A cloud-edge collaborative digital human sharing management method and system

Abstract

The application discloses a cloud-edge collaborative digital person sharing management method and system, and relates to the technical field of digital person interaction. The method comprises the following steps: an edge node collects voice emotion intensity time series data in a target user and a digital person interaction process in real time, extracts a first change slope of an emotion rising stage and a second change slope of a falling stage; a dynamic asymmetric coefficient and an emotion duration are calculated and uploaded to the cloud; the cloud outputs an emotion response speed matching factor according to historical records; the edge node calculates an accelerated transition duration of a digital person expression animation and a pitch dynamic range compression coefficient of voice synthesis; a digital person emotion output control sequence is generated and executed, and the difference between actual user emotion changes and expected changes after execution is returned to the cloud to update the historical records. Through the division of labor between the edge node and the cloud, the application solves the problem that the prior art relies on preset scripted responses, leading to a stereotyped and mechanical interactive experience and a lack of personalized emotional resonance.

Description

Technical Field

[0001] This invention relates to the field of digital human interaction technology, specifically to a cloud-edge collaborative digital human sharing management method and system. Background Technology

[0002] A digital human is a virtual avatar generated through 3D modeling and real-time rendering technology, capable of engaging in voice dialogue and emotional interaction with users on platforms such as mobile terminals and smart devices. Traditional digital human interaction methods primarily rely on pre-loading fixed 3D animations and scripted voice responses on the local device, triggering preset emotional responses based on keywords input by the user.

[0003] This pre-scripted response method has significant shortcomings in terms of interactive experience. First, users' genuine emotional expressions are highly personalized and dynamically changing; a fixed script cannot match the emotional fluctuations and reaction speeds of different users or the same user in different situations. Second, the natural process of user emotions from arousal to dissipation exhibits an asymmetrical characteristic: the intensity of emotion rises rapidly while its decline is slow. The pre-script cannot adaptively adjust the digital human's response rhythm according to this asymmetry, resulting in a disconnect between the digital human's reactions and the user's emotional state over time. Users feel that the digital human's interaction does not match their real psychological changes in emotional expression, lacking a deep emotional resonance.

[0004] Furthermore, existing technologies have failed to achieve the sharing of management functions through cloud-edge collaboration. In other words, the cloud cannot accumulate and reuse users' historical emotional response experience, and edge nodes cannot dynamically adjust the output parameters of digital humans based on cloud-shared knowledge, resulting in a lack of continuous optimization capabilities from historical feedback for each interaction. Summary of the Invention

[0005] This invention provides a cloud-edge collaborative digital human sharing management method and system to solve the technical problems in the prior art where digital human interaction relies on preset scripted responses, resulting in a rigid and mechanical interactive experience, a lack of personalized emotional resonance, and an ineffective sharing of management functions between cloud and edge nodes.

[0006] In a first aspect, the present invention provides a cloud-edge collaborative digital human sharing management method, the method comprising: The system collects real-time temporal data of voice emotion intensity during the interaction between the target user and the digital human through edge nodes, and extracts the first slope of the target user's emotion rise phase and the second slope of the emotion fall phase from the voice emotion intensity temporal data. Based on the first and second slopes of change, the dynamic asymmetry coefficient of the target user's emotion is calculated, and the duration of emotion before the current emotional peak is obtained. The dynamic asymmetry coefficient and the duration of emotion are uploaded to the cloud. The cloud outputs the emotion response speed matching factor for the target user based on the best response delay data corresponding to different durations of emotion in the historical emotion response records shared and managed in the cloud. Receive the emotional response speed matching factor, and combine it with the emotional intensity value at the current emotional peak moment to calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. Based on the accelerated transition duration and the dynamic range compression coefficient of tone, a digital human emotion output control sequence is generated and sent to the digital human terminal for execution. At the same time, the difference between the actual change and the expected change in the target user's emotion intensity after execution is sent back to the cloud to update the historical emotion response records shared and managed in the cloud.

[0007] Secondly, the present invention also provides a cloud-edge collaborative digital human sharing management system, the system comprising: The voice emotion acquisition and analysis module is used to collect real-time time-series data of voice emotion intensity during the interaction between the target user and the digital human through edge nodes, and to extract the first change slope of the target user's emotion rising phase and the second change slope of the emotion falling phase from the voice emotion intensity time-series data. The emotion dynamic analysis module is used to calculate the dynamic asymmetry coefficient of the target user's emotion based on the first change slope and the second change slope, and to obtain the duration of the emotion before the current emotional peak moment. The cloud-based sharing management module is used to upload the dynamic asymmetry coefficient and the duration of emotion to the cloud. The cloud outputs an emotion response speed matching factor for the target user based on the best response delay data corresponding to different durations of emotion in the historical emotion response records of the target user's cloud-based sharing management. The output parameter calculation module is used to receive the emotional response speed matching factor and, in combination with the emotional intensity value at the current emotional peak moment, calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. The emotion output control module is used to generate a digital human emotion output control sequence based on the accelerated transition duration and the dynamic range compression coefficient of the tone, and send it to the digital human terminal for execution. At the same time, it sends the difference between the actual change and the expected change in the target user's emotion intensity after execution back to the cloud and updates the historical emotion response records shared and managed in the cloud.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: First, this invention quantifies the asymmetric characteristics of user emotional changes by calculating the dynamic asymmetry coefficient between the upward and downward slopes of user emotions. When the asymmetry in user emotional changes is strong, the digital human's response speed increases accordingly to match the user's emotional rhythm; when the asymmetry is weak, the response speed decreases accordingly to maintain a smooth interaction. Compared with existing technologies that use a fixed response delay, the digital human's emotional response can dynamically adjust according to the user's real-time emotional changes, achieving personalized matching of the user's emotional state.

[0009] Secondly, this invention uploads the dynamic asymmetry coefficient and emotional duration to the cloud. Based on the target user's historical emotional response records shared and managed in the cloud, it outputs a personalized emotional response speed matching factor for that user. Compared to existing technologies where all users share the same set of interaction parameters and cannot accumulate individual experience, the cloud-shared historical records can continuously accumulate the user's best response parameters in different emotional scenarios. As the number of interactions increases, the matching factor gradually approaches the user's optimal response speed, making the digital human's emotional expression more closely match the user's personalized needs.

[0010] Third, after the digital human terminal performs emotional output, this invention sends the difference between the actual change and the expected change in the user's emotional intensity back to the cloud to update the shared historical records, forming a closed loop from perception, decision-making, execution to feedback. Compared with existing technologies where there is no feedback evaluation after digital human output and continuous optimization is impossible, this invention continuously updates the historical emotional response records shared in the cloud through closed-loop feedback, allowing subsequent matching factor calculations to refer to the latest response effects, thus realizing online iterative optimization of the digital human's emotional interaction strategy. Attached Figure Description

[0011] 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.

[0012] Figure 1 This is a flowchart illustrating a cloud-edge collaborative digital human sharing management method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a cloud-edge collaborative digital human sharing management method provided in an embodiment of the present invention; Figure 3 This is a data table diagram of effective voice segments and voice frames in a cloud-edge collaborative digital human sharing management method provided by an embodiment of the present invention; Figure 4This is a schematic diagram of the structure of a cloud-edge collaborative digital human sharing management system provided in an embodiment of the present invention; The diagram is labeled as follows: Voice Emotion Acquisition and Analysis Module 11, Emotion Dynamic Analysis Module 12, Cloud Sharing Management Module 13, Output Parameter Calculation Module 14, and Emotion Output Control Module 15. Detailed Implementation

[0013] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0014] Example 1, as Figure 1 As shown, this invention provides a flowchart illustrating a cloud-edge collaborative digital human sharing management method; as Figure 2 As shown, this invention provides a scheme logic diagram for a cloud-edge collaborative digital human sharing management method, the method comprising: The shared management described in this invention refers to the cloud and edge nodes jointly undertaking the management tasks of digital human emotional interaction. That is, the edge nodes perform real-time perception and parameter calculation, while the cloud performs historical storage and personalized matching. The two share management functions and make collaborative decisions.

[0015] S100: Real-time acquisition of voice emotion intensity time-series data during the interaction between the target user and the digital human via edge nodes, and extraction of the first change slope of the target user's emotion rising phase and the second change slope of the emotion falling phase from the voice emotion intensity time-series data. During voice interaction between a digital human and a target user, the user's voice signal contains rich emotional information. When the user is emotionally agitated, the loudness and pitch of the voice increase significantly; when the emotion subsides, the loudness and pitch gradually decrease. This change in emotional intensity exhibits a single-peak characteristic of rising and then falling over time, and the rate of rise and fall varies among different users and for different emotional types. In traditional solutions, the digital human's emotional response uses fixed scripted parameters. Regardless of how quickly the user's current emotion is aroused or how slowly it fades, the duration of the digital human's facial expression animation transition and the amplitude of pitch changes in the synthesized speech remain unchanged, causing the digital human's reaction to be disconnected from the user's actual emotional state over time.

[0016] Step S100 provided in this embodiment of the invention includes: Voice activity detection is performed on the target user's voice signal through edge nodes, and the detected valid voice segments are sampled at fixed time intervals to obtain voice frames at multiple consecutive sampling times. Extract the short-time energy value and fundamental frequency value of the speech frame at each sampling time; Based on the upper limit of the dynamic range of the audio acquisition at the edge node, the short-time energy value is normalized to a normalized energy value, wherein the upper limit of the dynamic range is the short-time energy value corresponding to the full-scale input amplitude of the audio analog-to-digital converter. Divide the fundamental frequency value by the mean fundamental frequency of the target user in the unemotional speech segment to obtain the normalized fundamental frequency, wherein the unemotional speech segment is the speech segment corresponding to the no-task confirmation speech issued by the target user according to the system prompt during the interaction initialization phase; The normalized energy and the normalized fundamental frequency are weighted and summed to obtain the emotional intensity value at the corresponding sampling time. The emotional intensity values ​​from multiple consecutive sampling times are arranged in chronological order and then subjected to median filtering to obtain the temporal data of the speech emotional intensity.

[0017] The specific implementation method is as follows: Voice activity detection is performed on the target user's speech signal using edge nodes. Voice activity detection distinguishes between speech segments and silence segments by analyzing the energy and zero-crossing rate of the audio signal. Continuous audio signals are processed in frames, for example, each frame is 20 milliseconds long with a 10-millisecond frame shift. For each frame, a short-time energy value and a zero-crossing rate are calculated. The short-time energy value is the mean of the sum of the squares of the amplitudes of all sampled points within the frame, and the zero-crossing rate is the ratio of the number of sign changes between adjacent sampled points to the total number of sampled points within the frame. When both the short-time energy value and the zero-crossing rate exceed a preset energy threshold and a preset zero-crossing rate threshold, the frame is determined to be a speech frame.

[0018] The preset energy threshold and preset zero-crossing rate threshold are adaptively set based on the background noise level of the environment in which the edge node is located. During the interaction initialization phase, the edge node acquires a one-second environmental silence signal. This silence signal is then divided into frames with a 20-millisecond frame length and a 10-millisecond frame shift. The short-time energy value and zero-crossing rate are calculated frame by frame. The arithmetic mean of the short-time energy values ​​of all frames is taken as the environmental energy baseline, and the arithmetic mean of the zero-crossing rates of all frames is taken as the environmental zero-crossing rate baseline. The preset energy threshold can be set to three times the environmental energy baseline. This setting classifies speech frames whose short-time energy values ​​exceed three times the environmental energy baseline as speech frames. Since the energy of environmental noise typically fluctuates slightly around the baseline value, and the short-time energy value of speech signals is significantly higher than that of environmental noise, setting it to three times effectively distinguishes speech from noise. The preset zero-crossing rate threshold can be set to twice the environmental zero-crossing rate baseline. This setting classifies speech frames whose zero-crossing rate exceeds twice the environmental zero-crossing rate baseline as speech frames. The zero-crossing rate of ambient noise typically fluctuates slightly around the baseline value. Human speech, due to the quasi-periodicity of vocal cord vibration, significantly increases the zero-crossing rate; doubling this rate effectively distinguishes speech from noise. When using two thresholds for joint discrimination, a frame is only identified as speech if both the short-time energy value and the zero-crossing rate exceed their respective thresholds. When the short-time energy values ​​of multiple consecutive frames are all below the energy threshold, a silence segment is entered. Consecutive speech frames are then merged into a valid speech segment.

[0019] Detected valid speech segments are sampled at fixed time intervals to obtain speech frames from multiple consecutive sampling times. The fixed time interval is set based on the smallest resolvable time scale of human emotional changes. In natural human interactions, significant changes in emotional intensity typically require hundreds of milliseconds or more. Too short a sampling interval introduces unnecessary computational overhead, while too long an interval may lose detailed features of emotional changes. For example, a fixed time interval of 100 milliseconds means that 10 speech frames are uniformly collected per second. Each speech frame has a duration of 20 milliseconds.

[0020] The short-time energy value and fundamental frequency value of the speech frame at each sampling moment are extracted, and the short-time energy value is normalized to normalized energy. Based on the upper limit of the dynamic range of the edge node audio acquisition, the short-time energy value is divided by this upper limit; the resulting ratio is the normalized energy. The upper limit of the dynamic range is the short-time energy value corresponding to the full-scale input amplitude of the audio analog-to-digital converter (AADC). The AADC converts the analog voltage signal acquired by the microphone into a digital sample value; its full-scale input amplitude is the maximum voltage peak value that can be converted without distortion, corresponding to a defined maximum short-time energy value. This maximum short-time energy value is used as the upper limit of the dynamic range. The normalized energy value ranges from 0 to 1; the closer the value is to 1, the closer the energy of the current speech frame is to the maximum dynamic range of the acquisition hardware, and the louder the user's voice.

[0021] The fundamental frequency value is normalized to a normalized fundamental frequency. The normalized fundamental frequency is obtained by dividing the fundamental frequency value by the mean fundamental frequency of the target user in the unemotional speech segment. The unemotional speech segment corresponds to the speech segment without task meaning issued by the target user according to system prompts during the interaction initialization phase, such as the short confirmation speech "okay" or "successful" issued by the user to confirm a normal device connection. This speech segment is issued under conditions without emotional stimulation, and its fundamental frequency reflects the user's pitch reference in a calm state. This speech segment is collected during the interaction initialization phase, and the arithmetic mean of the fundamental frequency values ​​of all speech frames within this segment is calculated as the user's mean fundamental frequency and stored locally on the edge node. The normalized fundamental frequency value is centered at 1; a value greater than 1 indicates that the fundamental frequency of the current speech frame is higher than the user's calm reference, i.e., the pitch is higher; a value less than 1 indicates that the fundamental frequency is lower than the calm reference, i.e., the pitch is lower.

[0022] The normalized energy and normalized fundamental frequency are weighted and summed to obtain the emotional intensity value at the corresponding sampling time. The weighting coefficients for the weighted summation are obtained through offline calibration. The offline calibration method is as follows: voice samples from multiple users under controlled emotional arousal conditions are collected, and the normalized energy and normalized fundamental frequency are extracted for each sample. The actual emotional intensity level at each sampling time is then manually labeled. Using normalized energy and normalized fundamental frequency as independent variables and the manually labeled emotional intensity level as the dependent variable, the optimal combination of weighting coefficients is determined through least squares linear regression. The sum of the energy weighting coefficient and the fundamental frequency weighting coefficient obtained from offline calibration is 1. The emotional intensity value ranges from 0 to 1, with a value closer to 1 indicating a higher emotional intensity at that sampling time.

[0023] The emotional intensity values ​​from multiple consecutive sampling times are arranged chronologically and then subjected to median filtering to obtain temporal data of speech emotional intensity. Median filtering uses each sampling point as the center, taking the emotional intensity values ​​of that sampling point and several adjacent sampling points to form a filtering window. The values ​​within the filtering window are sorted by size, and the median value is used to replace the original value of that sampling point. The purpose of median filtering is to eliminate isolated outliers caused by transient noise or short-term non-stationarity of the speech signal, while preserving the overall trend and peak-valley transition characteristics of emotional intensity changes. The width of the filtering window is set to 5 sampling points, and the window width is related to the sampling frequency and the time scale of emotional changes. With a fixed sampling interval of 100 milliseconds, the 5-sampling-point window covers a time span of 500 milliseconds. In natural human interaction, significant changes in emotional intensity typically require hundreds of milliseconds or more; a 500-millisecond window width is comparable to the smallest resolvable time scale of emotional changes. Within this time span, emotional intensity does not undergo drastic changes, and taking the median effectively removes isolated noise spikes while preserving the overall trend and peak-valley transition characteristics of emotional intensity changes.

[0024] The extraction of short-time energy and fundamental frequency values ​​of the speech frame at each sampling time includes: Calculate the sum of squared amplitudes of all sampling points within the speech frame, and divide the sum of squared amplitudes by the total number of sampling points in the speech frame to obtain the short-time energy value; Calculate the arithmetic mean of the amplitudes of all sampling points within the speech frame, and subtract the arithmetic mean from the amplitude of each sampling point to obtain a zero-mean corrected sampling point sequence. Arrange the corrected sampling point sequence in chronological order, and calculate the autocorrelation coefficient corresponding to each delay point. The delay point is an integer between the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period. The autocorrelation coefficient is the inner product of the corrected sampling point sequence and the sequence after shifting itself backward by the delay point. Divide the autocorrelation coefficient corresponding to each delay point by the autocorrelation coefficient of the modified sampling point sequence at zero delay to obtain the normalized autocorrelation coefficient. Arrange the normalized autocorrelation coefficients corresponding to all non-zero delay points in ascending order of delay points to form an autocorrelation function sequence; Within the search interval defined by the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period, the global maximum peak position of the autocorrelation function sequence is searched, and the sampling rate is divided by the number of delay points corresponding to the global maximum peak position to obtain the fundamental frequency value.

[0025] The specific implementation method is as follows: The short-time energy value is calculated by summing the squares of the amplitudes of all sampled points within a speech frame and dividing this sum by the total number of sampled points in the speech frame. The amplitude of a sampled point refers to the signed integer value output by the analog-to-digital converter (ADC) at the sampling moment, which can be directly read from the ADC's data register. This value represents the digital quantity of the sound pressure level signal received by the microphone at that moment after amplification and quantization. The short-time energy value reflects the average power level of the speech signal within the speech frame, and its unit is the dimensionless mean of the sum of squares. The greater the loudness of the speech signal, the greater the short-time energy value.

[0026] Calculate the arithmetic mean of the amplitudes of all sampling points within the speech frame. Subtract this arithmetic mean from the amplitude of each sampling point to obtain a zero-mean corrected sampling point sequence. The purpose of zero-mean is to remove the DC component from the speech signal, making the corrected sampling point sequence symmetrically distributed around the zero point, thus eliminating the influence of DC bias on subsequent autocorrelation calculations.

[0027] The corrected sampling point sequence is arranged in chronological order, and the autocorrelation coefficient corresponding to each delay point is calculated. The delay point number is an integer between the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period. The highest possible fundamental frequency of human speech is approximately 400Hz, corresponding to a minimum possible pitch period of 2.5 milliseconds; the lowest possible fundamental frequency of human speech is approximately 80Hz, corresponding to a maximum possible pitch period of 12.5 milliseconds. When the sampling rate of the analog-to-digital converter is 16000Hz, the number of sampling points corresponding to the minimum possible pitch period is 40, and the number of sampling points corresponding to the maximum possible pitch period is 200. Therefore, the search range for the delay point number is an integer between 40 and 200. The autocorrelation coefficient is calculated as follows: starting from the first sampling point of the corrected sampling point sequence, up to the sequence length minus the number of delay points, multiply the amplitude of each sampling point by the amplitude of the corresponding sampling point after shifting the corrected sampling point sequence by the number of delay points. Then, sum all the products. The sum is the autocorrelation coefficient corresponding to that delay point. The autocorrelation coefficient is the inner product of the corrected sampling point sequence and its own sequence after shifting the delay point.

[0028] The normalized autocorrelation coefficient is obtained by dividing the autocorrelation coefficient corresponding to each delay point by the autocorrelation coefficient of the corrected sampling point sequence at zero delay. At zero delay, the sequence completely overlaps with itself, resulting in the largest autocorrelation coefficient. The normalized autocorrelation coefficient ranges from zero to one. The normalized autocorrelation coefficients corresponding to all non-zero delay points are arranged in ascending order of delay point number to form an autocorrelation function sequence.

[0029] Within a search interval defined by the number of sampling points corresponding to the minimum and maximum possible pitch periods, the global maximum peak position of the autocorrelation function sequence is searched. The autocorrelation function sequence exhibits a local peak at the delay point corresponding to the pitch period; this peak reflects the periodicity of the speech signal. The delay point corresponding to the position with the highest autocorrelation function value within the search interval is taken. The sampling rate is divided by the number of delay points corresponding to this global maximum peak position; the quotient is the fundamental frequency. For example, when the sampling rate is 16000Hz and the number of delay points corresponding to the global maximum peak position is 100, the fundamental frequency is 16000 ÷ 100 = 160Hz. The fundamental frequency reflects the vibration frequency of the speaker's vocal cords; when emotionally agitated, the vocal cord tension increases, and the fundamental frequency rises.

[0030] Step S100 provided in this embodiment of the invention further includes: The arithmetic mean of multiple consecutive emotion intensity values ​​in the speech emotion intensity time series data is calculated, and the result of the arithmetic mean is used as the smoothed intensity value at the corresponding center time to obtain the smoothed intensity sequence. Calculate the difference in emotional intensity values ​​between adjacent sampling times in the smoothed intensity sequence, and mark the sampling time corresponding to the turning point where the difference changes from a positive value to a negative value as the emotional peak time; For each emotional peak moment, in the smooth intensity sequence, the first turning point where the difference turns from negative to positive is retrieved forward, the sampling time corresponding to the first turning point is marked as the starting point of the rise, and the time period between the starting point of the rise and the emotional peak moment is determined as the emotional rise phase. The first turning point where the difference turns from negative to positive is retrieved backward, and the sampling time corresponding to the first turning point is marked as the end point of the decline. The time period between the emotional peak and the end point of the decline is defined as the emotional decline phase. Based on the emotional intensity value at each sampling moment during the emotional rise phase and the corresponding time, a linear fit is performed, and the slope obtained from the fit is used as the first change slope. Based on the linear fitting of the emotional intensity value at each sampling moment during the emotional decline phase with the corresponding time, the slope obtained by fitting is used as the second change slope.

[0031] The specific implementation method is as follows: First, an arithmetic mean is calculated for multiple consecutive emotion intensity values ​​in the time-series data of speech emotion intensity. The result of this arithmetic mean is used as the smoothed intensity value at the corresponding center moment, resulting in a smoothed intensity sequence. The width of the smoothing window can be 5 sampling points, corresponding to a time span of 500 milliseconds, matching the time scale of human emotion changes. Smoothing assigns equal weights to each sampling point within the window through arithmetic averaging, homogenizing local fluctuations and making the main trend of emotion changes clearer and more discernible, providing a smoother input for subsequent inflection point detection and slope fitting. If the window is too small, the smoothing effect is insufficient, and subsequent difference sequences may produce false inflection points due to local fluctuations; if the window is too large, over-smoothing may obscure the true location of emotion peaks and troughs, leading to positioning errors at peak moments. The purpose of smoothing is to eliminate short-term random fluctuations in the time-series data of emotion intensity, making the main trend of emotion changes clearer and more discernible, and providing a more reliable input for subsequent inflection point detection and slope fitting.

[0032] Then, the difference in emotional intensity values ​​between adjacent sampling times in the smoothed intensity sequence is calculated, and the sampling time corresponding to the inflection point where the difference changes from positive to negative is marked as the emotional peak time. A positive difference indicates that the emotional intensity is increasing, and a negative difference indicates that the emotional intensity is decreasing. When the difference changes from positive to negative, it indicates that the emotional intensity has undergone a transition from increasing to decreasing, and this transition point corresponds to a local maximum of emotional intensity in time.

[0033] For each emotional peak moment, the first inflection point where the difference changes from negative to positive is retrieved forward in the smoothed intensity sequence. The sampling moment corresponding to this inflection point is marked as the starting point of the rise, and the period from the starting point of the rise to the emotional peak moment is defined as the emotional rise phase. The first inflection point where the difference changes from negative to positive is retrieved backward, and the sampling moment corresponding to this inflection point is marked as the ending point of the fall. The period from the emotional peak moment to the ending point of the fall is defined as the emotional fall phase. If no inflection point where the difference changes from negative to positive is found before the emotional peak moment, the first sampling moment of the continuous rising sequence containing the emotional peak moment is taken as the starting point of the rise. If no inflection point where the difference changes from negative to positive is found after the emotional peak moment, the last sampling moment of the continuous falling sequence containing the emotional peak moment is taken as the ending point of the fall.

[0034] Furthermore, a linear fit is performed between the emotional intensity values ​​at each sampling moment during the emotional escalation phase and the corresponding time, and the resulting slope is used as the first slope of change. Linear fitting is a mathematical method that aims to find a straight line that minimizes the sum of the squares of the vertical distances between the line and each sampling point. The unit of the first slope of change is emotional intensity values ​​per second, with a positive value indicating that the emotional intensity increases over time. The larger the first slope of change, the faster the emotional arousal, and the more rapidly the user's emotions rise in a short period of time.

[0035] A linear fit is performed between the emotional intensity values ​​at each sampling point during the emotional decline phase and the corresponding time, and the resulting slope is used as the second slope of change. The second slope of change is usually negative, indicating that the emotional intensity decreases over time. The larger the absolute value of the second slope of change, the faster the emotional decline, and the user quickly recovers to calm after the emotional peak; the smaller the absolute value, the slower the emotional decline, and the user maintains a high emotional level for a longer period of time after the emotional peak.

[0036] For example, Figure 3 The table diagram of sample audio frames for effective audio segments provided in this embodiment of the invention shows the original data and processing results of audio frames at each sampling time within the effective audio segment for a target user in a single emotional event.

[0037] The following technical effects were achieved through this step: First, this step collects real-time temporal data of users' voice emotional intensity at the edge nodes. Loudness and pitch features are normalized and weighted to transform them into a multi-dimensional comprehensive quantitative index of emotional intensity. The slope of the user's emotional rise and fall phases is then extracted through smoothing and linear fitting. Compared to using only a single voice feature or discrete emotion classification labels as input, the slope quantifies the dynamic speed and asymmetry of user emotional changes, providing dynamic temporal features for personalized matching and response parameter generation in subsequent cloud-based shared management.

[0038] Second, feature extraction is completed at the edge and the extracted lightweight slope parameters are uploaded to the cloud. The amount of communication data is much smaller than that of uploading the original audio stream, realizing a collaborative division of labor between edge feature extraction and cloud-based sharing and management.

[0039] S200: Calculate the dynamic asymmetry coefficient of the target user's emotion based on the first and second slopes of change, and obtain the duration of emotion before the current emotional peak moment; In genuine emotional expression, the rise and fall of user emotions typically exhibit asymmetry: emotional arousal is often rapid, with a steep initial slope and a short rise time; emotional decline, on the other hand, is relatively slow, with a smaller absolute value of the second change slope and a longer fall time. This asymmetry reflects the individual characteristics of user emotional changes. If a digital human's emotional response cannot perceive this asymmetry, a situation may arise where the user's emotional peak has begun to subside while the digital human continues to intensify its facial expressions, resulting in a disconnect in the interaction.

[0040] Step S200 provided in this embodiment of the invention includes: Take the absolute value of the second slope of change, and use the absolute value as the descent rate value; Add the value of the first slope change to the value of the descent rate to obtain the sum of the absolute values ​​of the slopes, and subtract the value of the first slope change from the value of the descent rate to obtain the difference in the absolute values ​​of the slopes. The dynamic asymmetry coefficient is determined based on the ratio of the difference of the absolute values ​​of the slopes to the sum of the absolute values ​​of the slopes. Read the sampling time index corresponding to the current emotional peak moment and the sampling time index corresponding to the rising starting point, and calculate the index difference between the two sampling time indices. The sampling time index is a sequence number that is sequentially incremented according to the time order of consecutive sampling moments. Multiply the index difference by a fixed time interval to obtain the duration of the emotion.

[0041] The specific implementation method is as follows: First, the absolute value of the second slope obtained in step S100 is taken as the rate of decline. The second slope is typically negative during the emotional decline phase; taking its absolute value converts it to a positive rate of decline, representing the rate of decrease in emotional intensity per unit time, expressed in emotional intensity values ​​per second. A larger rate of decline indicates a faster decline in emotional intensity. The first slope is then added to the rate of decline, resulting in the sum of the absolute values ​​of the slopes. Since both the first slope and the rate of decline are positive, their sum represents the total rate of change in emotional intensity during both the rising and falling phases.

[0042] Then, the value of the first slope of change is subtracted from the value of the rate of decline to obtain the difference in the absolute values ​​of the slopes. Since the first slope of change represents the rate of emotional ascent and the rate of decline represents the rate of emotional descent, the difference reflects the degree of difference between the rates of ascent and decline. A positive difference indicates that the rate of ascent is greater than the rate of decline, meaning that emotional arousal is faster than its decay; a negative difference indicates that the rate of decline is greater than the rate of ascent, meaning that emotional decay is faster than its arousal; and a zero difference indicates that the rate of ascent equals the rate of decline, meaning that the rates of emotional arousal and decay are symmetrical.

[0043] The dynamic asymmetry coefficient is determined by the ratio of the difference in the absolute values ​​of the slopes to the sum of the absolute values ​​of the slopes. The dynamic asymmetry coefficient = difference in absolute values ​​of slopes ÷ sum of absolute values ​​of slopes. This formula normalizes the difference between the rate of increase and the rate of decrease by dividing by their sum, resulting in a dimensionless coefficient ranging from -1 to 1. A positive value indicates that the rate of emotional arousal is faster than the rate of emotional decay; the larger the absolute value, the stronger the asymmetry, meaning the more pronounced the characteristic of rapid arousal followed by slow decay. A negative value indicates that the rate of emotional decay is faster than the rate of arousal, meaning the user's emotion accumulates slowly and then dissipates rapidly. A value of zero indicates that the rates of emotional arousal and decay are completely symmetrical. The dynamic asymmetry coefficient quantifies the degree of imbalance in the rates of increase and decrease during the two phases of user emotional change.

[0044] Further, the sampling time index corresponding to the current emotional peak moment and the sampling time index corresponding to the starting point of the rise are read. The sampling time index is a sequence number that increments sequentially for consecutive sampling moments according to time order, starting from 0 for the first sampling and increasing by 1 for each sampling. The starting point of the rise is obtained by searching backward in step S100 and is the first sampling moment of the emotional rise phase. The index difference between the two sampling time indices is calculated, that is, the index of the current emotional peak moment minus the index of the starting point of the rise. The difference represents the number of sampling cycles that the emotion goes through from the starting point of the arousal to the peak.

[0045] Multiplying the index difference by a fixed time interval yields the emotion duration. The fixed time interval is 100 milliseconds, as set in the previous steps, representing the time interval between each sample. The unit of emotion duration is seconds or milliseconds, indicating the length of time it takes for a user's emotion to reach its peak. When the fixed time interval is 100 milliseconds, the emotion duration = index difference × 0.1 seconds. For example, when the index difference is 15, the emotion duration is 1.5 seconds. Emotion duration reflects the speed at which a user accumulates emotion in the current emotional event and is directly related to the length of the corresponding emotional escalation phase.

[0046] For example, in a certain emotional event obtained in step S100, the first change slope is 0.42, and the second change slope is -0.18. The rate of decline = |-0.18| = 0.18. The sum of the absolute values ​​of the slopes = 0.42 + 0.18 = 0.60. The difference in the absolute values ​​of the slopes = 0.42 - 0.18 = 0.24. The dynamic asymmetry coefficient = 0.24 ÷ 0.60 = 0.40. This coefficient is positive, indicating that the user's emotional arousal speed is significantly faster than its dissipation speed, exhibiting an asymmetric characteristic of rapid emotional arousal followed by slow dissipation. The sampling time index of the rising starting point is 30, and the sampling time index of the emotional peak is 45, with an index difference of 45 - 30 = 15. The fixed time interval is 100 milliseconds, and the emotional duration = 15 × 0.1 = 1.5 seconds.

[0047] The following technical effects were achieved through this step: By calculating the ratio between the absolute values ​​of the first and second slopes, the difference between the speed of emotional arousal and dissipation is quantified as a dynamic asymmetry coefficient. This coefficient captures the relative speed of emotional arousal and dissipation, enabling the digital human to distinguish between two different emotional dynamic patterns: a rapid emotional outburst followed by slow dissipation, and a slow emotional accumulation followed by rapid dissipation. Simultaneously, the emotional duration quantifies the time span from arousal to peak emotion, forming, together with the dynamic asymmetry coefficient, two independent dimensions characterizing the dynamic features of user emotions.

[0048] S300: Upload the dynamic asymmetry coefficient and the duration of emotion to the cloud. The cloud outputs an emotion response speed matching factor for the target user based on the best response delay data corresponding to different durations of emotion in the historical emotion response records shared and managed by the target user in the cloud. The dynamic asymmetry coefficient and emotional duration quantify the speed asymmetry between the arousal and decay of user emotions and the time span from emotional arousal to peak, respectively. Different users have different expectations regarding the speed of the digital human's emotional response: some users prefer a rapid response from the digital human after the emotional peak, while others prefer an appropriate reaction interval. Applying the same response delay to all users would fail to match the personalized interaction rhythms of each user. Shared and managed historical emotional response records are stored in the cloud, containing feedback-verified optimal response delay data for each user under different emotional durations.

[0049] Step S300 provided in this embodiment of the invention includes: The cloud receives the dynamic asymmetry coefficient and the duration of the emotion uploaded by the edge node; All records in the historical emotional response records shared and managed in the cloud are sorted in ascending order according to the absolute value of the difference between the emotional duration field and the received emotional duration, and the first half of the records are taken as the reference record set. Extract the best response delay data stored in each record in the reference record set, wherein the best response delay data is the value of the current response delay field corresponding to the record in the historical emotional response records that is closest to the duration of the emotion and has the smallest response deviation; Calculate the arithmetic mean of all extracted best response delay data to obtain the baseline response delay; An emotional response speed matching factor is determined based on the baseline response delay and the dynamic asymmetry coefficient, wherein the emotional response speed matching factor is negatively correlated with the baseline response delay and positively correlated with the dynamic asymmetry coefficient.

[0050] The specific implementation method is as follows: The cloud server receives the dynamic asymmetry coefficient and sentiment duration uploaded by edge nodes. The edge nodes then transmit these two parameters in a structured data format to the cloud server via the communication network. The data packet contains the target user's unique identifier, the session identifier of the current interaction session, the value of the dynamic asymmetry coefficient, and the value of the sentiment duration. Upon receiving the data, the cloud server uses the target user's identifier as an index to load the corresponding shared historical sentiment response record from the cloud database.

[0051] Each record in the historical emotion response log represents the processing result of a particular emotion event in a user's past interactions. Each record contains multiple fields, among which the emotion duration field records the length of time from the trigger to the peak of the emotion event. The received dynamic asymmetry coefficient and emotion duration are real-time parameters extracted by the edge node in the current interaction session. The cloud needs to find the scenario with the closest emotion duration to the current one from the historical records in order to utilize the response parameters validated by historical feedback.

[0052] All historical sentiment response records of the target user are sorted in ascending order based on the absolute value of the difference between the sentiment duration field and the received sentiment duration. The absolute value of the difference indicates how close the sentiment duration in the historical record is to the current sentiment duration; the smaller the difference, the closer the two are. After ascending sorting, the first half of the records are selected as the reference record set. The reason for selecting the first half, rather than a fixed number, is that when the number of historical records is limited, selecting the first half can dynamically adapt to changes in the total number of records. When the total number of records is large, the first half still ensures a sufficient amount of historical sample references; when the total number of records is small, the first half still provides a certain number of reference samples. Selecting the first half strikes a balance between sample sufficiency and screening rigor.

[0053] Extract the best response delay data stored in each record of the reference record set. Each historical sentiment response record stores a best response delay data field, which records the best response delay verified by user feedback for that historical sentiment event, i.e., the response delay corresponding to the sentiment response speed matching factor used at that time. This current response delay is a valid response delay parameter verified by actual interaction feedback.

[0054] The baseline response delay is obtained by calculating the arithmetic mean of all extracted optimal response delay data. The baseline response delay is expressed in seconds and represents the average response delay level preferred by the target user in historical interactions under conditions of similar emotional duration. The baseline response delay is a statistical estimate of the target user's personalized response rhythm.

[0055] The emotional response speed matching factor is determined based on the baseline response delay and the dynamic asymmetry coefficient. The emotional response speed matching factor is negatively correlated with the baseline response delay and positively correlated with the dynamic asymmetry coefficient. This means that a larger baseline response delay indicates that the target user historically prefers a longer response delay, and the digital human should not respond too quickly, thus decreasing the emotional response speed matching factor. Conversely, a larger dynamic asymmetry coefficient indicates that the target user's emotional arousal speed is significantly faster than its dissipation speed, resulting in a rapid emotional escalation. The digital human should respond faster to match the user's emotional rhythm, thus increasing the emotional response speed matching factor. The specific formula for calculating the emotional response speed matching factor is: Emotional Response Speed ​​Matching Factor = (1 + Dynamic Asymmetry Coefficient) ÷ (Baseline Response Delay + Normal Number). The normal number in the denominator is a preset positive value to prevent the matching factor from being abnormally amplified due to an excessively small baseline response delay, thus enhancing the stability of the calculation. The normal number is determined based on the statistical distribution of the baseline response delay in historical records, and can be taken as one-tenth of the minimum baseline response delay in historical records to ensure it is small enough to accommodate the range of variation in the baseline response delay.

[0056] It should be noted that when the target user's cloud-shared historical emotional response records are empty, meaning this interaction is the user's first interaction, the reference record set is empty, and the baseline response delay cannot be calculated using the arithmetic mean of historical records. In this case, the baseline response delay is set to the preset default baseline response delay. The default baseline response delay is preset according to the design goals of the digital human interaction system: it takes the universal comfort response delay value built into the digital human interaction system at the factory. This value is obtained by collecting a large amount of user comfort rating data on the digital human's response speed in controlled experiments, statistically determining the response delay parameter that gives the highest rating to most users. For example, the statistically obtained universal comfort response delay is 0.80 seconds, therefore the default baseline response delay is set to 0.80 seconds.

[0057] Substituting the default baseline response delay into the formula for calculating the matching factor, we get the default matching factor = (1 + dynamic asymmetry coefficient) ÷ (default baseline response delay + normal number). This default matching factor does not rely entirely on the target user's personal historical data, but it retains the real-time adjustment effect of the user's emotional dynamic characteristics in the current interaction through the dynamic asymmetry coefficient. In the cold start phase, it can use the general comfort delay as a benchmark and the user's current emotional asymmetry as a correction to output a matching factor that balances safety and personalization.

[0058] While outputting the emotional response speed matching factor, the cloud generates a source marker based on the number of historical records used for this matching: when the number of records in the reference record set reaches or exceeds the record sufficiency threshold, the source marker is set to sufficient historical data; when the number of reference records is below the record sufficiency threshold but greater than zero, the source marker is set to sparse historical data; when no historical records are available, the source marker is set to the default initial value. The record sufficiency threshold is preset according to the statistical stability requirements of the emotional matching factor calculation, taking the minimum sample size corresponding to the relative fluctuation amplitude of the baseline response delay not exceeding the preset fluctuation tolerance as the record sufficiency threshold. The relative fluctuation amplitude is calculated as follows: assuming the current reference record set contains n records, the corresponding baseline response delay is R_n. From the user's historical emotional response records, n records are randomly selected for a total of m times (e.g., m is 100 times), and the baseline response delay R_n^(k), k=1,2,…,m is calculated for each selection. Calculate the mean R_n_mean and standard deviation R_n_std of these m baseline response delays. Then, the relative volatility for a sample size of n is calculated as R_n_std ÷ R_n_mean. A preset volatility tolerance of, for example, 0.05, means that the baseline response delay is allowed to fluctuate within a 5% range. Starting with n=1, the sample size is gradually increased. When n first satisfies the condition that the relative volatility is ≤0.05, this n is the sufficient record threshold. For example, performing the above statistical stability analysis on a user's historical sentiment response records, the relative volatility when the sample size n=8 is 0.062, which is greater than the volatility tolerance of 0.05; the relative volatility when the sample size n=10 is 0.043, which is less than the volatility tolerance of 0.05. Therefore, the sufficient record threshold is set to 10 records.

[0059] For example, the target user has 20 historical emotional response records. These are sorted in ascending order by the absolute value of the difference between the current emotional duration and the previous one (1.5 seconds). The top 10 records are then used as a reference set. The optimal response delay data for these 10 records are 0.8 seconds, 0.9 seconds, 0.7 seconds, 1.0 seconds, 0.85 seconds, 0.95 seconds, 0.75 seconds, 1.1 seconds, 0.82 seconds, and 0.88 seconds, respectively. The baseline response delay is the arithmetic mean of these ten values, approximately 0.875 seconds. The dynamic asymmetry coefficient is 0.40, the normality factor is set to 0.05, and the emotional response speed matching factor is calculated as (1 + 0.40) ÷ (0.875 + 0.05) = 1.40 ÷ 0.925 ≈ 1.51.

[0060] The above example assumes the user already has 20 historical sentiment response records. For a user's first interaction with an empty history, the baseline response delay is set to the default 0.80 seconds. If the dynamic asymmetry coefficient for this interaction is 0.30 and the normal coefficient is 0.05, then the default matching factor = (1 + 0.30) ÷ (0.80 + 0.05) = 1.30 ÷ 0.85 ≈ 1.53. After this interaction is completed, the feedback difference is sent back to the cloud to generate the first historical record. Subsequent interactions can then use the baseline response delay based on the user's own historical data for calculation.

[0061] The following technical effects were achieved through this step: First, by leveraging the target user's historical emotional response records shared and managed in the cloud, the system filters the optimal response delay data within scenarios with similar emotional durations, outputting a personalized emotional response speed matching factor for that user. Compared to solutions that share the same set of response parameters with all users, the historical records accumulated in the cloud-shared management system allow the matching factor to reflect the user's preferred response speed, validated through feedback, in different emotional scenarios.

[0062] Second, the matching factor integrates historical baseline delay and current dynamic asymmetry coefficient. Based on the user's personalized historical preferences, it is adjusted in real time according to the degree of current emotional asymmetry, so that the digital human's response speed can maintain the user's consistent interaction rhythm and habits, while also making adaptive fine adjustments to the dynamic characteristics of current emotions.

[0063] S400: Receive the emotional response speed matching factor, and combine it with the emotional intensity value at the current emotional peak moment to calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. The emotional response speed matching factor integrates the target user's historical preferences and current emotional asymmetry characteristics, determining the overall speed tendency of the digital human's emotional response at a global level. However, the emotional response speed matching factor only determines the "speed of response" and has not yet been translated into specific execution parameters that can directly drive the digital human's animation and speech synthesis. The emotional output of the digital human involves dynamic adjustments in two core dimensions: first, the transition duration between adjacent keyframes in facial expression animation, which determines the speed at which the expression switches from the current state to the target state; and second, the range of pitch variation in speech synthesis, which determines the intonation of the digital human's speech. These two parameters need to be scaled and adjusted according to the peak intensity of the current emotion; the stronger the emotion and the higher the matching factor, the faster the animation transition and the wider the range of pitch variation should be.

[0064] Step S400 provided in this embodiment of the invention includes: The emotional intensity value at the current emotional peak time is determined based on the smooth intensity value corresponding to the smooth intensity sequence at the emotional peak time. Read the duration of the emotion, the first slope of change, and the second slope of change; Multiply the emotional response speed matching factor by the emotional duration to obtain the emotional response transition base, and combine it with the emotional intensity value at the current emotional peak moment to obtain the initial transition duration; Based on the second slope of change, the denominator of the falling damping is calculated, wherein the denominator of the falling damping increases as the absolute value of the second slope of change increases; Take the reciprocal of the denominator of the descent damping to obtain the damping coefficient of the descent stage, and combine it with the initial transition time to determine the acceleration transition time; Based on the emotional intensity value at the current emotional peak moment, the emotional response speed matching factor, and the absolute value of the second change slope, a pitch dynamic range compression coefficient is determined, wherein the pitch dynamic range compression coefficient is positively correlated with the emotional intensity value, positively correlated with the emotional response speed matching factor, and negatively correlated with the absolute value of the second change slope.

[0065] The specific implementation method is as follows: First, determine the emotional intensity value at the current emotional peak moment. Based on the smoothed intensity value of the smoothed intensity sequence corresponding to the emotional peak moment in step S100, the emotional intensity value at the current emotional peak moment is determined after peak correction. The specific determination method for peak correction will be explained in detail in subsequent steps. The corrected emotional intensity value can more accurately reflect the true peak level of this emotional event.

[0066] Then, the duration of the emotion, the first slope of change, and the second slope of change are read to calculate the duration of the accelerated transition.

[0067] Calculate the emotional response transition baseline. Emotional response transition baseline = Emotional response speed matching factor × Emotional duration. The emotional response transition baseline combines two factors: the larger the matching factor, the faster the digital human's response should be, and the larger the transition baseline accordingly; the longer the emotional duration, the more fully the user's emotional accumulation process, and the faster the digital human's facial expression transition can be to match the user's emotional investment, thus increasing the transition baseline accordingly. The unit of the emotional response transition baseline is seconds, reflecting the facial expression transition time benchmark initially determined based on the user's emotional dynamic characteristics and personalized matching factors. Calculate the initial transition duration: Initial transition duration = Emotional response transition baseline ÷ Emotional intensity value at the current emotional peak moment. The larger the emotional intensity value, the shorter the initial transition duration. The physical meaning is: when the user's emotional peak is high, the digital human's expression should switch to the corresponding emotional state more quickly to express timely emotional resonance; when the emotional peak is low, the facial expression transition can be appropriately slowed down to maintain a natural and smooth interaction.

[0068] Calculate the descent phase damping coefficient. Based on the second slope of change, calculate the descent damping denominator. Descent damping denominator = 1 + |second slope of change|. The descent damping denominator increases as the absolute value of the second slope of change increases. The second slope of change represents the slope of the emotional descent phase, usually negative, and its absolute value represents the speed of emotional fading. Take the reciprocal of the descent damping denominator to obtain the descent phase damping coefficient. Descent phase damping coefficient = 1 ÷ (1 + |second slope of change|). The descent phase damping coefficient ranges from 0 to 1. When the absolute value of the second slope of change is large, i.e., the user's emotion fades quickly, the descent damping denominator is large, the descent phase damping coefficient is small, the compression effect on the initial transition time is stronger, the transition time is accelerated to a shorter duration, and the digital human's expression transitions to the target state more quickly to match the user's rapidly fading emotional rhythm. When the absolute value of the second change slope is small, that is, the user's emotion fades slowly, the denominator of the descent damping is small, the damping coefficient of the descent phase is large, the compression effect on the initial transition time is weak, the acceleration transition time is long, and the digital human's expression transition maintains a slow rhythm.

[0069] The accelerated transition duration is calculated as follows: Accelerated transition duration = Initial transition duration × Descent phase damping coefficient. The accelerated transition duration is the optimal facial expression transition time value obtained by combining the emotion duration, matching factor, peak intensity, and descent slope, and is measured in seconds. The accelerated transition duration determines the time length for the transition from the current facial expression keyframe to the target facial expression keyframe in digital human facial expression animation. This duration is ultimately mapped to the transition time parameter between facial expression keyframes in the digital human rendering engine.

[0070] Finally, the pitch dynamic range compression coefficient is calculated. The pitch dynamic range compression coefficient = (emotional intensity value at the current emotional peak × emotional response speed matching factor) ÷ (1 + |second change slope|). This coefficient is positively correlated with the emotional intensity value and the emotional response speed matching factor, and negatively correlated with the absolute value of the second change slope. The higher the emotional intensity value, the stronger the user's emotion, and the larger the pitch variation range of the digital human's speech should be to express richer tonal emotions. The higher the matching factor, the faster the overall response rhythm of the digital human, and the greater the amplitude of pitch variation. The larger the absolute value of the second change slope, the faster the user's emotion fades, and the pitch variation range should be moderately compressed to avoid the digital human speaking in a high pitch after the user's emotion has quickly faded, causing a sense of incongruity. The pitch dynamic range compression coefficient is used to adjust the scaling range of the fundamental frequency in the speech synthesis module. When the coefficient is greater than 1, the pitch range is widened; when the coefficient is less than 1, the pitch range is narrowed.

[0071] For example, the emotional response speed matching factor obtained in the above example is 1.51. The emotional intensity value at the current emotional peak is 0.68. The first change slope is 0.42, and the second change slope is -0.18. The emotional duration is 1.5 seconds. Emotional response transition base = 1.51 × 1.5 = 2.265. Initial transition duration = 2.265 ÷ 0.68 ≈ 3.33 seconds. Descent damping denominator = 1 + |-0.18| = 1.18. Descent phase damping coefficient = 1 ÷ 1.18 ≈ 0.847. Accelerated transition duration = 3.33 × 0.847 ≈ 2.82 seconds. Pitch dynamic range compression coefficient = (0.68 × 1.51) ÷ 1.18 = 1.027 ÷ 1.18 ≈ 0.87.

[0072] During the aforementioned calculation process, edge nodes achieve state coordination with the cloud in the following ways: After receiving the emotional response speed matching factor from the cloud, the edge node parses the source marker of the factor from the matching factor. This marker is attached by the cloud when outputting the matching factor and is used to indicate whether the current matching factor is calculated based on sufficient historical data, sparse historical data, or a default initial value. When the source marker indicates sparse historical data or a default initial value, it indicates that the credibility of the current matching factor is limited. In this case, the edge node performs a weighted average of the initially calculated accelerated transition time and the built-in conservative parameter value. The conservative parameter value is preset to a general safety value suitable for most users, and the weight in the weighted average is dynamically determined according to the confidence level of the source marker. When the source marker is the default initial value, the weight of the conservative parameter value is 0.6, and the weight of the initially calculated value is 0.4; when the source marker is sparse historical data, the weight of the conservative parameter value is 0.4, and the weight of the initially calculated value is 0.6; when the source marker is sufficient historical data, the initially calculated value is used directly without weighted averaging. This collaborative approach enables edge nodes to automatically tilt towards a safe and conservative control strategy when their personalized matching capabilities in the cloud are not yet sufficient, ensuring that the lower limit of the interactive experience is not breached. At the same time, as user historical data accumulates, it automatically and gradually transitions to fully personalized parameter control.

[0073] For example, in this interaction, the source returned by the cloud is marked as having sufficient historical data. The edge nodes directly use the preliminary calculated value without weighted averaging, keeping the accelerated transition time at 2.82 seconds. If a new user interacts for the first time, the cloud returns a default matching factor due to a lack of historical data, with the source marked as the default initial value. The edge nodes initially calculate the accelerated transition time as 3.50 seconds. This is weighted and averaged with the conservative parameter value of 2.80 seconds, resulting in a final accelerated transition time of 0.4 × 3.50 + 0.6 × 2.80 = 3.08 seconds. This convergence compared to the original calculated value avoids the interaction rhythm deviating from the comfortable range for most users due to a lack of historical data support for the matching factor.

[0074] Specifically, the emotional intensity value at the current emotional peak time is determined based on the smoothing intensity value corresponding to the emotional peak time in the smoothing intensity sequence, including: In the smooth intensity sequence, the smooth intensity value corresponding to the emotional peak moment is read as the center peak intensity; Read the smooth intensity value corresponding to the sampling time before the emotional peak moment as the previous neighbor intensity, and read the smooth intensity value corresponding to the sampling time after the emotional peak moment as the next neighbor intensity. Subtracting the preceding intensity from the central peak intensity yields the instantaneous change at the end of the rising phase; subtracting the following intensity from the central peak intensity yields the instantaneous change at the beginning of the falling phase. Based on the first slope of change and the fixed time interval, the theoretical change amount in the rising phase is calculated, and based on the absolute value of the second slope of change and the fixed time interval, the theoretical change amount in the falling phase is calculated. When the instantaneous change at the end of the rising phase is less than the theoretical change during the rising phase, and the instantaneous change at the beginning of the falling phase is less than the theoretical change during the falling phase, the central peak intensity is taken as the emotional intensity value at the current emotional peak moment. Otherwise, the arithmetic mean of the preceding neighbor intensity, the central peak intensity, and the following neighbor intensity is taken, and the result of the arithmetic mean is taken as the emotional intensity value at the current emotional peak moment.

[0075] The specific implementation method is as follows: In the smoothed intensity sequence obtained in the above steps, the smoothed intensity value corresponding to the emotional peak moment is read as the center peak intensity. The smoothed intensity value corresponding to the sampling moment before the emotional peak moment is read as the previous neighbor intensity. The smoothed intensity value corresponding to the sampling moment after the emotional peak moment is read as the next neighbor intensity.

[0076] Subtracting the preceding intensity from the central peak intensity yields the instantaneous change in intensity at the end of the ascent. This instantaneous change reflects the actual increase in emotional intensity during the last sampling interval before reaching the peak. Subtracting the following intensity from the central peak intensity yields the instantaneous change in intensity at the beginning of the descent. This instantaneous change reflects the actual decrease in emotional intensity during the first sampling interval after passing the peak.

[0077] Based on the first slope of change and a fixed time interval, the theoretical change in the rising phase is calculated. Theoretical change in the rising phase = First slope of change × Fixed time interval. The theoretical change in the rising phase represents the expected increase in emotional intensity within a single sampling interval, following a linearly fitted upward trend. Based on the absolute value of the second slope of change and a fixed time interval, the theoretical change in the falling phase is calculated. Theoretical change in the falling phase = |Second slope of change| × Fixed time interval. The theoretical change in the falling phase represents the expected decrease in emotional intensity within a single sampling interval, following a linearly fitted downward trend.

[0078] The instantaneous change at the end of the ascent is compared with the theoretical change during the ascent phase, and the instantaneous change at the beginning of the descent is also compared with the theoretical change during the descent phase. When the instantaneous change at the end of the ascent is less than the theoretical change during the ascent phase, and the instantaneous change at the beginning of the descent is less than the theoretical change during the descent phase, it indicates that the actual intensity change within the two sampling intervals before and after the emotional peak is less than the change predicted by the linear fitting. That is, the peak exhibits a gentle top in the local neighborhood, and this peak truly reflects the apex of the user's emotion. The intensity of the central peak is then taken as the emotional intensity value at the current emotional peak moment.

[0079] When the above conditions are not met—that is, when the instantaneous change at the end of the ascent is greater than or equal to the theoretical change during the ascent phase, or when the instantaneous change at the beginning of the descent is greater than or equal to the theoretical change during the descent phase—it indicates that the actual intensity change on at least one side before or after the peak exceeds the linear fitting prediction. This peak may be a local sharp spike caused by sampling fluctuations or short-term noise, and not a true emotional peak. In this case, the arithmetic mean of the preceding intensity, the central peak intensity, and the following intensity is taken, and the result is used as the emotional intensity value at the current emotional peak moment to smooth out the impact of local fluctuations on the peak intensity estimation.

[0080] The following technical effects were achieved through this step: First, this step combines the personalized emotional response speed matching factor delivered from the cloud with parameters such as the local emotional peak intensity and descent slope extracted from the edge. Through concatenated calculations, the matching factor is transformed into two directly executable dynamic parameters: the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. The calculation of the accelerated transition duration introduces a descent phase damping coefficient, ensuring that the facial expression transition speed is influenced not only by the matching factor and emotional intensity but also by the user's emotional decay speed. When the descent is rapid, the transition is accelerated to match the user's emotional decay rhythm. Simultaneously, in determining the emotional peak intensity value, by comparing the actual change before and after the peak with the theoretical change, spurious peaks caused by sampling fluctuations are automatically identified and corrected, improving the accuracy of peak intensity estimation.

[0081] Second, this step establishes an adaptive transition mechanism from cloud-based matching confidence to edge-end control parameters by parsing source tags at edge nodes and weighting them with conservative parameter values. When historical data in the cloud is insufficient, the source tags represent sparse historical data, and the statistical stability of the matching factors is limited. In this case, edge nodes automatically tilt towards conservative parameter values, using general safety parameters to constrain the deviation of personalized matching factors, preventing the control parameters from deviating significantly from the comfort range of most users due to accidental deviations in a small amount of historical data. As user interactions increase and historical records gradually accumulate, the source tags are automatically upgraded, and the weight of conservative parameter values ​​continuously decreases until it reaches zero. The control parameters smoothly transition from a general safety mode to a fully personalized mode, avoiding inconsistencies in the interactive experience caused by parameter mutations.

[0082] S500: Based on the accelerated transition duration and the dynamic range compression coefficient of tone, generate a digital human emotion output control sequence and send it to the digital human terminal for execution. At the same time, send the difference between the actual change and the expected change in the target user's emotion intensity after execution back to the cloud and update the historical emotion response records shared and managed in the cloud.

[0083] The accelerated transition duration of digital human facial expression animation and the compression coefficient of the dynamic range of tone in speech synthesis determine the dynamic adjustment of the digital human's emotional output from the visual and auditory dimensions, respectively. However, these two parameters are currently in numerical form and have not yet been translated into an executable instruction format for the digital human terminal. The digital human terminal needs to receive structured control sequences to execute the corresponding facial expression rendering and speech synthesis operations according to the specified transition duration and tone parameters. Furthermore, after the emotional response is executed, the actual change in the user's emotional intensity reflects the actual effect of the response. If the response is good, the actual change should be close to the expected change; if the response is too fast or too slow, causing user discomfort, the actual change will deviate from the expectation.

[0084] Step S500 provided in this embodiment of the invention includes: The accelerated transition duration is used as the transition time parameter between facial expression keyframes in the digital human facial expression animation, and the pitch dynamic range compression coefficient is used as the pitch scaling parameter for digital human speech synthesis. The digital human emotion output control sequence is generated by combining these parameters. The digital human emotion output control sequence is sent to the digital human terminal for execution. After the digital human terminal finishes execution, the edge node collects the time-series data of the target user's voice emotion intensity again, and extracts the emotion intensity value of the first emotional peak moment in time sequence as the emotion intensity value of the emotional peak moment after execution. Based on the emotional intensity value at the emotional peak moment after execution and the emotional intensity value at the current emotional peak moment, the actual change is determined; Based on the emotional response speed matching factor, the emotional intensity value at the current emotional peak moment, and the duration of the emotion, the expected change is determined; Calculate the difference between the actual change and the expected change, and send the difference back to the cloud; The cloud will use the received difference as the response deviation field, the emotion duration as the emotion duration field, the dynamic asymmetry coefficient as the dynamic asymmetry coefficient field, and the emotion response speed matching factor as the current response delay field, and store them together as a new historical emotion response record, adding it to the historical emotion response record shared and managed by the cloud.

[0085] The specific implementation method is as follows: Accelerated transition duration is used as the transition time parameter between facial expression keyframes in digital human facial expression animation. In the digital human rendering engine, facial expression animation consists of multiple preset sequences of facial expression keyframes, and the transition time between adjacent keyframes is a fixed value by default. The accelerated transition duration replaces this default value and is written into the transition time register of the facial expression animation controller, making the time for the digital human's expression to switch from the current keyframe to the target keyframe equal to the accelerated transition duration. The shorter the accelerated transition duration, the faster the expression switching, and the more agile the digital human's facial reactions.

[0086] The pitch dynamic range compression factor is used as the pitch scaling parameter in the digital human's speech synthesis. In the digital human's speech synthesis module, the fundamental frequency of the synthesized speech typically has a default natural fluctuation range. The pitch dynamic range compression factor is used as a scaling factor for this fundamental frequency range and written into the pitch parameter register of the speech synthesis controller. When the factor is greater than 1, the fundamental frequency variation range of the synthesized speech is widened, the intonation and rhythm of the speech are more pronounced, and richer emotional expression is conveyed. When the factor is less than 1, the fundamental frequency variation range is narrowed, and the speech intonation tends to be smoother.

[0087] The aforementioned transition time and pitch scaling parameters are combined and encapsulated into a digital human emotion output control sequence. This control sequence is represented in a structured data format, containing transition time and pitch scaling fields. The control sequence also includes identifiers for the target facial expression keyframes and identifiers for the synthesized speech text, enabling the digital human terminal to fully execute this emotion output task.

[0088] The digital human's emotion output control sequence is sent to the digital human terminal via the communication link between the edge node and the digital human terminal. After receiving the control sequence, the digital human terminal executes the expression animation controller to switch expression keyframes according to the accelerated transition duration, and the speech synthesis controller synthesizes and plays the speech according to the pitch dynamic range compression coefficient. The two outputs are executed synchronously to complete this digital human emotion output.

[0089] After the digital human terminal completes its execution, the target user's voice emotion intensity time-series data is collected again via edge nodes. From the newly collected voice emotion intensity time-series data, the emotion intensity value of the first occurrence of the emotion peak in time sequence is extracted as the emotion intensity value of the post-execution emotion peak moment. The emotion intensity value of the post-execution emotion peak moment reflects the intensity of the user's emotional response to the output after the digital human performs the emotional output.

[0090] The actual change is determined based on the emotional intensity value at the peak of emotion after execution and the emotional intensity value at the current peak of emotion. Actual change = Emotional intensity value at the peak of emotion after execution - Emotional intensity value at the current peak of emotion. A positive actual change indicates that the user's emotional intensity increased after execution, and the digital human's output stimulated the user's emotions. A negative actual change indicates that the user's emotional intensity decreased after execution, and the digital human's output soothed the user's emotions. A zero actual change indicates that the user's emotional intensity did not change significantly.

[0091] The expected change is determined based on the emotional response speed matching factor, the emotional intensity value at the current emotional peak, and the duration of the emotion. Expected Change = (Emotional Response Speed ​​Matching Factor × Emotional Intensity Value at the Current Emotional Peak) ÷ (Emotional Duration + 1). This formula shows that the expected change is directly proportional to the matching factor and peak intensity, and inversely proportional to the duration of the emotion. A larger matching factor indicates a more positive response from the digital human, and a greater expected magnitude of the user's emotional change. A higher peak intensity indicates a stronger current emotion from the user, and a stronger expectation of a positive response to the digital human's output. A longer duration of emotion indicates a longer period of high emotional engagement, leading to a smoother emotional change and a correspondingly smaller expected magnitude of change. Adding 1 to the denominator prevents an abnormally amplified expected change value when the duration of the emotion is too short.

[0092] Calculate the difference between the actual change and the expected change, and send the difference back to the cloud. A difference of zero or close to zero indicates that the actual effect of the response matches the expectation, and the current matching factor is appropriate for the user in the current emotional context. A large positive difference indicates that the actual emotional improvement exceeds expectations, the matching factor may be too large, and the digital human's response is too strong. A large negative difference indicates that the actual emotional improvement falls short of expectations, the matching factor may be too small, and the digital human's response is too weak.

[0093] The cloud uses the received difference as the response deviation field, the emotion duration as the emotion duration field, the dynamic asymmetry coefficient as the dynamic asymmetry coefficient field, and the emotion response speed matching factor as the response delay field, storing them together as a new historical emotion response record and adding it to the cloud-shared management history. This record is then stored in the target user's corresponding record table in the cloud-shared management database. As the number of user interactions increases, historical emotion response records accumulate, providing a more sufficient data foundation for filtering the reference record set and extracting the optimal response delay, allowing the emotion response speed matching factor to gradually approach the user's optimal response parameters.

[0094] For example, the accelerated transition duration obtained from the above steps is 2.82 seconds, and the pitch dynamic range compression coefficient is 0.87. The digital human emotion output control sequence includes a transition time field of 2.82 seconds and a pitch scaling field of 0.87. After the control sequence is sent to the digital human terminal, the facial expression animation controller sets the transition time of the facial expression keyframe to 2.82 seconds, and the speech synthesis controller scales the base frequency range to 0.87 times the default range. After the output is completed, the edge node collects and extracts the emotion intensity value of the emotional peak moment after execution, which is 0.52. The current emotion peak moment has an emotion intensity value of 0.68, and the actual change is 0.52 - 0.68 = -0.16, which is negative, indicating that the user's emotion intensity has decreased after this digital human output. The expected change is (1.51 × 0.68) ÷ (1.50 + 1) = 1.027 ÷ 2.50 ≈ 0.41. The difference between the actual change and the expected change is -0.16 - 0.41 = -0.57, which is negative and has a large absolute value, indicating that the actual effect of this response is far lower than expected. The cloud will store the difference of -0.57, the emotional duration of 1.50, the dynamic asymmetry coefficient of 0.40, and the delay of this response of 1.51 as a new historical emotional response record and add it to the historical emotional response records shared and managed in the cloud.

[0095] The following technical effects were achieved through this step: First, this step encapsulates the acceleration transition duration and pitch dynamic range compression coefficient into a structured control sequence recognizable by the digital human terminal, transforming abstract emotional response parameters into specific instructions executable by the rendering engine and speech synthesis module, thus achieving a complete closed loop from emotion analysis to digital human output. Simultaneously, by transmitting the feedback difference after execution back and updating the historical emotional response records shared and managed in the cloud, continuous accumulation and optimization of emotional response strategies across multiple rounds of interaction are achieved.

[0096] Second, after this step is completed on the digital human terminal, the user's emotional intensity data is collected again. The difference between the actual change and the expected change is calculated and transmitted back to the cloud, forming a continuous optimization mechanism from perception, decision-making, execution to feedback. The feedback records stored in the cloud provide continuously updated user data for subsequent matching factor calculations, enabling the digital human's emotional response strategy to gradually approach the user's optimal parameters as the number of interactions increases.

[0097] Example 2, as Figure 4 As shown, based on the same inventive concept as the cloud-edge collaborative digital human sharing management method provided in Embodiment 1, this embodiment of the invention also provides a cloud-edge collaborative digital human sharing management system, the system comprising: The voice emotion acquisition and analysis module 11 is used to acquire real-time time-series data of voice emotion intensity during the interaction between the target user and the digital human through edge nodes, and extract the first change slope of the target user's emotion rising phase and the second change slope of the emotion falling phase from the voice emotion intensity time-series data. The emotion dynamic analysis module 12 is used to calculate the dynamic asymmetry coefficient of the target user's emotion based on the first change slope and the second change slope, and to obtain the duration of emotion before the current emotional peak moment. The cloud-based sharing management module 13 is used to upload the dynamic asymmetry coefficient and the duration of emotion to the cloud. The cloud outputs the emotion response speed matching factor for the target user based on the best response delay data corresponding to different durations of emotion in the historical emotion response records of the target user's cloud-based sharing management. The output parameter calculation module 14 is used to receive the emotional response speed matching factor and, in combination with the emotional intensity value at the current emotional peak moment, calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. The emotion output control module 15 is used to generate a digital human emotion output control sequence based on the accelerated transition duration and the dynamic range compression coefficient of the tone, and send it to the digital human terminal for execution. At the same time, it sends back the difference between the actual change and the expected change in the target user's emotion intensity after execution to the cloud and updates the historical emotion response records shared and managed in the cloud.

[0098] In one embodiment, the voice emotion acquisition and analysis module 11 is further configured to perform voice activity detection on the voice signal of the target user through edge nodes, sample the detected effective voice segments at fixed time intervals, and obtain voice frames at multiple consecutive sampling times. Extract the short-time energy value and fundamental frequency value of the speech frame at each sampling time; Based on the upper limit of the dynamic range of the audio acquisition at the edge node, the short-time energy value is normalized to a normalized energy value, wherein the upper limit of the dynamic range is the short-time energy value corresponding to the full-scale input amplitude of the audio analog-to-digital converter. Divide the fundamental frequency value by the mean fundamental frequency of the target user in the unemotional speech segment to obtain the normalized fundamental frequency, wherein the unemotional speech segment is the speech segment corresponding to the no-task confirmation speech issued by the target user according to the system prompt during the interaction initialization phase; The normalized energy and the normalized fundamental frequency are weighted and summed to obtain the emotional intensity value at the corresponding sampling time. The emotional intensity values ​​from multiple consecutive sampling times are arranged in chronological order and then subjected to median filtering to obtain the temporal data of the speech emotional intensity.

[0099] Extract the short-time energy and fundamental frequency values ​​of the speech frame at each sampling time, including: Calculate the sum of squared amplitudes of all sampling points within the speech frame, and divide the sum of squared amplitudes by the total number of sampling points in the speech frame to obtain the short-time energy value; Calculate the arithmetic mean of the amplitudes of all sampling points within the speech frame, and subtract the arithmetic mean from the amplitude of each sampling point to obtain a zero-mean corrected sampling point sequence. Arrange the corrected sampling point sequence in chronological order, and calculate the autocorrelation coefficient corresponding to each delay point. The delay point is an integer between the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period. The autocorrelation coefficient is the inner product of the corrected sampling point sequence and the sequence after shifting itself backward by the delay point. Divide the autocorrelation coefficient corresponding to each delay point by the autocorrelation coefficient of the modified sampling point sequence at zero delay to obtain the normalized autocorrelation coefficient. Arrange the normalized autocorrelation coefficients corresponding to all non-zero delay points in ascending order of delay points to form an autocorrelation function sequence; Within a search interval defined by the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period, the global maximum peak position of the autocorrelation function sequence is searched. The sampling rate is divided by the number of delay points corresponding to the global maximum peak position to obtain the fundamental frequency value. The first slope of change during the rising phase and the second slope of change during the falling phase of the target user's emotion intensity are extracted from the speech emotion intensity time-series data, including: The arithmetic mean of multiple consecutive emotion intensity values ​​in the speech emotion intensity time series data is calculated, and the result of the arithmetic mean is used as the smoothed intensity value at the corresponding center time to obtain the smoothed intensity sequence. Calculate the difference in emotional intensity values ​​between adjacent sampling times in the smoothed intensity sequence, and mark the sampling time corresponding to the turning point where the difference changes from a positive value to a negative value as the emotional peak time; For each emotional peak moment, in the smooth intensity sequence, the first turning point where the difference turns from negative to positive is retrieved forward, the sampling time corresponding to the first turning point is marked as the starting point of the rise, and the time period between the starting point of the rise and the emotional peak moment is determined as the emotional rise phase. The first turning point where the difference turns from negative to positive is retrieved backward, and the sampling time corresponding to the first turning point is marked as the end of the decline. The period between the emotional peak time and the end of the decline is determined as the emotional decline stage. Based on the emotional intensity value at each sampling moment during the emotional rise phase and the corresponding time, a linear fit is performed, and the slope obtained from the fit is used as the first change slope. Based on the linear fitting of the emotional intensity value at each sampling moment during the emotional decline phase with the corresponding time, the slope obtained by fitting is used as the second change slope.

[0100] In one embodiment, the emotional dynamic analysis module 12 is further configured to take the absolute value of the second change slope and use the absolute value as the descent rate value; Add the value of the first slope change to the value of the descent rate to obtain the sum of the absolute values ​​of the slopes, and subtract the value of the first slope change from the value of the descent rate to obtain the difference in the absolute values ​​of the slopes. The dynamic asymmetry coefficient is determined based on the ratio of the difference of the absolute values ​​of the slopes to the sum of the absolute values ​​of the slopes. Read the sampling time index corresponding to the current emotional peak moment and the sampling time index corresponding to the rising starting point, and calculate the index difference between the two sampling time indices. The sampling time index is a sequence number that is sequentially incremented according to the time order of consecutive sampling moments. Multiply the index difference by a fixed time interval to obtain the duration of the emotion.

[0101] In one embodiment, the cloud sharing management module 13 is also used to receive the dynamic asymmetry coefficient and the duration of the emotion uploaded by the edge node in the cloud; All records in the historical emotional response records shared and managed in the cloud are sorted in ascending order according to the absolute value of the difference between the emotional duration field and the received emotional duration, and the first half of the records are taken as the reference record set. Extract the best response delay data stored in each record in the reference record set, wherein the best response delay data is the value of the current response delay field corresponding to the record in the historical emotional response records that is closest to the duration of the emotion and has the smallest response deviation; Calculate the arithmetic mean of all extracted best response delay data to obtain the baseline response delay; An emotional response speed matching factor is determined based on the baseline response delay and the dynamic asymmetry coefficient, wherein the emotional response speed matching factor is negatively correlated with the baseline response delay and positively correlated with the dynamic asymmetry coefficient.

[0102] In one embodiment, the output parameter calculation module 14 is also used for The emotional intensity value at the current emotional peak time is determined based on the smooth intensity value corresponding to the smooth intensity sequence at the emotional peak time. Read the duration of the emotion, the first slope of change, and the second slope of change; Multiply the emotional response speed matching factor by the emotional duration to obtain the emotional response transition base, and combine it with the emotional intensity value at the current emotional peak moment to obtain the initial transition duration; Based on the second slope of change, the denominator of the falling damping is calculated, wherein the denominator of the falling damping increases as the absolute value of the second slope of change increases; Take the reciprocal of the denominator of the descent damping to obtain the damping coefficient of the descent stage, and combine it with the initial transition time to determine the acceleration transition time; Based on the emotional intensity value at the current emotional peak moment, the emotional response speed matching factor, and the absolute value of the second change slope, a pitch dynamic range compression coefficient is determined, wherein the pitch dynamic range compression coefficient is positively correlated with the emotional intensity value, positively correlated with the emotional response speed matching factor, and negatively correlated with the absolute value of the second change slope.

[0103] In the smooth intensity sequence, the smooth intensity value corresponding to the emotional peak moment is read as the center peak intensity; Read the smooth intensity value corresponding to the sampling time before the emotional peak moment as the previous neighbor intensity, and read the smooth intensity value corresponding to the sampling time after the emotional peak moment as the next neighbor intensity. Subtracting the preceding intensity from the central peak intensity yields the instantaneous change at the end of the rising phase; subtracting the following intensity from the central peak intensity yields the instantaneous change at the beginning of the falling phase. Based on the first slope of change and the fixed time interval, the theoretical change amount in the rising phase is calculated, and based on the absolute value of the second slope of change and the fixed time interval, the theoretical change amount in the falling phase is calculated. When the instantaneous change at the end of the rising phase is less than the theoretical change during the rising phase, and the instantaneous change at the beginning of the falling phase is less than the theoretical change during the falling phase, the central peak intensity is taken as the emotional intensity value at the current emotional peak moment. Otherwise, the arithmetic mean of the preceding neighbor intensity, the central peak intensity, and the following neighbor intensity is taken, and the result of the arithmetic mean is taken as the emotional intensity value at the current emotional peak moment.

[0104] In one embodiment, the emotion output control module 15 is further configured to use the accelerated transition duration as the transition time parameter between expression keyframes in the digital human expression animation, and the pitch dynamic range compression coefficient as the pitch scaling parameter for digital human speech synthesis, and combine them to generate the digital human emotion output control sequence. The digital human emotion output control sequence is sent to the digital human terminal for execution. After the digital human terminal finishes execution, the edge node collects the time-series data of the target user's voice emotion intensity again, and extracts the emotion intensity value of the first emotional peak moment in time sequence as the emotion intensity value of the emotional peak moment after execution. Based on the emotional intensity value at the emotional peak moment after execution and the emotional intensity value at the current emotional peak moment, the actual change is determined; Based on the emotional response speed matching factor, the emotional intensity value at the current emotional peak moment, and the duration of the emotion, the expected change is determined; Calculate the difference between the actual change and the expected change, and send the difference back to the cloud; The cloud will use the received difference as the response deviation field, the emotion duration as the emotion duration field, the dynamic asymmetry coefficient as the dynamic asymmetry coefficient field, and the emotion response speed matching factor as the current response delay field, and store them together as a new historical emotion response record, adding it to the historical emotion response record shared and managed by the cloud.

Claims

1. A cloud-edge collaborative digital human sharing management method, characterized in that, The method includes: The system collects real-time temporal data of voice emotion intensity during the interaction between the target user and the digital human through edge nodes, and extracts the first slope of the target user's emotion rise phase and the second slope of the emotion fall phase from the voice emotion intensity temporal data. Based on the first and second slopes of change, the dynamic asymmetry coefficient of the target user's emotion is calculated, and the duration of emotion before the current emotional peak is obtained. The dynamic asymmetry coefficient and the duration of emotion are uploaded to the cloud. The cloud outputs the emotion response speed matching factor for the target user based on the best response delay data corresponding to different durations of emotion in the historical emotion response records shared and managed in the cloud. Receive the emotional response speed matching factor, and combine it with the emotional intensity value at the current emotional peak moment to calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. Based on the accelerated transition duration and the dynamic range compression coefficient of tone, a digital human emotion output control sequence is generated and sent to the digital human terminal for execution. At the same time, the difference between the actual change and the expected change in the target user's emotion intensity after execution is sent back to the cloud to update the historical emotion response records shared and managed in the cloud.

2. The cloud-edge collaborative digital human sharing management method according to claim 1, characterized in that, Real-time temporal data of voice emotional intensity during the interaction between the target user and the digital human is collected through edge nodes, including: Voice activity detection is performed on the target user's voice signal through edge nodes, and the detected valid voice segments are sampled at fixed time intervals to obtain voice frames at multiple consecutive sampling times. Extract the short-time energy value and fundamental frequency value of the speech frame at each sampling time; Based on the upper limit of the dynamic range of the audio acquisition at the edge node, the short-time energy value is normalized to a normalized energy value, wherein the upper limit of the dynamic range is the short-time energy value corresponding to the full-scale input amplitude of the audio analog-to-digital converter. Divide the fundamental frequency value by the mean fundamental frequency of the target user in the unemotional speech segment to obtain the normalized fundamental frequency, wherein the unemotional speech segment is the speech segment corresponding to the no-task confirmation speech issued by the target user according to the system prompt during the interaction initialization phase; The normalized energy and the normalized fundamental frequency are weighted and summed to obtain the emotional intensity value at the corresponding sampling time. The emotional intensity values ​​from multiple consecutive sampling times are arranged in chronological order and then subjected to median filtering to obtain the temporal data of the speech emotional intensity.

3. The cloud-edge collaborative digital human sharing management method according to claim 2, characterized in that, Extract the short-time energy and fundamental frequency values ​​of the speech frame at each sampling time, including: Calculate the sum of squared amplitudes of all sampling points within the speech frame, and divide the sum of squared amplitudes by the total number of sampling points in the speech frame to obtain the short-time energy value; Calculate the arithmetic mean of the amplitudes of all sampling points within the speech frame, and subtract the arithmetic mean from the amplitude of each sampling point to obtain a zero-mean corrected sampling point sequence. Arrange the corrected sampling point sequence in chronological order, and calculate the autocorrelation coefficient corresponding to each delay point. The delay point is an integer between the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period. The autocorrelation coefficient is the inner product of the corrected sampling point sequence and the sequence after shifting itself backward by the delay point. Divide the autocorrelation coefficient corresponding to each delay point by the autocorrelation coefficient of the modified sampling point sequence at zero delay to obtain the normalized autocorrelation coefficient. Arrange the normalized autocorrelation coefficients corresponding to all non-zero delay points in ascending order of delay points to form an autocorrelation function sequence; Within the search interval defined by the number of sampling points corresponding to the minimum possible pitch period and the number of sampling points corresponding to the maximum possible pitch period, the global maximum peak position of the autocorrelation function sequence is searched, and the sampling rate is divided by the number of delay points corresponding to the global maximum peak position to obtain the fundamental frequency value.

4. The cloud-edge collaborative digital human sharing management method according to claim 1, characterized in that, Extracting the first slope of change during the rising phase and the second slope of change during the falling phase of the target user's emotion from the time-series data of the voice emotion intensity includes: The arithmetic mean of multiple consecutive emotion intensity values ​​in the speech emotion intensity time series data is calculated, and the result of the arithmetic mean is used as the smoothed intensity value at the corresponding center time to obtain the smoothed intensity sequence. Calculate the difference in emotional intensity values ​​between adjacent sampling times in the smoothed intensity sequence, and mark the sampling time corresponding to the turning point where the difference changes from a positive value to a negative value as the emotional peak time; For each emotional peak moment, in the smooth intensity sequence, the first turning point where the difference turns from negative to positive is retrieved forward, the sampling time corresponding to the first turning point is marked as the starting point of the rise, and the time period between the starting point of the rise and the emotional peak moment is determined as the emotional rise phase. The first turning point where the difference turns from negative to positive is retrieved backward, and the sampling time corresponding to the first turning point is marked as the end of the decline. The period between the emotional peak time and the end of the decline is determined as the emotional decline stage. Based on the emotional intensity value at each sampling moment during the emotional rise phase and the corresponding time, a linear fit is performed, and the slope obtained from the fit is used as the first change slope. Based on the linear fitting of the emotional intensity value at each sampling moment during the emotional decline phase with the corresponding time, the slope obtained by fitting is used as the second change slope.

5. The cloud-edge collaborative digital human sharing management method according to claim 1, characterized in that, Based on the first and second slopes of change, the dynamic asymmetry coefficient of the target user's emotion is calculated, and the duration of emotion preceding the current emotional peak is obtained, including: Take the absolute value of the second slope of change, and use the absolute value as the descent rate value; Add the value of the first slope change to the value of the descent rate to obtain the sum of the absolute values ​​of the slopes, and subtract the value of the first slope change from the value of the descent rate to obtain the difference in the absolute values ​​of the slopes. The dynamic asymmetry coefficient is determined based on the ratio of the difference of the absolute values ​​of the slopes to the sum of the absolute values ​​of the slopes. Read the sampling time index corresponding to the current emotional peak moment and the sampling time index corresponding to the rising starting point, and calculate the index difference between the two sampling time indices. The sampling time index is a sequence number that is sequentially incremented according to the time order of consecutive sampling moments. Multiply the index difference by a fixed time interval to obtain the duration of the emotion.

6. The cloud-edge collaborative digital human sharing management method according to claim 1, characterized in that, The dynamic asymmetry coefficient and the duration of the emotion are uploaded to the cloud. Based on the optimal response delay data corresponding to different durations of emotion in the target user's historical emotion response records shared and managed in the cloud, the cloud outputs an emotion response speed matching factor for the target user, including: The cloud receives the dynamic asymmetry coefficient and the duration of the emotion uploaded by the edge node; All records in the historical emotional response records shared and managed in the cloud are sorted in ascending order according to the absolute value of the difference between the emotional duration field and the received emotional duration, and the first half of the records are taken as the reference record set. Extract the best response delay data stored in each record in the reference record set, wherein the best response delay data is the value of the current response delay field corresponding to the record in the historical emotional response records that is closest to the duration of the emotion and has the smallest response deviation; Calculate the arithmetic mean of all extracted best response delay data to obtain the baseline response delay; An emotional response speed matching factor is determined based on the baseline response delay and the dynamic asymmetry coefficient, wherein the emotional response speed matching factor is negatively correlated with the baseline response delay and positively correlated with the dynamic asymmetry coefficient.

7. The cloud-edge collaborative digital human sharing management method according to claim 1, characterized in that, Receive the emotional response speed matching factor, and combine it with the emotional intensity value at the current emotional peak moment to calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis, including: The emotional intensity value at the current emotional peak time is determined based on the smooth intensity value corresponding to the smooth intensity sequence at the emotional peak time. Read the duration of the emotion, the first slope of change, and the second slope of change; Multiply the emotional response speed matching factor by the emotional duration to obtain the emotional response transition base, and combine it with the emotional intensity value at the current emotional peak moment to obtain the initial transition duration; Based on the second slope of change, the denominator of the falling damping is calculated, wherein the denominator of the falling damping increases as the absolute value of the second slope of change increases; Take the reciprocal of the denominator of the descent damping to obtain the damping coefficient of the descent stage, and combine it with the initial transition time to determine the acceleration transition time; Based on the emotional intensity value at the current emotional peak moment, the emotional response speed matching factor, and the absolute value of the second change slope, a pitch dynamic range compression coefficient is determined, wherein the pitch dynamic range compression coefficient is positively correlated with the emotional intensity value, positively correlated with the emotional response speed matching factor, and negatively correlated with the absolute value of the second change slope.

8. The cloud-edge collaborative digital human sharing management method according to claim 7, characterized in that, Based on the smoothing intensity value of the smoothing intensity sequence corresponding to the emotional peak moment, the emotional intensity value at the current emotional peak moment is determined, including: In the smooth intensity sequence, the smooth intensity value corresponding to the emotional peak moment is read as the center peak intensity; Read the smooth intensity value corresponding to the sampling time before the emotional peak moment as the previous neighbor intensity, and read the smooth intensity value corresponding to the sampling time after the emotional peak moment as the next neighbor intensity. Subtracting the preceding intensity from the central peak intensity yields the instantaneous change at the end of the rising phase; subtracting the following intensity from the central peak intensity yields the instantaneous change at the beginning of the falling phase. Based on the first slope of change and the fixed time interval, the theoretical change amount in the rising phase is calculated, and based on the absolute value of the second slope of change and the fixed time interval, the theoretical change amount in the falling phase is calculated. When the instantaneous change at the end of the rising phase is less than the theoretical change during the rising phase, and the instantaneous change at the beginning of the falling phase is less than the theoretical change during the falling phase, the central peak intensity is taken as the emotional intensity value at the current emotional peak moment. Otherwise, the arithmetic mean of the preceding neighbor intensity, the central peak intensity, and the following neighbor intensity is taken, and the result of the arithmetic mean is taken as the emotional intensity value at the current emotional peak moment.

9. The cloud-edge collaborative digital human sharing management method according to claim 1, characterized in that, Based on the accelerated transition duration and pitch dynamic range compression coefficient, a digital human emotion output control sequence is generated and sent to the digital human terminal for execution. Simultaneously, the difference between the actual and expected changes in the target user's emotional intensity after execution is transmitted back to the cloud, updating the historical emotion response records shared and managed in the cloud, including: The accelerated transition duration is used as the transition time parameter between facial expression keyframes in the digital human facial expression animation, and the pitch dynamic range compression coefficient is used as the pitch scaling parameter for digital human speech synthesis. The digital human emotion output control sequence is generated by combining these parameters. The digital human emotion output control sequence is sent to the digital human terminal for execution. After the digital human terminal finishes execution, the edge node collects the time-series data of the target user's voice emotion intensity again, and extracts the emotion intensity value of the first emotional peak moment in time sequence as the emotion intensity value of the emotional peak moment after execution. Based on the emotional intensity value at the emotional peak moment after execution and the emotional intensity value at the current emotional peak moment, the actual change is determined; Based on the emotional response speed matching factor, the emotional intensity value at the current emotional peak moment, and the duration of the emotion, the expected change is determined; Calculate the difference between the actual change and the expected change, and send the difference back to the cloud; The cloud will use the received difference as the response deviation field, the emotion duration as the emotion duration field, the dynamic asymmetry coefficient as the dynamic asymmetry coefficient field, and the emotion response speed matching factor as the current response delay field, and store them together as a new historical emotion response record, adding it to the historical emotion response record shared and managed by the cloud.

10. A cloud-edge collaborative digital human sharing management system, characterized in that, The system for implementing the method according to any one of claims 1 to 9, the system comprising: The voice emotion acquisition and analysis module is used to collect real-time time-series data of voice emotion intensity during the interaction between the target user and the digital human through edge nodes, and to extract the first change slope of the target user's emotion rising phase and the second change slope of the emotion falling phase from the voice emotion intensity time-series data. The emotion dynamic analysis module is used to calculate the dynamic asymmetry coefficient of the target user's emotion based on the first change slope and the second change slope, and to obtain the duration of the emotion before the current emotional peak moment. The cloud-based sharing management module is used to upload the dynamic asymmetry coefficient and the duration of emotion to the cloud. The cloud outputs an emotion response speed matching factor for the target user based on the best response delay data corresponding to different durations of emotion in the historical emotion response records of the target user's cloud-based sharing management. The output parameter calculation module is used to receive the emotional response speed matching factor and, in combination with the emotional intensity value at the current emotional peak moment, calculate the accelerated transition duration of the digital human facial expression animation and the pitch dynamic range compression coefficient of the speech synthesis. The emotion output control module is used to generate a digital human emotion output control sequence based on the accelerated transition duration and the dynamic range compression coefficient of the tone, and send it to the digital human terminal for execution. At the same time, it sends the difference between the actual change and the expected change in the target user's emotion intensity after execution back to the cloud and updates the historical emotion response records shared and managed in the cloud.

Images