:

[0180] "You gave rather brief answers four times just now, perhaps because you weren't prepared or because of internet issues. Therefore, today's assessment may not have fully reflected your true state. Thank you for coming, and I look forward to our next meeting so we can complete the rest of the puzzle together. I hope you have a more relaxed evening than when you arrived."

[0181] Step 6: Zero-intrusion silent recording (if counter_f < 4)

[0182] When the S6.1 session ends and counter_f < 4, the system only writes the trajectory table to the background without playing any additional audio, achieving zero-intrusion silent recording.

[0183] Step 7: Reset and wait for the next consultation.

[0184] S7.1: counter_f←0 (the number of reminders has been reduced to zero), clear the gear cache, return to step 2 and wait for the next round of inquiries from new users or the same user.

[0185] This concludes the demonstration of the operation process, fully covering the collaborative workflow of "detection → reminder → feedback".

[0186] Reference Figures 1-4 The present invention proposes an AI-based psychological interview response quality management system, comprising:

[0187] The response detection module is used to acquire user-input voice data, perform endpoint detection and word segmentation on the voice data, obtain the word sequence, frame number and number of filler words of the current sentence, calculate the semantic features of the current sentence based on the rule path of zero floating-point operation, and obtain the first quality score. Zero floating-point operation is an operation that uses only integers.

[0188] The response detection module is also used to start a trained lightweight neural network model to analyze the energy envelope features of the speech data to obtain a second quality score when the first quality score is in the preset review interval. The energy envelope features refer to the characteristics of the speech signal energy changing over time. The preset review interval is the score range between the preset lower limit and the upper limit threshold.

[0189] The response detection module is also used to fuse the first quality score and the second quality score to obtain the final quality score, and output a single-bit flag that represents the current response quality level.

[0190] The response reminder module is used to obtain a single-bit flag. When the single-bit flag indicates a low-quality response, it queries the cumulative number of low-quality responses stored in memory. If the cumulative number of low-quality responses is less than a first threshold, it generates and plays a polite reminder message, and increments the cumulative number of low-quality responses by one and updates the storage. If the cumulative number of low-quality responses is greater than or equal to the first threshold, it records the low-quality response event in the background.

[0191] The closed-loop feedback module is used to obtain the cumulative number of low-quality responses at the end of the session. If the cumulative number of low-quality responses is greater than or equal to the second threshold, a low-quality response trajectory table is generated, which records detailed data of all low-quality responses in this session. A gentle closing statement is generated and played based on the low-quality response trajectory table. At the same time, a list containing user identifiers and the cumulative number of low-quality responses is generated and pushed to the management terminal.

[0192] Specifically, the response detection module is responsible for identifying low-quality responses in real time at the edge. It uses a dual approach of "rule-driven + micro-model supplementation" to quantify and fuse acoustic (energy envelope fluctuations) and semantic (sentence length, filler word ratio, content word density) to achieve zero-floating-point speech "focus" judgment, providing a single-bit credible flag for downstream intervention strategies.

[0193] In this embodiment, the activation conditions for the response reminder module are as follows:

[0194] ① First check the first quality score and the final quality score to confirm that this round is indeed "gear 1";

[0195] ② Check counter_f (number of reminders): If counter_f < 4, use the reminder agent to generate reminder messages. If counter_f ≥ 4, only record the user's low-quality process in the background and do not use the reminder agent to remind the user, so as to avoid excessive reminders that may cause user resentment and hinder the psychological counseling process.

[0196] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for quality management of AI-based psychological interview responses, characterized in that, Includes the following steps: S1. Obtain the voice data input by the user, perform endpoint detection and word segmentation on the voice data to obtain the word sequence, frame number and number of filler words of the current sentence, calculate the semantic features of the current sentence based on the rule path of zero floating-point operation, and obtain the first quality score. The zero floating-point operation is an operation that uses integers completely. S2. When the first quality score is within the preset review interval, a trained lightweight neural network model is started to analyze the energy envelope features of the speech data to obtain the second quality score. The energy envelope features refer to the features of speech signal energy changing over time. The preset review interval is the score range between the preset lower limit and the upper limit threshold. S3. Combine the first quality score and the second quality score to obtain the final quality score, and output a single-bit flag representing the current response quality level; S4. Obtain the single-bit flag. When the single-bit flag indicates a low-quality response, query the cumulative number of low-quality responses stored in memory. If the cumulative number of low-quality responses is less than the first threshold, generate and play a polite reminder message, and increment the cumulative number of low-quality responses by one and update the storage. If the cumulative number of low-quality responses is greater than or equal to the first threshold, then the low-quality response event will be recorded in the background.

2. The AI-based psychological interview response quality management method according to claim 1, characterized in that, Before step S1, the following is also included: S0. During the session initialization phase, the user's response voice to the preset welcome message is acquired. After endpoint detection and word segmentation, the average number of words and average content word density of the user during the baseline period are calculated. The rule path based on zero-floating-point arithmetic calculates the semantic features of the current sentence and obtains the first quality score based on the word sequence, frame number, and number of filler words. Specifically, it includes: The average number of words and average content word density of users during the baseline period are used as personal baselines; Sentence length feature markers are obtained by comparing the current sentence's word count and frame count with the individual baseline; The filler word ratio is calculated based on the number of filler words in the current sentence and the total number of words. The number of content words is identified and their density is calculated based on the word sequence of the current sentence, and the content word sparsity marker is obtained by comparing it with the personal baseline. The first quality score is calculated using a weighted summation formula based on sentence length feature markers, filler word ratio, and content word sparsity markers.

3. The AI-based psychological interview response quality management method according to claim 1, characterized in that, The lightweight neural network model is a one-dimensional convolutional neural network model quantized to INT8 precision. The structure of the one-dimensional convolutional neural network model is as follows: an input layer for receiving a 64-point energy envelope sequence, a first convolutional layer with 8 convolutional kernels of size 8 and a stride of 2, a second convolutional layer with 16 convolutional kernels of size 4 and a stride of 2, a global max pooling layer, and a fully connected layer.

4. The AI-based psychological interview response quality management method according to claim 1, characterized in that, The training process of the lightweight neural network model specifically includes: Obtain a training dataset, which contains multiple voice samples from real psychological interview scenarios, each voice sample being labeled with a corresponding response quality label; Construct an initial one-dimensional convolutional neural network, whose structure includes an input layer, a first convolutional layer, a second convolutional layer, a global max pooling layer, and a fully connected layer; Based on the initial one-dimensional convolutional neural network, a quantization-aware training method is used to train the initial one-dimensional convolutional neural network. By simulating the INT8 quantization process in the forward inference of the training, the network weights are adapted to low-precision calculations. The simulated quantized network is iteratively optimized using the training dataset to minimize the loss between the prediction results and the response quality label; After training, the network weights are fixed to INT8 precision, and a corresponding Sigmoid lookup table is generated, ultimately yielding the lightweight neural network model.

5. The AI-based psychological interview response quality management method according to claim 4, characterized in that, The Sigmoid lookup table is a read-only memory with a capacity of 256 bytes, and 256 bytes of data are pre-programmed into the Sigmoid lookup table. The byte-type data is configured in the following way: the standard Sigmoid function is uniformly sampled in a predetermined input range, and the function output value is mapped to an integer range of 0 to 255, thereby establishing a non-linear mapping relationship from the 16-bit integer pre-stage output to the 8-bit probability value.

6. The AI-based psychological interview response quality management method according to claim 1, characterized in that, The final quality score is obtained by fusing the first quality score and the second quality score, specifically as follows: Final quality score = ×First Quality Score+ × (Second quality score / 255); in, The weights associated with the first quality score; Weights associated with the second quality score; .

7. The AI-based psychological interview response quality management method according to claim 6, characterized in that, The step of outputting a single-bit flag representing the current response quality level specifically includes: The final quality score is compared with a preset final decision threshold; If the final quality score is greater than or equal to the final decision threshold, a binary '1' representing a low-quality response is output as a single-bit flag. If the final quality score is less than the final decision threshold, a binary '0' representing a high-quality response is output as a single-bit flag.

8. The AI-based psychological interview response quality management method according to claim 2, characterized in that, The steps for generating the polite reminder message specifically include: Construct a context vector, which includes sentence length feature markers, filler word ratio, content word sparsity markers, second quality score, first quality score, final quality score, cumulative number of low-quality responses, current session round, and the user's response text from the last three rounds. The context vector is combined with a preset prompt template and input into a preset large language model to obtain the text output by the large language model. The text output by the large language model is used as the euphemistic reminder. The prompt template is used to limit the large language model to generate an empathetic response that does not contain negative words and has a limited word count.

9. The AI-based psychological interview response quality management method according to claim 1, characterized in that, Also includes: S5. At the end of the session, obtain the cumulative number of low-quality responses. If the cumulative number of low-quality responses is greater than or equal to the second threshold, generate a low-quality response trajectory table that records detailed data of all low-quality responses in this session. Generate and play a gentle closing statement based on the low-quality response trajectory table. At the same time, generate a list containing user identifiers and the cumulative number of low-quality responses and push it to the management terminal.

10. An AI-powered psychological interview response quality management system, characterized in that, include: The response detection module is used to acquire user-input voice data, perform endpoint detection and word segmentation on the voice data, obtain the word sequence, frame number and number of filler words of the current sentence, calculate the semantic features of the current sentence based on the rule path of zero floating-point operation, and obtain the first quality score. The zero floating-point operation is an operation that uses only integers. The response detection module is also used to start a trained lightweight neural network model to analyze the energy envelope features of the speech data to obtain a second quality score when the first quality score is in a preset review interval. The energy envelope features refer to the features of speech signal energy changing over time. The preset review interval is a score range between a preset lower limit and an upper limit threshold. The response detection module is also used to fuse the first quality score and the second quality score to obtain the final quality score, and output a single-bit flag that represents the current response quality level. The response reminder module is used to obtain the single-bit flag. When the single-bit flag indicates a low-quality response, it queries the cumulative number of low-quality responses stored in memory. If the cumulative number of low-quality responses is less than a first threshold, it generates and plays a polite reminder message, and increments the cumulative number of low-quality responses by one and updates the storage. If the cumulative number of low-quality responses is greater than or equal to the first threshold, then the low-quality response event will be recorded in the background.

Images