A method and system for generating fine-grained digital human motion based on user intent awareness

By capturing user facial expressions and voice features in real time using a value-policy neural network, and dynamically adjusting the selection of digital human template frames, the problem of inaccurate prediction of user interruption intentions is solved, enabling fine-grained control of digital human actions and improving the realism and continuity in interactive scenarios.

CN117576273BActive Publication Date: 2026-11-17HANGZHOU YIWISE INTELLIGENT TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311447322.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-02
Publication Date
2026-11-17
Estimated Expiration
2043-11-02

AI Technical Summary

Technical Problem

In existing digital human interactions, the prediction of user interruption intentions is inaccurate, resulting in insufficient granularity in action generation and an inability to make timely adjustments, which affects the realism of digital humans in interactive scenarios.

Method used

A fine-grained digital human motion generation method based on user intent perception is adopted. The value-policy neural network captures the user's facial expressions and voice features in real time and dynamically adjusts the selection of digital human template frames to achieve fine-grained control of motion.

Benefits of technology

It improves the accuracy of predicting user interruption intentions, enables flexible adjustment of digital human actions, and enhances the realism and continuity of digital humans in interactive scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117576273B_ABST
    Figure CN117576273B_ABST
Patent Text Reader

Abstract

This invention discloses a high-fine-grained digital human action generation method and system based on user intent perception, belonging to the field of artificial intelligence technology. It includes initializing templates, a trajectory list, and a value-policy neural network; receiving interaction commands and calculating the current state; generating template frames through the policy network to select directions; synthesizing the template frame at the determined digital human template frame position with the current speech features to form the current digital human video frame and playing it; obtaining the current state, action value, reward value, and end flag tuple based on the current digital human state and adding them to the trajectory list; updating the parameters of the value-policy neural network based on the data in the trajectory list when the digital human is idle, and clearing the trajectory list after the update is complete. This invention intervenes in the selection of digital human template frames in a timely manner according to the constantly changing user state during the response process, dynamically generating digital human template fragments, and achieving fine-grained control over action generation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, and in particular relates to a method and system for generating highly fine-grained digital human actions based on user intent perception. Background Technology

[0002] Digital human technology has advanced rapidly in recent years. By inputting text or voice, virtual avatars can be driven to speak. Combined with natural language processing (NLP) technology, digital humans can also interact with people naturally and fluently, showing broad application prospects in e-commerce live streaming, intelligent customer service, intelligent shopping guides, online education, and many other fields. Current digital human technology can achieve a lip-sync accuracy of nearly 99%, and facial movements and expressions can be regenerated based on new text or voice content, resulting in a high degree of realism. However, current technology is somewhat lacking in the area of ​​digital human body movements. Currently, digital human movements are achieved by replaying actions from template videos. Specifically, current digital human technology sequentially plays body movements from template videos, then uses voice to drive the generation of a new face, replacing the face in the template to obtain a new digital human video. In interactive scenarios, this approach has a problem: after the digital human answers a question, its body movements should return to their normal state. This means that the movements should return to the state they were in when not speaking, such as being placed naturally on a table or hanging naturally in front of the body, assuming a posture of waiting to be asked a question. Instead, the body movements should remain in the posture of answering after the digital human has finished answering, as this would create inconsistency between the digital human's dialogue and its body movements, greatly affecting the realism of the digital human in interactive scenarios.

[0003] Current digital human motion generation primarily relies on replaying actions from template videos. However, the length of these template videos is limited. To ensure the continuity of the actions, most digital human synthesis methods in practice generate actions through repeated looping playback. That is, the actions are played from the first frame to the last frame of the template, then reversed back to the first frame, ensuring visual continuity of the body movements. To achieve motion repositioning, existing methods pre-calculate the length of the digital human's speech content and select segments of the repetitive digital human template based on this length. This ensures that when the digital human finishes speaking, the actions return to the initial first frame, achieving the effect of motion repositioning when the digital human finishes speaking.

[0004] However, the methods described above only address the issue of reverting actions when the digital human finishes its natural speech, without considering the scenario where the interactive digital human might be interrupted by the user. In practice, the user might ask follow-up questions or change the question before the digital human finishes its response, causing the digital human's answer to change accordingly. In this case, the digital human's body language should also change accordingly to make the digital human behave more like a real person in interactive applications. Currently, there are very few publicly available methods in this area. Some methods attempt to predict the likelihood of interruption based on the content of the digital human's dialogue, thereby pre-selecting a reasonable length for the template segment to be played repeatedly, so that the digital human can return to the first frame exactly when it finishes speaking, achieving the reverting of actions. However, these methods have two problems: First, they utilize very limited information, often only using dialogue-related text and voice content, failing to accurately predict and recognize the user's intention to interrupt; second, the granularity of decision-making is insufficient, meaning that the selection of the template segment length is made at the beginning of a speech, and cannot be adjusted in time according to the actual situation, thus affecting the flexibility of the digital human's action selection.

[0005] In summary, existing action generation methods based on predicting user interruption intentions in digital human interaction have two main problems:

[0006] (1) The information used is very limited. It often only uses the text and voice content related to the dialogue. It is not enough to make a more accurate prediction of the user's intention to interrupt the digital human's speech based solely on the voice and text content.

[0007] (2) The decision-making is not granular enough. That is, the selection of template fragment length is made at the beginning of a speech and cannot be corrected in time according to the actual situation. Even if the user’s strong intention to interrupt is detected in the middle, the action generation of the current answer content cannot be changed. Summary of the Invention

[0008] To address the issue of inaccurate repositioning of interactive digital human motion generation when interrupted by the user, this invention provides a high-fine-grained digital human motion generation method and system based on user intent perception.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] In a first aspect, the present invention provides a method for generating highly fine-grained digital human actions based on user intent awareness, comprising:

[0011] Step 1: Capture a video template of the digital human speaking state, initialize the trajectory list and the value-policy neural network for selecting the direction of the template frame, and pre-train the policy network in the value-policy neural network;

[0012] Step 2: Receive interactive instructions, convert the current user's voice input into a text question, generate the answer text, synthesize speech based on the answer text, and extract the speech feature sequence;

[0013] Step 3: Capture the user's facial image in real time, update the feature sequence of the voice to be played and its length, as well as the current digital human template frame position, based on the voice frame number of the current response, and calculate the current state by combining the user's current facial image information;

[0014] Step 4: For the current state, generate action values ​​representing the selection direction of the template frame through the policy network in the value-policy neural network, obtain the position of the digital human template frame corresponding to the next speech frame number, synthesize the current digital human video frame based on the template frame at the determined digital human template frame position and the current speech features, and play it.

[0015] Step 5: Based on the current state of the digital human, record the end marker and reward value, and add the current state, action value, reward value, and end marker tuple to the trajectory list; the digital human state includes natural end, interruption by the user during the answering process, and no interruption by the user during the normal answering process;

[0016] Step 6: Select different operations based on the current state of the digital human: if the digital human's state is that the answer ends naturally, proceed to step 7; if the digital human's state is that the answer was interrupted by the user, return to step 2; if the digital human's state is that the answer was given normally and was not interrupted by the user, increment the voice frame number by 1 and return to step 3.

[0017] Step 7: Determine if the digital human is idle. If so, update the parameters of the policy neural network based on the data in the trajectory list. After the update is complete, clear the trajectory list and return to step 2. If not, return to step 2.

[0018] Further, step 4, which generates action values ​​representing the selection direction of the template frame through the policy network in the value-policy neural network to obtain the position of the digital human template frame corresponding to the next speech frame number, includes:

[0019] Calculation via policy network ,in This is the policy probability vector. The i-th element This represents selecting the i-th action. The probability of choosing to use the template frame forward or backward is determined based on the probability value, thereby determining the position of the digital human template frame corresponding to the next voice frame number.

[0020] Furthermore, the step of determining whether to use the template frame forward or backward based on the probability value, and then determining the position of the digital human template frame corresponding to the next voice frame number, is expressed as follows:

[0021]

[0022]

[0023] in, This indicates the first character in the digital human's response to the synthesized speech. Motion values ​​at frame time, This indicates that a template frame is used forward. This indicates that the template frame will be used backward. This represents the probability value calculated by the policy network for choosing to use the template frame forward or backward. This indicates the action value corresponding to the higher probability value. This indicates the first part of the digital human's response to the synthesized speech. The corresponding digital human template frame position at frame time. Indicates the length of the digital human template. This indicates the first part of the digital human's response to the synthesized speech. The frame position of the digital human template corresponding to the frame.

[0024] Furthermore, the value network and policy network in the value-policy neural network adopt the same network structure but have different output dimensions in the last layer. The output dimension of the last layer in the value network is 1, and the output dimension of the last layer in the policy network is 2. Each of the value network and policy network includes a two-branch backbone network and several fully connected layers. The first branch of the two-branch backbone network is based on the input state. The concatenated input is used as the second branch, which takes the input state as its input. As input, the two branches extract features from the input data respectively, and then the output features of the two branches are unfolded into one dimension and spliced ​​together as the output of the two-branch backbone network.

[0025] Furthermore, the first branch is composed of several fully connected layers, and the second branch is composed of convolutional layers, max pooling layers, convolutional layers, max pooling layers, several residual blocks, and max pooling layers connected in sequence.

[0026] Secondly, the present invention provides a high-granularity digital human motion generation system based on user intent perception, for implementing the above-mentioned online digital human motion generation method.

[0027] Thirdly, the present invention provides an electronic device including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the above-described high-fine-grained digital human motion generation method.

[0028] Fourthly, the present invention provides a machine-readable storage medium storing machine-executable instructions, which, when called and executed by a processor, are used to implement the above-described high-fine-grained digital human motion generation method.

[0029] The beneficial effects of this invention are:

[0030] (1) On digital human interaction devices, the camera is used to acquire the user’s facial movements and expression input in real time. By capturing information such as the user’s facial expressions, the user’s intention to interrupt is analyzed, thereby enabling a more accurate prediction of the possibility of the user interrupting the digital human’s speech.

[0031] (2) The present invention enables more refined selection of digital human actions. Instead of selecting the length of the digital human template fragment at the beginning of each answer, the present invention intervenes in the selection of the digital human template frame in a timely manner according to the constantly changing user state during the answering process, dynamically generating digital human template fragments and realizing refined control of action generation. Attached Figure Description

[0032] Figure 1 This is a diagram of the backbone neural network structure shown in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of the structure of a value network (V-network) as shown in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the structure of a policy network (P network) as shown in an embodiment of the present invention;

[0035] Figure 4 This is a flowchart illustrating a high-granularity digital human motion generation method based on user intent perception, as shown in an embodiment of the present invention.

[0036] Figure 5 This is a schematic diagram of the electronic device terminal structure of the high-fine-grained digital human motion generation method based on user intent perception, as shown in an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0039] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used in this invention includes any and all combinations of one or more of the associated listed items.

[0040] like Figure 4 As shown in this embodiment, a high-fine-grained digital human action generation method based on user intent awareness is proposed, including:

[0041] Step 1: Video template of the digital human speaking. ,in This is the length of the video template.

[0042] Step two: Conduct customized training on the digital human to obtain its voice-driven model. .

[0043] The aforementioned voice-driven model This is prior art in the field, used to synthesize digital human images for real-time playback based on a selected template frame and the current speech feature input. The speech-driven model will not be discussed further here. To elaborate further.

[0044] Step 3: Initialize the value-policy neural network used for template frame selection. Let the policy network be denoted as... It is used to generate target action values ​​based on different input states, where the target action value is either a forward selection template frame or a backward selection template frame; the value network is denoted as... It is used to score different input states and reduce the prediction score and advantage value during training. The difference between states is considered; states with more rewards have higher scores, thus the optimal state is selected from the set. These are the parameters for the policy network and the value network, respectively. The value-policy neural network represents the state vector input to the input state. When using a value-policy neural network to calculate the result corresponding to the input state, it can be simplified to... and In this form, network parameters are omitted to simplify the expression.

[0045] At the same time, initialize the trajectory list traj = [].

[0046] The policy network initialized above Pre-training with offline data is required to enable the system to output repetitive digital human template segment selection results regardless of interruption. The training samples of the offline data are state vectors. The label is the action value. The action value It means selecting template frames forward or backward.

[0047] The aforementioned repetitive digital human template refers to a method that uses loop playback. When the digital human template has progressed from the initial frame to the last frame, instead of jumping back to the initial frame, it reverses the order, going back from the last frame to the first frame, repeating this process to ensure the continuity of the digital human's movements. This repetitive digital human template is existing technology in this field and will not be elaborated upon further here.

[0048] Step 4: Receive interactive commands based on the current user's voice input. Transformed into a text problem using an ASR model Then, generate the answer text based on the existing question-and-answer system. Finally, speech is synthesized using an existing speech synthesis system, and the corresponding feature sequences are extracted using an ASR model. ,in Indicates the first Frame speech features Frame number t corresponds to response time t. The initial speech length is... The initial voice frame number of the response The initial position of the digital human template frame used. .

[0049] In the above process, the voice input is sequentially converted into text questions, answer text, and synthesized speech, all of which can be achieved using existing technologies in this field, and will not be elaborated here.

[0050] Step 5: Capture the user's facial image in real time, based on the current voice frame number of the response. Update the feature sequence of the audio to be played and its length and the current digital human template frame position. According to the audio to be played length Current digital human template frame position and the user's facial images captured by the camera. The current state is calculated. .

[0051] Step 6: Generate the current state through the policy network P. Corresponding action value , ,in This indicates that the next frame is selected in a positive forward direction. This represents selecting the previous frame backwards, through the action. To determine the position of the digital human template frame in the next frame. Then, based on the currently selected frame and current speech features Using voice-driven models The digital human video frames are synthesized to generate digital human images for real-time playback.

[0052] Step 7: Determine the current state of the digital human, including natural termination, interruption by the user during the response, and no interruption during a normal response; calculate the end flag and reward value based on the digital human state, and add the current state, action value, reward value, and end flag tuple to the trajectory list; and select different steps to execute based on the current state of the digital human: if the digital human state is natural termination of the response, proceed to step 8; if the digital human state is interruption by the user during the response, return to step 4; if the digital human state is no interruption during a normal response, then the voice frame number... Add 1, then return to step five;

[0053] Step 8: Determine if the digital human is idle. If so, update the parameters of the policy neural network based on the data in the trajectory list. After the update is complete, clear the trajectory list and return to step 4. If not, return to step 4.

[0054] In one specific embodiment of the present invention, step three can be implemented as follows:

[0055] Step 3.1: Initialize a value network V and a policy network P, with the following model structure: Figure 2 and Figure 3 As shown, the policy network in a value-policy neural network Sum Value Network A similar network structure can be used, consisting of a backbone neural network and three fully connected layers. The first two fully connected layers are each followed by a ReLU activation function. The backbone neural network is then configured to operate in the current state. As input, a 4352-dimensional feature is generated. The first two fully connected layers adjust the feature dimensions to 1024 and 256 respectively. The last fully connected layer generates different dimensional results from the 256-dimensional feature based on different purposes: in the value network, the value corresponding to the input state is generated based on the 256-dimensional feature. The dimension is 1; in the policy network, the policy probability vector corresponding to the input state is generated based on the 256-dimensional features. , dimension .

[0056] refer to Figure 1 The backbone neural network has a two-branch structure. The first branch takes vector data as input and reduces the dimensionality of the input vector data to 256 dimensions through two fully connected layers. Preferably, the two fully connected layers in the first branch adjust the dimensions of the input vector data to 1024 and 256 respectively. The second branch takes image data as input. Taking a 256×256×3 image as an example, the second branch is composed of a convolutional layer (128×128×32), a max pooling layer (64×64×32), a convolutional layer (32×32×64), a max pooling layer (16×16×64), five residual blocks (16×16×64), and a max pooling layer (8×8×64) connected sequentially. The output of the second branch is unfolded into a one-dimensional vector, concatenated with the one-dimensional vector output from the first branch, and used as the output of the backbone neural network. The parameters of each layer in the second branch can be adaptively adjusted for practical applications.

[0057] Step 3.2: Initialize the trajectory list traj to be empty: .

[0058] Step 3.3: Using the pre-collected offline data, calculate the state using the data from Step 5. The method is used to calculate the state, and the P network is pre-trained with offline data until the model converges, so that the model can output the digital human template fragment selection result without considering the interruption situation.

[0059] In one specific embodiment of the present invention, the current state is calculated in step five. One possible way to achieve this is as follows:

[0060] Step 5.1: At the current moment, the digital human answers the first sentence of the synthesized speech. Frame, obtain the remaining speech feature sequence to be played. , And, obtain the position of the currently used digital human template frame. Update the length of the remaining audio. .

[0061] Step 5.2: Obtain the current user's facial image through the camera. It is used to extract the user's current mental state.

[0062] Step 5.3: Extract user inquiry speech using existing speech emotion recognition models. The emotional feature vector in the image is used to estimate the user's psychological state and serves as one of the sources for predicting the likelihood of interruption. This vector is denoted as... In the same interaction, only one user inquiry voice needs to be calculated. The emotion feature vector in the data.

[0063] Step 5.4: Combine the various information data at the current moment into the current state to determine the direction of the next digital human template frame selection. .

[0064] In one specific embodiment of the present invention, step six can be implemented as follows:

[0065] Step 6.1: Calculate via policy network ,in This is the policy probability vector. The i-th element P represents the probability of choosing the i-th action. This refers to the probability value of choosing to use the template frame forward or backward. Based on... The probability value of each component is randomly selected, and an element is selected according to the probability. Its index It's the final action value you choose. That is, the selection direction of the next template frame:

[0066]

[0067] in, This refers to selecting one of the component indices based on probability for each component of the probability vector.

[0068] Step 6.2: Based on the action value selected in the previous step Select the digital human template frame position for the next frame. :

[0069]

[0070] Using a speech-driven model According to the characteristics of the input speech and template frame To synthesize digital humans , .

[0071] In one specific embodiment of the present invention, step seven can be implemented as follows:

[0072] Step 7.1: Determine the current state of the digital human, including natural termination, interruption by the user during the response process, and no interruption by the user during the normal response process; calculate the end flag and the first reward value based on the digital human state; specifically:

[0073] If the digital human's current response ends naturally, that is, at the digital human template frame position... The corresponding frame is the last frame, assuming that this frame corresponds to the original template. For the j-th frame, record frame number j as the end position and end marker. Set to 1 to calculate the first return value. .

[0074] If the digital human is interrupted by the user during the current moment of answering, i.e., at the current position of the digital human template frame... The corresponding frame is the frame at the time of the interruption, assuming that this frame corresponds to the original template. For the k-th frame, record the frame number. End position, end marker Set to 1 to calculate the first return value. .

[0075] If the digital human is currently responding normally and has not been interrupted by the user, the end signal will be displayed. Set to 0, first reward value Set to 0.

[0076] Step 7.2: Determine the template direction selection at the current moment. Whether the choice made at the previous moment is the same, a second reward value is obtained based on different situations, specifically:

[0077] If the template direction is selected at the current moment If the choice is different from the previous moment, then the second reward value will be... Set to -0.1, otherwise This is to prevent the model's output actions from causing the template frame selection to repeatedly recur around the initial frame, thus keeping the penalty consistently around 0. By adding a penalty for changing the direction of template frame usage, the model can fully utilize the entire template instead of remaining stationary.

[0078] Step 7.3: Calculate the final return value at the current moment. .

[0079] Step 7.4: Combine the current state, action, reward, and end flag tuples. Add it to the track list traj.

[0080] Step 7.5: At any point in an interactive task, there will inevitably be one of the following scenarios: natural termination, interruption by the user during the response, or no interruption during the normal response. Different scenarios correspond to different decisions. Based on the current state of the digital human, select different steps to execute, selectively performing the following operations:

[0081] If the digital human's state is "responds naturally and ends," then the current response is considered complete, and step eight is executed. This involves determining whether it's appropriate to update the parameters of the value-policy neural network at this moment. It's determined that updating the parameters of the value-policy neural network is only appropriate when the digital human is in an idle state. In non-idle states, it's necessary to return to step four to initiate a new interaction. Here, "idle" means there is no new interaction, and the time required to update the parameters of the value-policy neural network is extremely short and will not affect the next interaction.

[0082] If the digital human is interrupted by the user during the answering process, it is considered that the current answer is completed, and the process returns directly to step four to start a new interaction.

[0083] If the digital human is in a normal response state and is not interrupted by the user, then the voice frame number will be... Add 1, then return to step five to continue completing the current answer.

[0084] In one specific embodiment of the present invention, when the digital human is idle and the parameters of the value-policy neural network are updated based on the data in the trajectory list, the parameters of each state are first calculated using the standard GAE (Generalized Advantage Estimation) method based on the trajectory data in traj. In the current value network Advantage value ,in , Let be the length of the trajectory traj; then, the proximal policy gradient optimization (PPO) training method is used to optimize the neural networks P and V respectively using stochastic gradient descent until convergence on the current trajectory data. The aforementioned GAE and PPO are conventional techniques.

[0085] In this embodiment, the value network is optimized first. The process is carried out by minimizing the loss function of the value network:

[0086]

[0087] Further optimize the policy network The process is carried out by minimizing the policy substitution loss:

[0088]

[0089]

[0090] in, The model representing the current moment uses parameters that have already undergone several gradient descent optimizations. , This represents the action selected previously. The strategy model used at that time used the old model parameter values. These two models have different parameter values; The operator took Clip to range Within the size, when insufficient When, let it equal 0.8; when If it exceeds 1.2, set it to 1.2. This represents the probability ratio between the old and new strategies.

[0091] After the update, the policy and value network has been optimized to better select action template frames based on the external environment.

[0092] To verify the effectiveness of this invention, the method of this invention will be compared with a method that directly selects segments based on a preset answer length, and a method that does not require user facial image input and does not make decisions every frame. The training set includes 5000 user question voice recordings and interruption data, including 2000 uninterrupted voice recordings and 3000 question voice recordings and interruption locations. The test set includes 1000 question data, including 800 interrupted questions and 200 uninterrupted questions. The dataset was collected by multiple users in a simulated environment to simulate real-world usage scenarios as closely as possible.

[0093] Discontinuity of action refers to the ratio of the distance between the video frame when the digital human interrupts its current response and the initial static frame to the length of the complete template; continuity = 1 – discontinuity.

[0094] Table 1

[0095]

[0096] As shown in Table 1, it can be seen that, without interruption, directly selecting the segment length based on the preset answer length can obviously achieve perfect continuity. However, user interruptions are unavoidable in practice, and in such cases, the continuity of actions using this simple method becomes very low. The method provided by this invention, however, maintains good continuity even without interruption, and also maintains a high degree of action continuity even when interrupted by the user, thus providing a better digital human interaction experience. This indicates that this invention can fully utilize information from the user's facial image to predict the user's interruption behavior, and makes template usage decisions for each frame, allowing for timely adjustments to template usage based on the user's state, achieving more rational template application and making the digital human's positioning more accurate.

[0097] This embodiment also provides a high-granularity digital human motion generation system based on user intent awareness, which is used to implement the above embodiments. The terms "module," "unit," etc., used below refer to combinations of software and / or hardware that perform a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible.

[0098] This embodiment provides a high-granularity digital human motion generation system based on user intent awareness, including:

[0099] The preprocessing module is used to capture video templates of the digital human's speaking state, initialize the trajectory list and the value-policy neural network for selecting the direction of the template frame, and pre-train the policy network in the value-policy neural network.

[0100] The interaction module is used to receive interaction commands, convert the current user's voice input into a text question, generate the answer text, synthesize speech based on the answer text and extract speech feature sequences, and capture the user's facial image in real time.

[0101] The state calculation module is used to update the feature sequence of the speech to be played and its length, as well as the current digital human template frame position, based on the current speech frame number of the current answer, and to calculate the current state in combination with the user's current facial image information.

[0102] The value-policy neural network module is used to generate an action value representing the direction of template frame selection through the policy network in the value-policy neural network for the current state, obtain the position of the digital human template frame corresponding to the next speech frame number, synthesize the current digital human video frame based on the template frame at the determined digital human template frame position and the current speech features, and play it; and, based on the digital human state at the current moment, record the end flag and the reward value, and add the current state, action value, reward value, and end flag tuple to the trajectory list; the digital human state includes natural end, interruption by the user during the answering process, and no interruption by the user during the normal answering process.

[0103] The loop control module is used to select different operations to perform based on the current state of the digital human: if the digital human's state is "respond naturally and end", it determines whether the digital human is idle. If it is idle, it updates the parameters of the policy neural network based on the data in the trajectory list. After the update is completed, it clears the trajectory list and returns to the interaction module to wait for new interaction instructions. If it is not idle, it directly returns to the interaction module to wait for new interaction instructions.

[0104] If the digital human is interrupted by the user during the response process, it will directly return to the interaction module to wait for new interaction instructions;

[0105] If the digital human is in a normal response state and is not interrupted by the user, the voice frame number is incremented by 1 and the system returns to the status calculation module to continue completing the response.

[0106] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0107] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0108] This invention also provides an electronic device, including a memory and a processor;

[0109] The memory is used to store computer programs;

[0110] The processor is configured to implement, when executing the computer program, the aforementioned method for generating fine-grained digital human actions based on user intent awareness.

[0111] From a hardware perspective, such as Figure 5 The diagram shown is a hardware structure diagram provided in this embodiment. In addition to the processor, memory, network interface and non-volatile memory shown in the diagram, any device with data processing capabilities in which the system is located in the embodiment may also include other hardware depending on the actual function of the device with data processing capabilities, which will not be described in detail here.

[0112] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the above-described method for generating fine-grained digital human actions based on user intent perception.

[0113] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0114] Obviously, the embodiments and accompanying drawings described above are merely some examples of this application. Those skilled in the art can apply this application to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, for those skilled in the art, certain design, manufacturing, or production modifications made based on the technical content disclosed in this application are merely conventional technical means and should not be considered as insufficient disclosure of this application. Several modifications and improvements can be made without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A method for generating highly fine-grained digital human actions based on user intent awareness, characterized in that, include: Step 1: Capture a video template of the digital human speaking state, initialize the trajectory list and the value-policy neural network for selecting the direction of the template frame, and pre-train the policy network in the value-policy neural network; Step 2: Receive interactive instructions, convert the current user's voice input into a text question, generate the answer text, synthesize speech based on the answer text, and extract the speech feature sequence; Step 3: Capture the user's facial image in real time, update the feature sequence of the voice to be played and its length, as well as the current digital human template frame position, based on the voice frame number of the current response, and calculate the current state by combining the user's current facial image information; The calculation of the current state includes: At the current moment, the digital human answers the first question in the synthesized speech. The frame updates the length of the remaining speech based on the remaining speech feature sequence to be played. And, obtain the position of the currently used digital human template frame. ; Capture the current user's facial image ; Extracting emotional feature vectors from user voice input ; Generate input states to determine the template frame selection direction. ; Step 4: For the current state, generate action values ​​representing the selection direction of the template frame through the policy network in the value-policy neural network, obtain the position of the digital human template frame corresponding to the next speech frame number, synthesize the current digital human video frame based on the template frame at the determined digital human template frame position and the current speech features, and play it. Step 5: Based on the current state of the digital human, record the end marker and reward value, and add the current state, action value, reward value, and end marker tuple to the trajectory list; the digital human state includes natural end, interruption by the user during the answering process, and no interruption by the user during the normal answering process; Step 6: Select different operations based on the current state of the digital human: if the digital human's state is that the answer ends naturally, proceed to step 7; if the digital human's state is that the answer was interrupted by the user, return to step 2; if the digital human's state is that the answer was given normally and was not interrupted by the user, increment the voice frame number by 1 and return to step 3. Step 7: Determine if the digital human is idle. If so, update the parameters of the policy neural network based on the data in the trajectory list. After the update is complete, clear the trajectory list and return to step 2. If not, return to step 2.

2. The high-fine-grained digital human motion generation method based on user intent perception according to claim 1, characterized in that, Step 4, which generates action values ​​representing the selection direction of the template frame through the policy network in the value-policy neural network, and obtains the position of the digital human template frame corresponding to the next speech frame number, includes: Calculation via policy network ,in This is the policy probability vector. The i-th element This represents selecting the i-th action. The probability of choosing to use the template frame forward or backward is determined based on the probability value, thereby determining the position of the digital human template frame corresponding to the next voice frame number.

3. The high-fine-grained digital human motion generation method based on user intent perception according to claim 1, characterized in that, Step 5 includes: Step 5.1: Determine the current state of the digital human, and calculate the end flag and first reward value based on the digital human's state; Step 5.2: Determine whether the template frame selection direction at the current moment is the same as that at the previous moment. If not, set the second reward value to -0.1; otherwise, set the second reward value to 0. Step 5.3: The sum of the first return value and the second return value is the total return value.

4. The high-fine-grained digital human motion generation method based on user intent perception according to claim 3, characterized in that, Step 5.1 specifically involves: If the digital human answers naturally at the current moment, record frame number j as the end position, set the end flag to 1, and calculate the first reward value. ; If the digital human is interrupted by the user during an answer, record the frame number. To interrupt the position, the end flag is set to 1, and the first reward value is calculated. ; If the digital human is currently responding normally and has not been interrupted by the user, the end flag will be set to 0 and the first reward value will be set to 0.

5. The high-fine-grained digital human motion generation method based on user intent perception according to claim 1, characterized in that, The parameters of the value-policy neural network that are updated based on data in the trajectory list include: First, calculate the current state, action value, reward value, and end flag tuples in the trajectory list within the current value network. Advantage value ,in , The length of the trajectory list; Next, the training method of proximal policy gradient optimization is used to first optimize the value network. The process is carried out by minimizing the loss function of the value network: ; Further optimize the policy network The process is carried out by minimizing the policy substitution loss: ; ; in, The parameters of the representative value model, Representative value model, The policy model representing the current moment uses parameters that have already undergone several gradient descent optimizations. , Represents the previously selected action The strategy model used at that time used the parameter values ​​of the old model. ; The representative put Clip to range , This represents the probability ratio between the old and new strategies.

6. The high-fine-grained digital human motion generation method based on user intent perception according to claim 1, characterized in that, The value network and policy network in the described value-policy neural network use the same network structure but have different output dimensions in their last layers. The output dimension of the last layer in the value network is 1, and the output dimension of the last layer in the policy network is 2. Each of the value network and policy network includes a two-branch backbone network and several fully connected layers. The first branch of the two-branch backbone network is based on the input state... The concatenated input is used as the second branch, which takes the input state as its input. As input, the two branches extract features from the input data respectively, and then the output features of the two branches are unfolded into one dimension and concatenated as the output of the two-branch backbone network.

7. A high-fine-grained digital human motion generation system based on user intent awareness, used to implement the high-fine-grained digital human motion generation method based on user intent awareness as described in claim 1, characterized in that, The system includes: The preprocessing module is used to capture video templates of the digital human's speaking state, initialize the trajectory list and the value-policy neural network for selecting the direction of the template frame, and pre-train the policy network in the value-policy neural network. The interaction module is used to receive interaction commands, convert the current user's voice input into a text question, generate the answer text, synthesize speech based on the answer text and extract speech feature sequences, and capture the user's facial image in real time. The state calculation module is used to update the feature sequence of the speech to be played and its length, as well as the current digital human template frame position, based on the current speech frame number of the current answer, and to calculate the current state in combination with the user's current facial image information. The value-policy neural network module is used to generate an action value representing the direction of template frame selection through the policy network in the value-policy neural network for the current state, obtain the position of the digital human template frame corresponding to the next speech frame number, synthesize the current digital human video frame based on the template frame at the determined digital human template frame position and the current speech features, and play it; and, based on the digital human state at the current moment, record the end flag and the reward value, and add the current state, action value, reward value, and end flag tuple to the trajectory list; the digital human state includes natural end, interruption by the user during the answering process, and no interruption by the user during the normal answering process. The loop control module is used to select different operations to perform based on the current state of the digital human: if the digital human's state is "respond naturally and end", it determines whether the digital human is idle. If it is idle, it updates the parameters of the policy neural network based on the data in the trajectory list. After the update is completed, it clears the trajectory list and returns to the interaction module to wait for new interaction instructions. If it is not idle, it directly returns to the interaction module to wait for new interaction instructions. If the digital human is interrupted by the user during the response process, it will directly return to the interaction module to wait for new interaction instructions; If the digital human is in a normal response state and is not interrupted by the user, the voice frame number is incremented by 1 and the system returns to the status calculation module to continue completing the response.

8. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the high-fine-grained digital human motion generation method based on user intent awareness as described in any one of claims 1 to 6.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, are used to implement the fine-grained digital human motion generation method based on user intent awareness as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Unified access method applying digital human code chain

    CN112669165A

  • Construction method and application of digital human cardiovascular system based on hemodynamics

    CN113160921A