A spoken dialogue dynamic generation method and system based on voice emotion feedback

Abstract

The application discloses a spoken dialogue dynamic generation method and system based on voice emotion feedback and belongs to the technical field of voice interaction. The application obtains emotion change trend parameters, current task load indexes and interactive load acceleration parameters by performing data processing on the obtained data, and then calculates the comprehensive intervention tendency value by taking the emotion change trend parameters and the interactive load acceleration parameters as input. If the comprehensive intervention tendency value reaches the early warning trigger threshold value and does not reach the intervention trigger threshold value, preventive protection state data is output. If the comprehensive intervention tendency value reaches the intervention trigger threshold value, real-time intervention state data is output. According to the data output by the judgment, dialogue generation parameters are dynamically adjusted to output target dialogue reply data. The application improves the synergy of user emotion fluctuation and interactive load state in the spoken dialogue generation process and solves the problem that the existing dialogue system lacks the function of synergistically adjusting user emotion fluctuation and interactive load state.

Description

Technical Field

[0001] This invention relates to the field of voice interaction technology, and in particular to a method and system for dynamically generating spoken dialogue based on voice emotion feedback. Background Technology

[0002] With the acceleration of globalization and the frequency of international exchanges, language proficiency, especially oral proficiency, has become a crucial factor in personal career development and social adaptation. Particularly in second language learning, many learners face emotional barriers such as anxiety, nervousness, and lack of confidence in oral expression, affecting their learning outcomes and oral proficiency improvement. Meanwhile, traditional oral learning methods often lack personalization and emotional support, failing to meet learners' psychological and emotional needs. Technological advancements, particularly innovations in speech recognition, sentiment analysis, and natural language generation, have offered new possibilities for addressing this issue. Methods for dynamically generating spoken dialogues based on emotional feedback in user speech have emerged, aiming to provide a personalized and highly interactive oral practice experience by dynamically adjusting dialogue content and feedback methods based on real-time identification of learners' emotional states. This technology not only alleviates learners' emotional barriers and enhances learning motivation but also improves the efficiency and enjoyment of oral learning. Therefore, with the increasing demand for personalized education and emotion-driven learning, the social demand for this method and system in language education, psychological counseling, and intelligent learning is growing.

[0003] Existing methods primarily utilize three core technologies: speech emotion recognition, dialogue generation, and emotion feedback modulation. First, the system uses speech emotion recognition technology to analyze learners' emotional states in real time, extracting speech features (such as pitch, intonation, speech rate, and pauses), and then transforming these features into specific emotion categories and intensities using machine learning models (such as LSTM and CNN). This emotion data is then passed to the dialogue generation module, which generates natural language dialogue content based on pre-trained language models (such as GPT and BERT). The generated dialogue considers not only grammar and semantics but also the learner's emotional state, adjusting the tone, speech rate, and content challenge to match the learner's emotional needs. During this process, the system dynamically adjusts its feedback strategy based on real-time emotion feedback, providing encouraging, reassuring, or challenging feedback to help learners alleviate negative emotions or stimulate learning motivation. Through this emotion-driven dynamic dialogue generation, the system can provide learners with a personalized, immersive oral practice experience, optimizing learning outcomes and promoting emotion modulation. This approach not only improves oral fluency but also opens up new avenues for the application of affective computing technology in education.

[0004] For example, Chinese invention patent CN120726996B discloses a method for processing spoken dialogue data based on an intelligent agent. This method includes: receiving a voice input file sent by a user; converting the voice signal into text data; extracting emotional features; generating emotional tags and speech prosody features; combining text and emotional information to form dialogue anchor information; constructing a path graph; calculating edge weights; constructing a propagation matrix based on the similarity between the current dialogue anchor and historical path nodes; performing intelligent path reasoning; calculating path scores to ensure that emotional fluctuations do not affect task path selection; the intelligent agent selects the optimal path for subsequent task processing based on the path score; and making real-time corrections based on user feedback; adjusting the current dialogue anchor information in real-time using sentiment analysis technology; and adjusting the reasoning process based on feedback to form a closed-loop adaptive adjustment mechanism, thereby optimizing the system's response quality and enhancing user trust in the system.

[0005] However, the above-mentioned technologies have at least the following technical problems:

[0006] In existing technologies, traditional dialogue systems largely rely on rules and model dialogue templates, lacking emotion perception and dynamic adjustment mechanisms. While semantic understanding can process user intent, it ignores the influence of factors such as emotional changes and psychological states. Furthermore, the interaction design of most systems fails to adjust the frequency and complexity of feedback in real time, leading to excessive pressure on users under high load or negative emotions, or a lack of challenge when users are in a good mood. Therefore, existing dialogue systems suffer from the problem of generating responses solely based on semantic content and lacking the ability to coordinate and adjust to user emotional fluctuations and interaction load. Summary of the Invention

[0007] To address the technical problems of existing dialogue systems that generate responses solely based on semantic content and lack the ability to coordinate adjustments based on user emotional fluctuations and interaction load, this invention provides a method and system for dynamically generating spoken dialogue based on voice emotional feedback. The technical solution is as follows:

[0008] On the one hand, a method for dynamically generating spoken dialogue based on voice emotion feedback is provided. This method includes: acquiring user voice signals, semantic text data, and interaction behavior data for the current interaction round; inputting the user voice signals into a preset emotion representation model for feature extraction to output continuous emotion change trend parameters; and calculating the current task load index based on the semantic text data and interaction behavior data according to preset rules; calculating the interaction load acceleration parameter based on the rate of change of the current task load index within a preset time window; constructing a correlation function between the emotion change trend parameter and the interaction load acceleration parameter; and calculating a comprehensive intervention tendency value through the correlation function; and acquiring a preset representation... The system uses a dual-threshold range between the warning trigger threshold and the intervention trigger threshold. If the comprehensive intervention tendency value reaches the warning trigger threshold but does not reach the intervention trigger threshold, preventive protection status data is output. If the comprehensive intervention tendency value reaches the intervention trigger threshold, real-time intervention status data is output. The preventive protection status data represents an overload warning before a user's emotions experience drastic fluctuations. Based on the output preventive protection status data or real-time intervention status data, preset dialogue generation parameters are dynamically adjusted. These parameters include a semantic dimensionality reduction step size set for the preventive protection status data. The adjusted dialogue generation parameters are input into the dialogue generation engine to generate and output target dialogue response data.

[0009] On the other hand, a dynamic oral dialogue generation system based on voice emotion feedback is provided. This system includes: a multi-dimensional data acquisition module for acquiring user voice signals, semantic text data, and interaction behavior data for the current interaction round; a feature quantization processing module for inputting user voice signals into a preset emotion representation model for feature extraction, outputting continuous emotion change trend parameters, and calculating the current task load index based on semantic text data and interaction behavior data according to preset rules; calculating the interaction load acceleration parameter based on the rate of change of the current task load index within a preset time window; constructing a correlation function between the emotion change trend parameter and the interaction load acceleration parameter, and calculating a comprehensive intervention tendency value through the correlation function; and determining the correlation threshold. The module is used to acquire a preset dual-threshold interval representing the range between the warning trigger threshold and the intervention trigger threshold. If the comprehensive intervention tendency value reaches the warning trigger threshold but does not reach the intervention trigger threshold, it determines and outputs preventive protection status data. If the comprehensive intervention tendency value reaches the intervention trigger threshold, it determines and outputs real-time intervention status data. The preventive protection status data represents a load overload warning before a user's emotions fluctuate drastically. The dynamic generation control module is used to dynamically adjust the preset dialogue generation parameters based on the determined preventive protection status data or real-time intervention status data. The dialogue generation parameters include the semantic dimensionality reduction step size set for the preventive protection status data. The adjusted dialogue generation parameters are input into the dialogue generation engine to generate and output the target dialogue response data.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0011] 1. This invention provides a method for dynamically generating spoken dialogue based on voice emotion feedback. By comprehensively analyzing the user's voice signal, semantic text data, and interactive behavior data, it achieves accurate monitoring and dynamic adjustment of the user's emotions and interactive load. Furthermore, through the calculation of emotion change trend parameters and interactive load acceleration parameters, the system can assess the user's emotional fluctuations and task load status in real time, thereby generating a comprehensive intervention tendency value. Based on this, the system outputs preventative protection or real-time intervention status data in a timely manner according to preset dual threshold intervals, preventing user discomfort when experiencing severe emotional fluctuations or overload. Simultaneously, this method optimizes dialogue generation and response by dynamically adjusting dialogue generation parameters, such as semantic dimensionality reduction step size. This allows the system to provide appropriate tone, content, and interaction methods when the user's load is too high or their emotions are unstable, effectively reducing user stress, improving the interactive experience, and enhancing user engagement. This improves the synergistic regulation of user emotional fluctuations and interactive load status, effectively solving the problem that existing dialogue systems only generate responses based on semantic content and lack synergistic regulation of user emotional fluctuations and interactive load status.

[0012] 2. By constructing a correlation function between continuous emotion change trend parameters and interaction load acceleration parameters, a multi-dimensional fusion perception of users' emotional-cognitive state is achieved. The dual threshold interval judgment mechanism can provide early warning before drastic emotional fluctuations, taking into account both the timeliness and accuracy of intervention. The hierarchical dynamic adjustment strategy of dialogue generation parameters ensures the precise adaptation of intervention measures. The adaptive feedback update mechanism effectively improves the system's long-term adaptive capability and personalized service level. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating a method for dynamically generating spoken dialogue based on voice emotion feedback, provided in an embodiment of this application;

[0015] Figure 2 This is a schematic diagram of the structure of a spoken dialogue dynamic generation system based on voice emotion feedback, provided in an embodiment of this application. Detailed Implementation

[0016] The technical solution provided in this application will now be described with reference to the accompanying drawings.

[0017] To facilitate understanding of the embodiments of this application, the following points will be explained first:

[0018] First, in this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but it does not exclude the possibility of indicating an "and" relationship; the specific meaning can be understood in context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0019] Second, the use of prefixes such as "first" and "second" in this application is solely for the purpose of distinguishing and describing different things belonging to the same category, and does not constrain the order, size, or quantity of things. For example, "first message" and "second message" are simply different messages, and there is no chronological, size, or priority relationship between them.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] like Figure 1 The diagram shown is a flowchart of a method for dynamically generating spoken dialogue based on voice emotion feedback provided in an embodiment of this application. The method includes the following steps:

[0022] S1: Obtain user voice signals, semantic text data, and interaction behavior data for the current interaction round.

[0023] Specifically, user voice signals are acquired in real time via microphone with a sampling rate of no less than 16kHz to preserve sufficient spectral details; semantic text data is generated by an automatic speech recognition (ASR) engine that transcribes the acquired speech. The ASR engine can be replaced with an implementation based on pre-trained models such as Wav2Vec2 and Whisper; interactive behavior data is extracted synchronously from front-end interaction logs, including quantifiable behavioral indicators such as user response latency and the number of times answers are modified.

[0024] The simultaneous collection of these three types of data provides a data guarantee for subsequent emotion-cognition dual-dimensional modeling. For example, taking Zhang's fifth round of dialogue as an example: the system collects his English speech signal with hesitation and pauses, transcribes it into corresponding text using ASR, and records the response delay time of 4.2 seconds for this round. The above three types of data are simultaneously sent into the subsequent processing flow; by establishing a multimodal data acquisition channel through step one, a more comprehensive raw input is provided for subsequent emotion feature extraction and cognitive load calculation.

[0025] S2 inputs the user's voice signal into a preset emotion representation model, extracts features, outputs continuous emotion change trend parameters, and calculates the current task load index based on semantic text data and interaction behavior data according to preset rules.

[0026] It should be noted that, in this embodiment, the preset emotion representation model is not a general discrete emotion classifier, but a dynamic evaluation model based on the user's "interaction baseline". For scenarios such as oral practice and language rehabilitation, the model establishes a speech feature baseline under the user's specific normal state (no stress state), and calculates the difference between the features extracted in real time to eliminate the interaction of individual speech differences on emotion judgment.

[0027] In the context of language rehabilitation: the model weights are tilted towards "physiological characteristics of pronunciation". By monitoring the jitter and shimmer of the user during the attempt to pronounce words, it is determined whether the user experiences frustration due to pronunciation obstruction.

[0028] In the context of spoken language practice: the model weights are tilted towards "prosodic logic features". By monitoring the rate of change of speed and the rationality of pause positions, it is determined whether the user is experiencing anxiety due to cognitive overload.

[0029] To ensure the model can capture the preceding signals of emotion, the input to the pre-defined emotion representation model is configured to receive a relevant feature tensor consisting of the following three components:

[0030] Physical acoustic features: including the first 40 Vimel frequency cepstral factors (MFCC) and the fundamental frequency (F0) fluctuation sequence, used to characterize the envelope and pitch fluctuations of the speech signal.

[0031] Prosodic temporal features include energy entropy, dynamic rate of speech change, and duration of unexpected pauses, used to map the representation of emotion on the "Arousal" dimension.

[0032] Icon-related features: Keyword repetition rate and error correction feedback frequency are extracted from speech recognition results as external reference data to help judge the stability of user emotions.

[0033] It should also be noted that the representation model employs a bidirectional recurrent neural network (Bi-LSTM) or a transformer encoder with a self-focusing mechanism. Its processing procedure and data association logic are as follows:

[0034] I. Feature Texture and Fusion: The model performs texture mapping on the time axis of the input physical acoustic features and prosodic temporal features to construct a high-dimensional feature space.

[0035] 2. Spatial coordinate mapping: The algorithm uses a nonlinear mapping function to map the fused features to the initial VA (Valence-Arousal) continuous sentiment space and outputs real-time sentiment coordinates.

[0036] III. Calculation of Change Trend: The system performs a first derivative attack on the emotion coordinates to obtain the parameters of emotion change trend.

[0037] Internal correlation logic: When the emotional change trend parameter points to the negative interval of the VA space (decreased valence) and the arousal value increases, the model variable user is shifting from "calm" to "anxious" or "depressed".

[0038] Stability determination: The model calculates the reciprocal of the angle of the emotion coordinate within the time window and outputs the emotion stability parameter. This parameter is used to modify the subsequent interference trend value: When the emotion stability parameter is low, it indicates that the user is emotionally agitated, and the system will automatically adjust the high intervention weight.

[0039] Furthermore, the model's output data directly serves as the core input for the next step, the "coupling agent." Its output structure is specifically set as follows:

[0040] Output 1: The rate and direction of movement of the emotion change trend parameter in two-dimensional space represent emotion.

[0041] Output 2: The sentiment stability parameter predicts the frequency of sentiment oscillation within the current trading segment.

[0042] Output 3: Emotional arousal intensity represents the intensity of the user's current emotional outburst and is used to trigger the emergency trigger mode.

[0043] In a specific embodiment, such as in a spoken language practice scenario, if the task burden index indicates that the user is processing long and complex sentences outside their boundary range, such as the English translation of Qu Yuan's poems, which contain complex emotions such as mythological imagery, political metaphors, and the imagery of fragrant herbs and beautiful women, and the interaction burden acceleration parameter is greater than the preset step size, it indicates that the user's cognitive pressure has accumulated. At this time, the reinforced emotion representation model does not yet have a clearly defined discrete label of "anger" or "give up," but due to the continuous negative shift in its output emotion change trend parameter, the algorithm will calculate a high comprehensive intervention tendency value. If this tendency value falls into the preset "warning trigger threshold" range, the system will determine that the current state is "preventive protection state" and actively adjust the semantic dimensionality reduction step size in the dialogue generation parameter vector. Specifically, the dialogue content generated in the next round will be automatically reconstructed from "open-ended essay questions" to "selection-guided questions with prompts," thereby unloading the cognitive load in advance before the user's emotions collapse.

[0044] This step encompasses two parallel core sub-processes: extracting emotion change trend parameters based on an emotion representation model, and calculating the task load index based on multi-dimensional behavioral indicators. Unlike traditional solutions that limit emotion recognition results to limited category labels (such as discrete states like happiness and anger), this invention outputs emotion change trend parameters in continuous numerical form, which is more conducive to capturing the fine-grained evolution of emotions over time. Simultaneously calculating the task load index also allows the system to consider the cognitive stress dimension in addition to the emotional dimension, avoiding a one-sided description of the user's overall state from a single emotion perception.

[0045] The process of inputting user voice signals into a pre-defined emotion representation model for feature extraction and outputting continuous emotion change trend parameters includes the following stages:

[0046] Acoustic feature extraction stage: Extract acoustic feature data from the user's speech signal. The acoustic feature data includes Mel frequency cepstral coefficients and fundamental frequency feature data.

[0047] It should be added that Mel frequency cepstral coefficients (MFCC) provide a compact representation of the speech spectrum envelope by simulating the nonlinear frequency scale perceived by the human ear, typically extracting coefficients with 13 to 39 dimensions; the fundamental frequency (F0) feature reflects pitch, and its mean and standard deviation within a short time window can effectively distinguish emotional states such as tension and calmness. These two types of features can be further extended into an extended acoustic feature vector that includes fundamental frequency jitter, amplitude perturbation (shimmer), and harmonic-to-noise ratio (HNR) to improve the accuracy of emotion mapping.

[0048] Two-dimensional spatial mapping stage: The acoustic feature data is mapped to the preset effective valence wake-up two-dimensional space to generate continuous coordinate data; the time derivative of the continuous coordinate data is calculated according to the preset time window to output the emotional change trend parameters.

[0049] It should be added that the Valence-Arousal two-dimensional sentiment model represents the emotional state as a continuous coordinate point V(t)=(v(t),a(t)) in the plane, where v(t)∈[-1,1] represents the positive or negative polarity of the emotion, and a(t)∈[0,1] represents the intensity of the emotional activation; the mapping function is implemented by a pre-trained sentiment regression model, which can be replaced by support vector regression (SVR) or a Transformer-based sentiment estimation model.

[0050] Trend parameter calculation stage: Calculate the reciprocal of the coordinates of continuous coordinate data within the time window, and output the sentiment stability parameter.

[0051] In a specific embodiment, let the time window length be T (e.g., T=3 seconds), and the emotion change trend parameter be defined as: Δv = v(t) − v(t − T), Δa = a(t) − a(t − T). Taking Zhang's fifth round as an example: let V(t) = (−0.3, 0.7), V(t − T) = (−0.1, 0.4), then ΔV(t) = ((−0.3 − (−0.1)) / 3, (0.7 − 0.4) / 3) = (−0.067, 0.10), indicating a shift in emotion towards negativity, a rapid increase in arousal, and an intensified anxiety trend. The emotional stability parameter S(t) is obtained by calculating the reciprocal of the coordinates of continuous coordinate data within the time window, and a smaller S(t) indicates a greater fluctuation in the emotional trajectory and a more unstable emotional state.

[0052] By replacing discrete emotion categories with continuous coordinate change rates, we can accurately quantify the direction and rate of emotion evolution, providing higher-quality emotion-side input for subsequent two-dimensional fusion calculations.

[0053] Meanwhile, based on semantic text data and interaction behavior data, the current task load index is calculated according to preset rules, as follows:

[0054] The first step is to extract keyword encoding data and encoding frequency data from the text data.

[0055] The second step is to extract user response latency data from the interaction data.

[0056] The third step involves weighting and summing the keyword encoding data, encoding frequency data, and user response delay time data with the set scenario weights to output the current task load index.

[0057] Specifically, keyword encoding data K reflects the number of types of technical terms or difficult words in the current text; encoding frequency data F reflects the total number of times these words appear in the current round of dialogue text; user response delay time data D reflects the time interval between when the user receives the system question and when they begin to answer, and the longer the delay, the heavier the cognitive processing burden usually is.

[0058] It should be added that the current task load index L(t) is calculated using the following formula: Where w1, w2, and w3 are preset scene weights for the pre-defined staff and experience rules, and satisfy the following conditions: , The normalized value of user response latency data is pre-configured according to the application scenario.

[0059] For example, in an oral English teaching scenario, the three weights can be set as w1=0.3, w2=0.3, and w3=0.4. Specific numerical demonstration: Continuing with Zhang's fifth round as an example, K=5 (including grammatical terms such as conditional sentence and subjunctive mood), F=8 times, and D=4.2 seconds (after normalization). =0.84), substituting into the formula: L(5)=0.3×5+0.3×8+0.4×0.84=1.5+2.4+0.336=4.236; after normalizing the dimensions, take L(5)=0.72, then this value will be used as the current task load index for this round and input into the subsequent acceleration calculation steps.

[0060] S3 calculates the interactive load acceleration parameters based on the rate of change of the current task load index within a preset time window.

[0061] This step enables the calculation of the second-order change (acceleration) of the current task load index, based on the first-order change (velocity), to quantify the urgency of recognizing the increasing load. Compared to a scheme that only uses the first-order load velocity, load acceleration is beneficial for more timely detection of signals of rapid load increases: when the load velocity is already high and is still accelerating, the system can issue an early warning before the load enters the danger zone, providing a window of time for intervention measures to take effect.

[0062] Specifically, the interactive load acceleration parameter is calculated based on the rate of change of the current task load index within a preset time window. The specific process is as follows:

[0063] First, calculate the difference between the current task load index and the historical task load index to obtain the first-order load rate data for the corresponding time point.

[0064] Specifically, first-order load speed data Defined as: , where L(t) is the current task load index and L(t−1) is the historical task load index of the previous round. This value represents the change in load between rounds. A positive value indicates that the load is increasing and a negative value indicates that the load is decreasing.

[0065] Next, the historical load speed data set within the preset time window is obtained; the historical load speed data set is read from the historical data set in the preset database.

[0066] Then, the rate of change between the first-order load rate data and the historical load rate data set is calculated.

[0067] Finally, the rate of change data is output as the interactive load acceleration parameter, which is used to characterize the urgency of the user's cognitive load increase.

[0068] It needs to be explained that the interactive load acceleration parameter Defined as: ,in Preset time window The mean of the historical load speed data set. The number of rounds the time window spans (e.g., taking...) =3 rounds).

[0069] Specific numerical demonstration: Taking Zhang as an example again, the load indices for the last three rounds are L(3)=0.50, L(4)=0.61, and L(5)=0.72. , , =0.11, =(0.11−0.11) / 1=0.00, indicating that the load is increasing at a constant rate; if L(6)=0.89 in the 6th round, then =0.17, =(0.17−0.11) / 1=0.06, indicating that the rate of load increase is also accelerating, and the level of cognitive stress urgency is increasing.

[0070] This step defines the layer-by-layer calculation process for the interactive load acceleration parameters, introducing the concept of "acceleration" from physics to accurately quantify the dynamic evolution trend of the cognitive load. The meanings of each calculation step are as follows. Simultaneously, the "inertia" information of load changes is introduced through second-order differentiation, enabling the system to identify accelerating deterioration trends before the absolute value of the load exceeds a threshold. This provides a basis for early intervention, thus forming a temporal synergy with the dual-threshold mechanism in subsequent steps.

[0071] S4. Based on the emotional change trend parameter and the interaction load acceleration parameter, a correlation function is constructed between them, and the comprehensive intervention tendency value is calculated through the correlation function.

[0072] This step maps the trend parameters from the emotion dimension and the load acceleration parameters from the cognitive dimension to a unified intervention tendency value through a correlation function, achieving joint perception along the emotion-cognition dual axes. Since relying solely on emotional signals may lead to misjudgments due to short-term fluctuations in emotion itself, or relying solely on cognitive load indicators may fail to perceive the user's subjective emotional stress, joint modeling of both can more accurately depict the user's true comprehensive stress state, thus supporting subsequent grading determination within dual threshold intervals.

[0073] It should be noted that the comprehensive intervention propensity value is calculated using a correlation function, and the specific steps are as follows:

[0074] Step 1: Normalize the emotion change trend parameter and the interaction load acceleration parameter to obtain normalized emotion data and normalized acceleration data.

[0075] Specifically, the normalization process uses the Min-Max method to map the original parameters to the [0,1] interval: Normalized sentiment data Normalized acceleration data ,in , , , These are the historical extreme values ​​of the two types of parameters, respectively.

[0076] Step 2: Obtain preset emotion-related weight data and load-related weight data; both emotion-related weight data and load-related weight data are preset data, which are set in advance by preset staff based on historical data and experience rules, and stored in preset data.

[0077] Step 3: Weight the normalized sentiment data with the sentiment-related weight data, and weight the normalized acceleration data with the load-related weight data. Combine the results of the two weighted calculations to output the comprehensive intervention tendency value.

[0078] It should be noted that the formula for calculating the propensity score I(t) for comprehensive intervention is as follows: ,in For emotion-related weighted data, For load-related weighted data, satisfy + =1. The initial weight can be set to 1. =0.5、 =0.5.

[0079] Specifically, taking Zhang's fifth round as an example, |ΔV(5)| after normalization =0.78 (Emotions rapidly shift towards anxiety). After normalization =0.65 (the load acceleration is moderately high), substituting into the formula: I(5)=0.5×0.78+0.5×0.65=0.390+0.325=0.715.

[0080] Through a configurable weighted fusion mechanism, the system can flexibly adjust the relative weights of emotional and cognitive factors in comprehensive intervention decisions based on different scenarios or individual user differences; and the dynamic updating capability of the weights enables the system to gradually adapt to the emotional-cognitive response characteristics of specific users, thereby improving the accuracy and adaptability of intervention decisions.

[0081] S5, obtain a preset dual threshold interval representing the range between the warning trigger threshold and the intervention trigger threshold. If the comprehensive intervention tendency value reaches the warning trigger threshold but does not reach the intervention trigger threshold, then determine to output preventive protection status data. If the comprehensive intervention tendency value reaches the intervention trigger threshold, then determine to output real-time intervention status data. The preventive protection status data represents the overload warning before the user's emotions fluctuate drastically.

[0082] It should be noted that the dual threshold interval is determined by the warning trigger threshold. Intervention trigger threshold Common definition ( < The comprehensive intervention propensity value I(t) is divided into three continuous intervals: the normal interval [0, ...]. ), warning interval [ , ) and mandatory intervention intervals [ ,1]. When I(t) < When the system does not output intervention status data, the dialogue generation parameters remain at their current settings; when ≤I(t)< When I(t) ≥ 1, the system outputs preventative protection status data, triggering semantic dimensionality reduction transition adjustment (only gradually reducing vocabulary difficulty while maintaining the speech rate); At that time, the system outputs real-time intervention status data, triggering a downgrade and reconstruction adjustment (synchronously reducing vocabulary difficulty, slowing down the speech rate, and injecting reassuring prompts).

[0083] By using a dual-threshold design, the single-threshold scheme is prone to frequent switching and jitter near the threshold. The introduction of the warning interval is equivalent to setting a buffer transition zone between the two states, which can prevent emotional breakdown in advance and avoid excessive interruption of the normal interaction process by forced intervention, thereby improving the stability of the system and the user experience.

[0084] Specific example: Taking student Zhang as an example, let... =0.60, =0.80. In the 5th round, I(5) = 0.715, falling into the warning interval [0.60, 0.80), and the system outputs preventive protection status data; if I(7) = 0.83 in the 7th round, it falls into the mandatory intervention interval [0.80, 1], and the system outputs real-time intervention status data, driving more powerful parameter adjustments. The initial threshold value can be configured according to the scenario and adaptively adjusted across rounds through the subsequent threshold dynamic update mechanism.

[0085] S6. Based on the preventive protection status data or real-time intervention status data output by the judgment, dynamically adjust the preset dialogue generation parameters. The dialogue generation parameters include the semantic dimensionality reduction step size set for the preventive protection status data.

[0086] This step, based on the data type of the intervention state output by S5, dynamically adjusts the dialogue generation parameters in a differentiated manner. It is a key control interface for transforming emotion-cognitive perception results into dialogue generation behavior. Among these parameters, the semantic dimensionality reduction step size is a core control parameter designed specifically for preventative protective states. Its function is to gradually reduce the vocabulary difficulty in the response content at a controlled rate, avoiding abrupt transitions that could disrupt the dialogue experience. This parameter is also applicable in real-time intervention states. Furthermore, by introducing a tiered parameter adjustment mechanism, the system can implement gentle, gradual interventions during the warning phase and mandatory multi-dimensional joint interventions during the real-time intervention phase. This ensures that each type of intervention matches the severity of the current user's state, achieving more precise adaptation.

[0087] It should be explained that the dialogue generation parameters include vocabulary difficulty indicators and speech rate control indicators. Based on the preventive protection status data or real-time intervention status data of the judgment output, the preset dialogue generation parameters are dynamically adjusted, as follows:

[0088] If preventive protection status data is received, the preventive intervention instructions are extracted from the preventive protection status data; based on the preventive intervention instructions, the vocabulary difficulty index in the dialogue generation parameters is updated by decreasing the value through semantic dimensionality reduction step size, while keeping the speech rate control index data unchanged, and the adjusted dialogue generation parameters after dimensionality reduction transition are generated.

[0089] Specifically, vocabulary difficulty index Numerical difficulty levels are used (e.g., levels 1-10), and the semantic dimensionality reduction step size δ_v is the decreasing magnitude of each adjustment (e.g., ...). =Level 1), the update formula is: That is, in the next round, the dialogue generation engine will be based on the reduced... Generate responses with lower lexical complexity to help users free up cognitive space, and control speech rate metrics. The reason for keeping the speaking speed unchanged is that slowing down the speaking speed would give users a noticeable perception of a change in the rhythm of the conversation. In the warning stage where there is no drastic emotional fluctuation, maintaining a stable speaking speed helps to maintain the natural fluency of the conversation and avoids giving users unnecessary psychological cues.

[0090] In a specific embodiment, taking Zhang's 5th round (warning state, I(5)=0.715) as an example: Currently (5) = Level 7 =1, then (6) = 7 − 1 = Level 6. In the 6th round, the system will use a more concise grammatical structure and basic vocabulary to respond. For example, questions that originally contained difficult grammar such as subjunctive mood will be changed to questions using basic if conditional sentences, while the speaking speed will remain the same.

[0091] By gradually reducing the difficulty of vocabulary, cognitive pressure is released in advance before users' emotions fluctuate drastically, which acts as a buffer against overload and helps reduce the probability of triggering real-time intervention, making the system's intervention behavior more transparent and natural to users.

[0092] It should be added that the dialogue generation parameters also include preset speech rate slowdown step size and soothing prompt label data.

[0093] If real-time intervention status data is received, the mandatory intervention instructions are extracted from the real-time intervention status data. Based on the mandatory intervention instructions, the vocabulary difficulty index in the dialogue generation parameters is updated by decreasing the semantic dimensionality reduction step size, and the speech rate control index is updated by decreasing the speech rate step size. The soothing prompt label data is injected into the dialogue generation parameters to generate the downgraded and reconstructed dialogue generation parameters.

[0094] It should be noted that the vocabulary difficulty adjustment is the same as in S6, and the execution... The formula for adjusting the speech rate control index is: ,in The speech rate is slowed down by a step size (e.g., a reduction of approximately 30 words per minute). After the data of soothing prompt labels is injected, the dialogue generation engine will embed emotionally guiding discourse structures into the generated responses, such as statements that guide students to slow down and provide positive encouragement. The difference between this and the "dimensional reduction transition" mode is that the "dimensional reduction transition" mode only adjusts vocabulary difficulty and is suitable for moderate stress situations where the overall intervention tendency value falls into the warning range; the "downgrade reconstruction" mode, in addition to decreasing vocabulary difficulty, also adds a slowdown in speech rate and the injection of soothing prompts, forming a three-dimensional synergistic intervention, suitable for situations where the overall intervention tendency value has exceeded the intervention trigger threshold. In highly urgent situations, the two modes form a gradient intervention response system driven by thresholds.

[0095] Specific example: Taking Zhang's 7th round (real-time intervention status, I(7)=0.83) as an example: class, (Speed ​​reduction) The soothing prompt tag is activated, and the system generates a gentle, concise response containing emotionally reassuring content. This multi-dimensional collaborative intervention helps reduce Zhang's cognitive and emotional stress. Simultaneously, by reducing vocabulary difficulty, speech speed, and emotional tone, the system alleviates the user's overall stress level to the greatest extent possible in the shortest time, preventing a complete breakdown in interaction due to emotional distress.

[0096] S7 inputs the adjusted dialogue generation parameters into the dialogue generation engine to generate and output the target dialogue response data.

[0097] This step involves passing the dynamically adjusted dialogue generation parameters from S6 to the dialogue generation engine, driving it to output target dialogue response data that matches the user's current emotional and cognitive state. It is the final execution stage to complete the single-round intervention loop. The dialogue generation engine can be implemented based on a Large Language Model (LLM). Through structured prompt engineering, parameters such as lexical difficulty indicators, speech rate control indicators, and soothing prompt labels are encoded as generation constraints and injected into the system prompt, thereby constraining the model's generation behavior. Alternative solutions include rule-based generation engines based on template filling (suitable for resource-constrained scenarios) and dialogue generation systems based on Retrieval Augmentation (RAG). By uniformly transforming the multi-dimensional intervention parameters calculated in the preceding steps into generation constraints for dialogue content, a complete closed-loop control chain from emotional and cognitive perception to dialogue behavior adjustment is achieved, ensuring that the output target dialogue response data is adapted to the user's current state in terms of semantic content, expression form, and emotional tone.

[0098] Continuing with Zhang as an example: The system will class, Input parameters such as the reduced set value and soothing prompt labels into the LLM to generate a target response that is concise, slow-paced, and contains emotionally comforting content, thus completing this round of intervention. The status parameters of this round are stored for use in subsequent threshold dynamic updates and adaptive feedback update steps.

[0099] It should be noted that generating and outputting the target dialogue response data, followed by a step of dynamically updating the threshold:

[0100] S71, obtain the latest task load index for the next interaction round and calculate the load difference data between the latest task load index and the current task load index.

[0101] S72, determine whether the load difference data meets the preset load reduction conditions. If so, generate strategy activation confirmation data.

[0102] It needs to be explained that the load difference data If the preset load reduction condition is met (e.g., ΔL < −0.05, meaning the load reduction exceeds 0.05), then policy activation confirmation data is generated, triggering an increase in the threshold.

[0103] S73, based on the policy activation confirmation data, adjusts and updates the warning trigger threshold and intervention trigger threshold by increasing the values ​​according to the preset threshold rollback compensation value.

[0104] It should be noted that after the threshold is raised, the system's overall judgment range for the comprehensive intervention tendency value shifts upward, which means that it shows a higher tolerance for intervention tendency values ​​of the same magnitude in subsequent interactions, reducing the frequency of unnecessary intervention triggers and preventing users from feeling uncomfortable due to excessive system intervention; if the load does not decrease, the threshold remains unchanged, maintaining sensitivity to continuous pressure.

[0105] The threshold update rule is as follows: , , where Δθ is the preset threshold rollback compensation value (e.g., Δθ=0.02).

[0106] S74 overwrites and stores the updated warning trigger threshold and the updated intervention trigger threshold locally in the system.

[0107] Taking Zhang as an example: After the real-time intervention in the 7th round, in the 8th round, L(8) = 0.52 (a decrease of 0.20 from L(7) = 0.72, satisfying ΔL < −0.05), the system generated policy effectiveness confirmation data and executed. , The data is overwritten and stored locally in the system. In subsequent rounds, the new threshold is used as the standard. After Zhang's emotions stabilize, the system no longer easily triggers warnings, thus improving the naturalness of the interaction.

[0108] In addition, the target dialogue response data is generated and output, followed by an adaptive feedback update step:

[0109] Obtain the emotional stability parameter for the next interaction round and perform a difference calculation with the emotional stability parameter for the current interaction round to obtain the emotional stability difference data.

[0110] Balance gain data is calculated based on the load difference data and the emotional stability difference data.

[0111] Based on the balanced gain data, the sentiment association weight data and load association weight data in the association function are dynamically updated.

[0112] Specifically, the meanings of the above parameters are: emotional stability difference data. (The difference in emotional stability parameters across rounds is used; ΔS>0 indicates that emotions are stabilizing, while ΔS<0 indicates further emotional fluctuations); the load difference data ΔL has been defined in the preceding steps. The formula for calculating the balance gain data G is: α and β are balancing factors (initially set to α=0.5, β=0.5). The larger −ΔL (the greater the load reduction) and ΔS (the more stable the emotions), the larger G is, indicating that the overall effect of this round of intervention is better.

[0113] The weight update rule is as follows: , ,in The learning rate (e.g., γ=0.01). The relative contribution ratio to emotion improvement is represented by a higher value, indicating a better effect on emotion improvement, and the emotion-related weight w_e increases accordingly. Specific numerical demonstration: Taking Zhang's 8th round as an example: S(8) increased by 0.15 compared to S(7) (emotional stabilization), ΔS=0.15, ΔL=−0.20. , , If Zhang's emotions continue to improve in subsequent rounds, The weighting will be gradually increased, and the system will assign higher weights to the emotional change trend parameters in future correlation function calculations, forming a personalized perception-weight configuration for Zhang.

[0114] In this embodiment, this step defines an adaptive update mechanism for the emotion association weight and load association weight in the association function. By introducing balanced gain data to dynamically correct the weights, the emotion-cognition trade-off in the calculation of the comprehensive intervention tendency value can be continuously optimized with the accumulation of historical interactions, enabling the invention to achieve a key closed-loop mechanism for long-term personalized adaptation. Furthermore, the dynamic weight update mechanism allows the system to learn from the actual effects of each intervention, gradually identifying whether a specific user is more sensitive to intervention in the emotion or cognition dimension. Together with the dynamic threshold update step, this constitutes the complete adaptive feedback system for cross-session continuous optimization in this invention, realizing personalized adaptive service.

[0115] like Figure 2 The diagram shown is a structural schematic of a spoken dialogue dynamic generation system based on voice emotion feedback provided in an embodiment of this application, comprising:

[0116] The multidimensional data acquisition module is used to acquire user voice signals, semantic text data, and interaction behavior data for the current interaction round.

[0117] The feature quantization processing module is used to input the user's voice signal into a preset emotion representation model, extract features, output continuous emotion change trend parameters, and calculate the current task load index based on semantic text data and interaction behavior data according to preset rules. Based on the rate of change of the current task load index within a preset time window, the interaction load acceleration parameter is calculated. Based on the emotion change trend parameter and the interaction load acceleration parameter, a correlation function is constructed between them, and the comprehensive intervention tendency value is calculated through the correlation function.

[0118] The associated threshold determination module is used to obtain a preset dual threshold interval representing the range between the warning trigger threshold and the intervention trigger threshold. If the comprehensive intervention tendency value reaches the warning trigger threshold but does not reach the intervention trigger threshold, it determines and outputs preventive protection status data. If the comprehensive intervention tendency value reaches the intervention trigger threshold, it determines and outputs real-time intervention status data. The preventive protection status data represents the overload warning before the user's emotions fluctuate drastically.

[0119] The dynamic generation control module is used to dynamically adjust the preset dialogue generation parameters based on the preventive protection status data or real-time intervention status data output by the judgment. The dialogue generation parameters include the semantic dimensionality reduction step size set for the preventive protection status data. The adjusted dialogue generation parameters are input into the dialogue generation engine to generate and output the target dialogue response data.

[0120] The various features and processes described above can be used independently of each other or can be combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of this disclosure. Furthermore, certain method or process blocks may be omitted in some embodiments. The methods and processes described herein are not limited to any particular order, and the blocks or states associated with them may be performed in other suitable orders. For example, the described blocks or states may be performed in an order different from the order specifically disclosed, or multiple blocks or states may be combined in a single block or state. Example blocks or states may be performed serially, in parallel, or in some other manner. Blocks or states may be added to or removed from the disclosed example embodiments. The exemplary systems and components described herein may be configured differently from those described. For example, elements may be added to, removed from, or rearranged compared to the disclosed example embodiments.

[0121] The various operations of the example methods described herein can be performed at least in part by an algorithm. This algorithm can be contained in program code or instructions stored in memory (e.g., the aforementioned non-transitory computer-readable storage medium). Such an algorithm may include a machine learning algorithm. In some embodiments, the machine learning algorithm may not be explicitly programmed into the computer to perform the function, but can learn from training data to create a predictive model that performs the function.

[0122] The various operations of the example methods described herein can be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors can constitute the engine of a processor implementation that operates to perform one or more of the operations or functions described herein.

[0123] Similarly, the methods described herein can be implemented at least in part by a processor, where one or more specific processors are examples of hardware. For example, at least some operations of a method can be performed by one or more processors or an engine implemented by a processor. Furthermore, one or more processors can also be operated to support the performance of related operations in a “cloud computing” environment or as “Software as a Service” (SaaS). For example, at least some operations can be performed by a set of computers (as an example of a machine including processors), where these operations are accessible via a network (e.g., the Internet) and via one or more suitable interfaces (e.g., application programming interfaces (APIs)).

[0124] The performance of certain operations can be distributed across processors, residing not only within a single machine but also deployed across multiple machines. In some example embodiments, the processor or processor-implemented engine may reside in a single geographic location (e.g., within a home environment, office environment, or server cluster). In other example embodiments, the processor or processor-implemented engine may be distributed across multiple geographic locations.

[0125] In this specification, multiple instances may implement components, operations, or structures described as single instances. Although individual operations of one or more methods are shown and described as separate operations, one or more of the separate operations may be performed simultaneously and do not need to be performed in the order shown. Structures and functions presented as separate components in the example configuration may be implemented as composite structures or components. Similarly, structures and functions presented as single components may be implemented as separate components. These and other variations, modifications, additions, and improvements fall within the scope of this document.

[0126] While an overview of the subject matter has been described with reference to specific example embodiments, various modifications and changes can be made to these embodiments without departing from the broader scope of embodiments of this disclosure. Such embodiments of the subject matter are referred to herein, individually or collectively, by the term "invention," and are used for convenience only and are not intended to limit the scope of this application to any single disclosure or concept, should more than one disclosure or concept be disclosed in fact.

[0127] The embodiments described herein have been described in sufficient detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the detailed description should not be construed as limiting, and the scope of the various embodiments is defined only by the appended claims and the full scope of their equivalents.

Claims

1. A method for dynamically generating spoken dialogue based on voice emotion feedback, characterized in that, Includes the following steps: Acquire user voice signals, semantic text data, and interaction behavior data for the current interaction round; The user's voice signal is input into a preset emotion representation model for feature extraction, and continuous emotion change trend parameters are output. Based on the semantic text data and the interaction behavior data, the current task load index is calculated according to preset rules. The interaction load acceleration parameter is calculated based on the rate of change of the current task load index within a preset time window. Based on the emotional change trend parameter and the interaction load acceleration parameter, a correlation function is constructed between them, and a comprehensive intervention tendency value is calculated through the correlation function. Obtain a preset dual threshold interval representing the range between the warning trigger threshold and the intervention trigger threshold. If the comprehensive intervention tendency value reaches the warning trigger threshold but does not reach the intervention trigger threshold, then preventive protection status data is determined to be output. If the comprehensive intervention tendency value reaches the intervention trigger threshold, then real-time intervention status data is determined to be output. The preventive protection status data represents a load overload warning before the user's emotions fluctuate drastically. Based on the preventive protection status data or the real-time intervention status data output by the determination, the preset dialogue generation parameters are dynamically adjusted. The dialogue generation parameters include the semantic dimensionality reduction step size set for the preventive protection status data. The adjusted dialogue generation parameters are input into the dialogue generation engine to generate and output the target dialogue response data.

2. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 1, characterized in that: The specific process of inputting the user's voice signal into a preset emotion representation model for feature extraction and outputting continuous emotion change trend parameters is as follows: Acoustic feature data is extracted from the user's speech signal, including Mel frequency cepstral coefficients and fundamental frequency feature data; The acoustic feature data is mapped to a preset effective valence wake-up two-dimensional space to generate continuous coordinate data; the time derivative of the continuous coordinate data is calculated according to a preset time window to output the emotion change trend parameter, and the reciprocal of the coordinates of the continuous coordinate data within the time window is calculated to output the emotion stability parameter.

3. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 1, characterized in that: The current task load index is calculated based on the semantic text data and the interaction behavior data according to preset rules, as follows: Extract keyword encoding data and encoding frequency data from the text data; Extract the user response delay time data from the interaction data; The keyword encoding data, the encoding frequency data, and the user response delay time data are weighted and summed with the set scenario weights to output the current task load index.

4. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 1, characterized in that: The interactive load acceleration parameter is calculated based on the rate of change of the current task load index within a preset time window. The specific process is as follows: Calculate the difference between the current task load index and the historical task load index to obtain the first-order load rate data for the corresponding time node; Obtain the set of historical load rate data within the preset time window; Calculate the rate of change between the first-order load rate data and the historical load rate data set; The rate of change data is output as an interactive load acceleration parameter, which is used to characterize the urgency of the user's cognitive load increase.

5. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 1, characterized in that: The specific steps for calculating the comprehensive intervention tendency value using the correlation function are as follows: The emotion change trend parameter and the interaction load acceleration parameter are normalized respectively to obtain normalized emotion data and normalized acceleration data. Obtain preset emotion-related weight data and load-related weight data; The normalized emotion data and the emotion-related weight data are weighted and calculated, and the normalized acceleration data and the load-related weight data are weighted and calculated. The results of the two weighted calculations are then merged to output the comprehensive intervention tendency value.

6. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 1, characterized in that: The dialogue generation parameters include vocabulary difficulty indicators and speech rate control indicators; The preset dialogue generation parameters are dynamically adjusted based on the preventive protection status data or the real-time intervention status data output by the determination, specifically as follows: If the preventive protection status data is received, then the preventive intervention instructions in the preventive protection status data are extracted; According to the prevention and intervention instructions, the vocabulary difficulty index in the dialogue generation parameters is updated by decreasing the value through the semantic dimensionality reduction step size, while keeping the data of the speech rate control index unchanged, thereby generating the adjusted dialogue generation parameters with dimensionality reduction transition.

7. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 6, characterized in that: The dialogue generation parameters also include preset speech rate slowdown step size and soothing prompt label data; If the real-time intervention status data is received, the mandatory intervention command in the real-time intervention status data is extracted; According to the mandatory intervention instruction, the vocabulary difficulty index in the dialogue generation parameters is updated by decreasing the semantic dimensionality reduction step size, the speech rate control index is updated by decreasing the speech rate based on the speech rate slowing step size, and the soothing prompt label data is injected into the dialogue generation parameters to generate the downgraded and reconstructed dialogue generation parameters.

8. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 1, characterized in that: The process of generating and outputting the target dialogue response data is followed by a threshold dynamic update step: Obtain the latest task load index for the next interaction round and calculate the load difference data between the latest task load index and the current task load index; Determine whether the load difference data meets the preset load reduction conditions; if so, generate strategy activation confirmation data. Based on the policy activation confirmation data, the warning trigger threshold and the intervention trigger threshold are numerically adjusted and updated upwards according to the preset threshold rollback compensation value. The updated warning trigger threshold and the updated intervention trigger threshold are overwritten and stored locally in the system.

9. The method for dynamically generating spoken dialogue based on voice emotion feedback as described in claim 2, characterized in that: The process of generating and outputting the target dialogue response data is followed by an adaptive feedback update step: Obtain the emotional stability parameter for the next interaction round and perform a difference calculation with the emotional stability parameter for the current interaction round to obtain the emotional stability difference data; Balance gain data is calculated based on load difference data and emotional stability difference data; Based on the balanced gain data, the sentiment association weight data and load association weight data in the association function are dynamically updated.

10. A system applying the method for dynamically generating spoken dialogue based on voice emotion feedback as described in any one of claims 1-9, characterized in that, include: The multi-dimensional data acquisition module is used to acquire user voice signals, semantic text data, and interaction behavior data in the current interaction round. The feature quantization processing module is used to input the user's voice signal into a preset emotion representation model, extract features, output continuous emotion change trend parameters, and calculate the current task load index according to preset rules based on the semantic text data and the interaction behavior data; calculate the interaction load acceleration parameter based on the rate of change of the current task load index within a preset time window; construct a correlation function between the emotion change trend parameter and the interaction load acceleration parameter, and calculate the comprehensive intervention tendency value through the correlation function. The associated threshold determination module is used to obtain a preset dual threshold interval representing the range between the warning trigger threshold and the intervention trigger threshold. If the comprehensive intervention tendency value reaches the warning trigger threshold but does not reach the intervention trigger threshold, it determines and outputs preventive protection status data. If the comprehensive intervention tendency value reaches the intervention trigger threshold, it determines and outputs real-time intervention status data. The preventive protection status data represents a load overload warning before the user's emotions fluctuate drastically. The dynamic generation control module is used to dynamically adjust preset dialogue generation parameters based on the preventive protection status data or the real-time intervention status data output by the judgment. The dialogue generation parameters include a semantic dimensionality reduction step size set for the preventive protection status data. The adjusted dialogue generation parameters are input into the dialogue generation engine to generate and output target dialogue response data.

Images