AI psychological interview response quality management method and system

Abstract

The invention discloses an AI psychological interview response quality management method and system, and relates to the technical field of artificial intelligence and digital psychological health. The method comprises the following steps: firstly, analyzing semantic features of user voice through a rule path of zero floating point operation to obtain a first quality score; when the score is in the section needing to be rechecked, starting a lightweight neural network model to analyze the voice energy envelope features to obtain a second quality score; and after fusing the two, outputting a single-bit quality flag bit. And the system decides whether to generate the gentle reminding verbal skill or not through a progressive gate according to the flag bit and the accumulated number of times, and automatically triggers zero-responsibility closed-loop feedback for the high-frequency low-quality response user when the session is ended, so as to generate a gentle closing declaration and management list. According to the invention, low-power-consumption and high-precision response quality real-time discrimination and low-invasion intervention on the edge terminal are realized, interview data quality and user experience are effectively considered, and a complete quality management closed loop is formed.

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

[0005] Acoustic disorder: the fluctuation of the voice energy envelope (i.e. the curve of the energy of the voice signal changing with time) is chaotic, lacking the rhythm and stress rules of normal conversation.

[0006] Currently, the industry's solutions to such problems mainly exist in the following technical routes, and all have obvious limitations: 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.

[0007] 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.

[0008] 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".

[0009] 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.

[0010] 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 precise, individualized, and low-invasion whole-process quality management on the terminal side without relying on the cloud. SUMMARY

[0011] To solve the technical problems in the background art, the present application proposes an AI psychological interview response quality management method and system.

[0012] The AI psychological interview response quality management method proposed by the present application comprises the following steps: S1, obtaining the user's input voice data, performing 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, calculating the semantic features of the current sentence and obtaining 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 being an operation using only integers; 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 referring to the features of the energy of the voice signal changing with time, and the preset review interval being a score range between a preset lower limit and an upper limit threshold; S3, fusing the first quality score and the second quality score to obtain a final quality score, and outputting a single-bit flag representing the current response quality level; S4, obtaining the single-bit flag, when the single-bit flag indicates a low-quality response, querying 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, generating and playing a tactful reminder script, and updating 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, recording the current low-quality response event in the background.

[0013] Preferably, before step S1, it further comprises: S0, in the session initialization phase, obtaining the user's response voice to the preset welcome word, calculating the average number of words and the average density of real words 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: obtaining the average number of words and the average density of real words of the user in the baseline period as the personal baseline; comparing the word number and frame number of the current sentence with the personal baseline to obtain a sentence length feature flag; According to the number of fillers in the current sentence and the total number of words, the filler proportion is calculated; According to the word sequence of the current sentence, the number of real substantive words is identified and the density is calculated, and the real substantive word sparsity mark is obtained by comparing with the personal baseline; Based on the sentence length feature mark, the filler proportion and the real substantive word sparsity mark, the first quality score is calculated by a weighted sum formula.

[0014] Preferably, the lightweight neural network model is a one-dimensional convolutional neural network model quantized to INT8 precision, and the structure of the one-dimensional convolutional neural network model is in turn: an input layer for receiving a 64-point energy envelope sequence, a first convolutional layer with 8 convolutional kernels of size 8 and a step of 2, a second convolutional layer with 16 convolutional kernels of size 4 and a step of 2, a global maximum pooling layer, and a fully connected layer.

[0015] Preferably, the training process of the lightweight neural network model specifically includes: Obtain a training data set, which contains multiple speech samples from real psychological interview scenes, and each speech sample is labeled with a corresponding response quality label; Construct an initial one-dimensional convolutional neural network, which includes an input layer, a first convolutional layer, a second convolutional layer, a global maximum pooling layer, and a fully connected layer; Based on the initial one-dimensional convolutional neural network, use a quantization-aware training method to train the initial one-dimensional convolutional neural network, simulate the INT8 quantization process in the forward inference before training, and adapt the network weights to low-precision calculation; Use the training data set to iteratively optimize the simulated quantized network to minimize the loss between the predicted results and the response quality labels; After training, the network weights are fixed to INT8 precision, and the corresponding Sigmoid lookup table is generated, and finally the lightweight neural network model is obtained.

[0016] Preferably, the Sigmoid lookup table is a read-only memory with a capacity of 256 bytes, and the Sigmoid lookup table has 256 byte-type data pre-burned inside; The byte-type data is configured by: uniformly sampling the standard Sigmoid function in a predetermined input interval, and mapping the function output value to the integer range of 0 to 255, thereby establishing a nonlinear mapping relationship from the 16-bit integer pre-stage output to the 8-bit probability value.

[0017] Preferably, the final quality score is obtained by fusing the first quality score and the second quality score, specifically: Final quality score = w1*first quality score + w2*second quality score w1+w2=1 X (second quality score / 255) ; wherein, is a weight related to the first quality score; is a weight related to the second quality score; .

[0018] Preferably, the single-bit flag representing the current response quality level is output, specifically comprising: comparing the final quality score with a preset final decision threshold; if the final quality score is greater than or equal to the final decision threshold, outputting binary '1' representing a low-quality response as the single-bit flag; if the final quality score is less than the final decision threshold, outputting binary '0' representing a high-quality response as the single-bit flag.

[0019] Preferably, the step of generating the polite reminder phrase specifically comprises: constructing a context vector containing a sentence length feature flag, a filler word proportion, a content word sparsity flag, a second quality score, a first quality score, a final quality score, a cumulative number of low-quality responses, a current session round, and a user's last three rounds of response texts; combining the context vector with a preset prompt template, inputting into a preset large language model to obtain a text output by the large language model, and taking the text output by the large language model as the polite reminder phrase, the prompt template being used to limit the large language model to generate an empathetic response without negative words and with limited word count.

[0020] Preferably, it further comprises: S5, obtaining 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 a second threshold, generating a low-quality response trajectory table recording all the detailed data of low-quality responses in the current session, generating and playing a gentle closing statement based on the low-quality response trajectory table, and generating a list containing the user identifier and the cumulative number of low-quality responses and pushing it to the management end.

[0021] The AI psychological interview response quality management system provided by the application comprises: a response detection module for obtaining voice data input by a user, performing endpoint detection and word segmentation processing on the voice data to obtain a word sequence, a frame number and a number of filler words of a current sentence, calculating semantic features of the current sentence and obtaining a first quality score based on a rule path of zero floating point operation, the zero floating point operation being an operation using only integers; The response detection module is further used to start a trained lightweight neural network model to analyze an energy envelope feature of the voice data to obtain a second quality score when the first quality score is in a preset review interval, the energy envelope feature refers to a feature that the energy of a voice signal changes over time, and the preset review interval is a score range between preset lower and upper threshold values. The response detection module is further used to fuse the first quality score and the second quality score to obtain a final quality score, and output a single-bit flag bit representing a current response quality level. The response reminding module is used to obtain the single-bit flag bit, query a low-quality response cumulative number stored in the memory when the single-bit flag bit indicates a low-quality response, generate and play a tactful reminder script if the low-quality response cumulative number is less than a first threshold value, update the low-quality response cumulative number after adding one, and record the low-quality response event in the background if the low-quality response cumulative number is greater than or equal to the first threshold value.

[0022] In the present application, the AI psychological interview response quality management method and system proposed first realizes high-precision, low-power real-time analysis of user response quality on the edge terminal side through the construction of a double-path cascade decision architecture of "zero floating point rule dominance + INT8 micro model supplement", which effectively distinguishes "low-quality response" and "naturally less talkative" based on individual baseline, greatly reduces the misjudgment rate, and at the same time, makes up for the blind area of pure semantic rules by accurately capturing the disordered features of acoustic energy envelope with the help of lightweight CNN network, thereby systematically improving the accuracy and robustness of screening. Secondly, in the interactive experience, a progressive gate intervention strategy based on a counter is innovatively introduced, which changes the harsh immediate warning into a data-driven, empathetic tactful reminder, ensuring the "low intrusion" and "high acceptance" of the intervention script, thereby significantly relieving the psychological pressure of the user and effectively ensuring the completion rate and data quality of the interview. Finally, in terms of system efficiency, a closed-loop feedback and management mechanism with "zero blame" as the core concept is established, which realizes gentle recall and accurate secondary interview of the missed screening population by automatically generating explanatory summaries and potential risk user lists, fundamentally solving the "interview ends, process ends" problem in the prior art, forming a whole-process quality guarantee closed loop of "screening-identification-intervention-recall", and finally achieving the comprehensive performance improvement of low power consumption, low misjudgment, low intrusion and high return flow. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A work flow diagram of an AI psychological interview response quality management method proposed in the present application; Figure 2A processing flow schematic diagram of an embodiment of an AI psychological interview response quality management method proposed in the present application; Figure 3 A score curve schematic diagram of an embodiment of an AI psychological interview response quality management method proposed in the present application; Figure 4 A structure schematic diagram of an embodiment of an AI psychological interview response quality management system proposed in the present application. DETAILED DESCRIPTION

[0024] Reference Figures 1-4 The AI psychological interview response quality management method proposed in the present application comprises the following steps: S0, in the conversation initialization stage, the response voice of the user to the preset welcome word is acquired, and after endpoint detection and word segmentation processing, the average word number and the average substantive word density of the user in the baseline period are calculated.

[0025] S1, the voice data input by the user is acquired, endpoint detection and word segmentation processing are performed on the voice data to obtain the word sequence, the frame number and the number of fillers of the current sentence, based on the rule path of zero floating point operation, the semantic features of the current sentence are calculated according to the word sequence, the frame number and the number of fillers, and the first quality score is obtained, and the zero floating point operation is an operation completely using integers.

[0026] In the present embodiment, based on the rule path of zero floating point operation, the semantic features of the current sentence are calculated according to the word sequence, the frame number and the number of fillers, and the first quality score is obtained, which specifically comprises: The average word number and the average substantive word density of the user in the baseline period are acquired as the personal baseline; According to the word number and the frame number of the current sentence, the sentence length feature flag is obtained by comparison with the personal baseline; According to the number of fillers and the total number of words of the current sentence, the filler proportion is calculated; According to the word sequence of the current sentence, the number of substantive words is identified and its density is calculated, and the substantive word sparsity flag is obtained by comparison with the personal baseline; Based on the sentence length feature flag, the filler proportion and the substantive word sparsity flag, the first quality score is calculated through a weighted summation formula.

[0027] It should be noted that the zero floating point operation means that after processing in the calculation model, no decimal data operation is involved, and only integer calculation is used to complete semantic feature extraction and score calculation, such as the above-mentioned rule engine which uses integer to determine short sentences, count the number of low-quality words, and convert the integer proportion, etc. The algorithm power consumption caused by floating point operation is avoided, and the lightweight device is adapted for fast processing.

[0028] Specifically, after the system is powered on, the pickup is continuously performed at 16 kHz, and the 640 ms sliding window buffer is moved forward by 10 ms frames, and the start / end of sentence interrupt signal is given by the front-end VAD. Then, the "zero floating point" rule engine is entered, and the pipeline is as follows: First, the system plays a welcome speech, for example, "Please introduce yourself briefly.", and then the user speaks, and the personal baseline is calculated within 30 seconds, including the average number of words , the average number of frames ; then for each sentence of the user, the word count W and the total frame count F are output by the word segmenter; Then, the relative short sentence judgment is performed, and the judgment formula is as follows: short_flag=(W≤ ×0.4)&&(F≤120)?1:0; The specific meaning of the above judgment formula is: when the current sentence is 60% shorter than the user's usual sentence and the duration is <1.2s, it is considered "too short", and short_flag represents the short sentence judgment value. short_flag is 1, indicating that the sentence is too short; short_flag is 0, indicating that the sentence is not too short, to prevent people who naturally speak less from being misjudged.

[0029] Pre-store 12 high-frequency low-quality words {um, uh, just, okay, don't know, random, average, nothing, see, all right, don't care, just like that}, and check the table while segmenting. 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; Specifically, the calculation formula of the filler word ratio is as follows: filler_ratio=filler_cnt / W; Where filler_cnt represents the number of low-quality words in the current sentence; W represents the total number of words in the current sentence; filler_ratio represents the low-quality word content in the current sentence, i.e. the filler word ratio.

[0030] In this embodiment, the number of content words is identified according to the word sequence of the current sentence, and the density is calculated, and the content word sparse flag is obtained by comparing with the personal baseline, which specifically includes: First, the local 200 kB static dictionary has a "content" flag (noun / verb / adjective) for each bit; content_cnt is the number of content word hits; density0= user's usual density = the content word density in the user's reply sentence within 30 seconds of the system welcome speech = content_cnt0 / W0; density= user's current sentence density = content word density in the current sentence = content_cnt / W; Wherein, the process of relative sparse determination is specifically: rel_density_flag = (density x 100 <= density0 x 60)? 1 : 0; Wherein, rel_density_flag is the current sentence density determination value, density0 is the average content word density, and density is the current sentence content word density. When the current sentence content word density is 40% lower than the user's usual density, it is considered "empty".

[0031] In the embodiment, the calculation formula of the first quality score is specifically: S = (20 x short_flag + 45 x filler_ratio + 35 x rel_density_flag) / 100; Wherein, 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 determination value.

[0032] 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 speech data to obtain a second quality score. The energy envelope feature refers to the feature of the energy of the speech signal changing with time, and the preset review interval is a score range between a preset lower limit and an upper limit threshold.

[0033] Specifically, the first quality score output is between 0 and 1, and the threshold value of the first quality score θ_rule = 0.70; the first quality score >= 0.70 is judged as "low quality".

[0034] It should be noted that the energy envelope feature refers to the feature of the energy of the speech signal changing with time, that is, the speech is divided according to a fixed time frame (such as 10 ms), the energy value of each frame is calculated (the greater the energy, the louder the speech), and the frame energy is arranged in time sequence to form a continuous energy change sequence (such as a 64-point sequence), which can reflect the strength of the speech. Fluctuation, is a key acoustic feature to determine whether the speech is clear and coherent.

[0035] In the embodiment, the lightweight neural network model is a one-dimensional convolutional neural network model quantized to INT8 precision, and the structure of the one-dimensional convolutional neural network model is in turn: an input layer for receiving a 64-point energy envelope sequence, a first convolutional layer with 8 convolutional kernels with a size of 8 and a step of 2, a second convolutional layer with 16 convolutional kernels with a size of 4 and a step of 2, a global maximum pooling layer, and a fully connected layer.

[0036] Specifically, the 64-point energy envelope sequence is specifically: ; wherein E is a 64-point energy envelope of 640ms speech with 10ms frame shift, single byte 0-255, the larger the value, the more responsive the frame is; is the envelope point number, taking values in [0, 63], a total of 64 points.

[0037] The first convolutional layer is specifically:

[0038] wherein is the height index of the output feature map; is the local offset of the convolution kernel on the input channel; is the input channel number; is the input sampling value of the 64-point energy envelope at the window position ; and are the weight of the first convolutional layer at the input channel, the kernel position and the shared bias of the output channel, both of which are continuously updated through gradient descent in the training stage; The linear convolution result is hard saturated and clipped to the interval [0, 6] to form the final output feature. is the activation output of the first convolutional layer at the output channel, the spatial position.

[0039] The second convolutional layer is specifically:

[0040] wherein is the height index of the output feature map; is the local offset of the convolution kernel on the height dimension; is the output channel number; is an 8-dimensional column vector of the output feature map of the first convolutional layer at all 8 input channels, height position (i.e., the values of the 8 channels of at the same spatial position are taken out at a time); and are the weight of the second convolutional layer at the output channel, the kernel inner offset and the shared bias of the second convolutional layer at the output channel, both of which are continuously updated through gradient descent in the training stage; The linear convolution result is hard saturated and clipped to the interval [0, 6] to form the final output feature. is the activation output of the second convolutional layer at the The output channel, the first The activation output is obtained from each spatial location.

[0041] The global max pooling layer is as follows: ; 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.

[0042] The fully connected layer is specifically:

[0043] 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.

[0044] In this embodiment, the training process of the lightweight neural network model specifically includes: 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; 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 was iteratively optimized using the training dataset to minimize the loss between the prediction results and the response quality label. After the training is completed, the network weights are solidified into INT8 precision, and the corresponding Sigmoid lookup table is generated, and finally a lightweight neural network model is obtained.

[0045] In this embodiment, the Sigmoid lookup table is a read-only memory with a capacity of 256 bytes, and 256 byte-type data are pre-burned in the Sigmoid lookup table internally; The byte-type data are configured in the following manner: the standard Sigmoid function is uniformly sampled in a predetermined input interval, and the function output value is mapped to the integer range of 0 to 255, thereby establishing a nonlinear mapping relationship from the 16-bit integer pre-stage output to the 8-bit probability value.

[0046] Specifically, considering that "energy is high and low, irregular" often accompanies casual responses, for example: Serious answer: The volume is stable, with stress and pauses, like "ECG" regular heartbeat.

[0047] Low-quality answer: Think of it and say it, sometimes shout a sentence, sometimes whisper, and the energy broken line is jumping around.

[0048] Therefore, when and only when the first quality score S is in the preset review interval [0.4, 0.7] and short_flag=0, the system determines that it is in the "regular edge area", and immediately starts the lightweight neural network model to quantify the low-quality typical feature of "energy envelope fluctuation disorder", and realizes "rules are not clear, machines listen again". In the end, it still needs to be calculated together with the sentence length and fillers, and it will not make a conclusion based on fluctuation alone.

[0049] S3, fusing the first quality score and the second quality score to obtain a final quality score, and outputting a single-bit flag representing the current response quality level.

[0050] In this embodiment, the first quality score and the second quality score are fused to obtain the final quality score, specifically: Final quality score= × first quality score+ × (second quality score / 255); Wherein, is the weight related to the first quality score; is the weight related to the second quality score; .

[0051] Specifically, 0.6, 0.4, then the final quality score = 0.6× first quality score + 0.4× (second quality score / 255); if the final quality score is greater than or equal to 0.55, it is judged as "low quality" and output gear 1; otherwise, gear 0.

[0052] The lightweight neural network model filling path of the embodiment captures the energy envelope disorder features in the "rule edge area", which is complementary to the hard rules, avoids misjudging low-volume users, accurately identifies low-quality answers with "not few words but absent-mindedness", and provides high-feasible concentration guarantee for self-help psychological counseling.

[0053] In the embodiment, a single-bit flag representing the current answer quality level is also output, specifically including: comparing the final quality score with a preset final decision threshold; if the final quality score is greater than or equal to the final decision threshold, outputting binary '1' representing a low-quality answer as the single-bit flag; if the final quality score is less than the final decision threshold, outputting binary '0' representing a high-quality answer as the single-bit flag.

[0054] S4, acquire the single-bit flag, when the single-bit flag indicates a low-quality answer, query the low-quality answer cumulative number stored in the memory, if the low-quality answer cumulative number is less than a first threshold, generate and play a tactful reminder script, and update the storage after adding one to the low-quality answer cumulative number; if the low-quality answer cumulative number is greater than or equal to the first threshold, record the current low-quality answer event in the background.

[0055] In the embodiment, the step of generating the tactful reminder script specifically includes: construct a context vector, the context vector containing a sentence length feature flag, a filler word proportion, a substantive word sparsity flag, a second quality score, a first quality score, a final quality score, a low-quality answer cumulative number, a current conversation turn, and a user's last three rounds of answer texts; combine the context vector with a preset prompt template, input it into a preset large language model, to obtain the text output by the large language model, and use the text output by the large language model as the tactful reminder script, the prompt template being used to limit the large language model to generate a co-empathetic response without negative words and with limited word count.

[0056] In the embodiment, it also includes: S5, acquire the low-quality answer cumulative number at the end of the conversation, if the low-quality answer cumulative number is greater than or equal to a second threshold, generate a low-quality answer trajectory table recording all low-quality answer details in the current conversation, generate and play a gentle closing statement based on the low-quality answer trajectory table, and generate a list containing the user identifier and the low-quality answer cumulative number and push it to the management end.

[0057] Embodiment 1: Take "the user completes a 15-round conversation in front of the AI psychological interview terminal, and the system identifies 4 low-quality answers" as an example, this application demonstrates how to complete the whole process of "low-power detection → low-intrusion reminder → zero-blame closed-loop feedback".

[0058] Step 1 system initialization: S1.1 terminal power on, load 200kB real word dictionary, 12 low-quality word hash table, 14kB CNN weight and 256B Sigmoid lookup table; S1.2 voice front end set 16kHz sampling, 10ms frame shift, 640ms sliding window; Initialize the reminded times counter_f to 0; Gear buffer flag initialization to 0.

[0059] Step 2 personal baseline collection (round 0): S2.1 terminal plays the welcome word "please introduce yourself"; S2.230s pickup, word segmentation output total word number , total frame number , real word number content_cnt0, calculate density0=content_cnt0 / W0; S2.3 personal baseline is written into SRAM for subsequent rule calling.

[0060] Step 3 real-time detection (round 1→15) per round: Perform 3.1-3.4 for each round of user voice: S3.1 voice input → VAD cut sentence → word segmentation → get W, F, filler_cnt, content_cnt; S3.2 rule path (zero floating point): short_flag=(W≤W0×0.4)&&(F≤120)?1:0; filler_ratio=filler_cnt / W; rel_density_flag=(content_cnt / W≤density0×0.6)?1:0; S=(20×short_flag+45×filler_ratio+35×rel_density_flag) / 100; S3.3 if S≥0.70 → directly judge "gear 1" jump to S3.5; If S∈[0.40,0.70] and short_flag=0, start lightweight neural network model: a) 640 ms energy envelope 64 points→INT8 quantization→CNN forward→get the second quality score ∈ [0, 255]; b) Final quality score = 0.6 * first quality score + 0.4 * (second quality score / 255); c) If the final quality score ≥ 0.55, judge "gear 1", otherwise "gear 0"; otherwise, judge "gear 0".

[0061] S3.4 Gear 0: End of this round, wait for the next sentence; S3.5 Gear 1: Enter the intervention process (step 4).

[0062] Step 4 Progressive intervention: S4.1 Check the number of reminders counter_f: If counter_f < 4, pack the 8-dimensional context vector, which includes: a) (short sentence judgment value, low quality word density, substantive word density judgment value, second quality score, first quality score, final quality score, number of reminders, and the number of the current round of dialogue in the whole psychological interview process and the original text of the last three rounds of dialogue); b) Feed the 8-dimensional vector into the "reminder coach" prompt word used by the large model, so that the large model returns a ≤28-character spoken response; In this embodiment, the corresponding prompt response for a low-quality response when applied is shown in Table 1: Table 1 Low-quality response reminder response comparison table:

[0063] c) The terminal plays the response; the number of reminders is increased by 1, i.e., counter_f++; d) Background database: record the current round of trajectory (round, text, each indicator, final quality score).

[0064] If counter_f ≥ 4: a) Only background database: record the current round of trajectory (round, text, each indicator, final quality score); b) Do not play any voice to avoid excessive reminders.

[0065] Step 5 Session end triggers feedback (round 15 ends): S5.1 When counter_f = 4, the system detects that the counter_f ≥ 4 "gentle ending" condition is met, and the user information is stored in the "low-quality response" user table for secondary screening interviews by the interview administrator. The number of low-quality responses is shown in Table 2: Table 2 Low-quality response frequency table:

[0066] S5.2 Background generation of "low quality track" table + snapshot of broken line, along with counter_f value, write into log, the log of this embodiment is shown in Table 3: Table 3 Low quality response log table:

[0067] As shown in Figure 3 , the black curve in the figure is the first quality score, and the coordinates of the four black points marked on the curve are (1, 1), (9, 0.45), (10, 1), (13, 0.45). The first quality score of the first round and the tenth round is 1, which is immediately judged as "low quality response". The first quality score of the ninth round and the thirteenth round is 0.45, which is between 0.4-0.7, and short_flag=0, which is within the "low quality response" fuzzy boundary, so the final quality score is continued to be calculated. The red point (9, 0.62) in the figure represents that the final quality score of the ninth round is 0.62, and (13, 0.62) represents that the final quality score of the thirteenth round is 0.62, both of which are greater than 0.55, so it is judged as "low quality response".

[0068] S5.3 In the feedback link, feed S5.2 "low quality track" table into the feedback prompt word used by the language large model, so that the large model outputs ≤80 words of gentle closing statement; S5.4 Terminal broadcast example statement: "Recently you have 4 times of short answers, which may be not ready or network inconvenience. Therefore, today's evaluation may not fully present your true state. Thank you for your willingness to come, and look forward to the next time we can complete the remaining puzzle together. I wish you a little more relaxed than when you come tonight." Step 6 Zero-intrusion silence recording (if counter_f < 4) S6.1 Session ends, counter_f < 4, system only writes track table into background, does not play any additional voice, realizes zero-intrusion silence recording.

[0069] Step 7 Reset and wait for the next consultation S7.1: counter_f <- 0 (number of reminders is reset to zero), gear buffer is cleared, and step 2 is returned to wait for a new user or the next round of consultation of the same user.

[0070] So far, the operation flow demonstration is complete, which fully covers the cooperative work of "detection → reminder → feedback".

[0071] Refer to Figures 1-4The present invention proposes an AI-based psychological interview response quality management system, comprising: 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. 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. 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 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.

[0072] 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.

[0073] 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.

[0074] In this embodiment, the activation conditions for the response reminder module are as follows: ① First check the first quality score and the final quality score to confirm that this round is indeed "gear 1"; 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.

[0075] 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