A voice assistant avatar generation method

Abstract

The application belongs to the technical field of artificial intelligence interaction, and provides a voice assistant image generation method, in the adaptation of candidate shapes and feature components, a structured keyword set and a style tendency vector indicating user intention are obtained by analyzing user language instructions, and then the comprehensive matching degree is calculated to obtain personalized adaptation; the action resources based on the safety level classification are associated, and the virtual image object that can be interacted is output, the current driving situation is analyzed by fusion to control the interaction, the scene performance is effectively strengthened, and high safety constraints are realized; the lightweight real-time rendering and driving based on the Unity engine can meet the low computing power requirement; in this way, a full automation, full quantization driving, rule and parameter determination technical closed loop from "user voice input" to "screen interaction image output" is constructed.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence interaction technology, specifically relating to a method for generating a voice assistant image. Background Technology

[0002] Currently, smart cockpits commonly incorporate virtual avatars for in-vehicle voice assistants to enhance the interactive experience, such as the industry's typical fixed-avatar solutions (like NIO's NOMI). However, limitations still exist in terms of personalization and intelligence, as follows: 1) Static and fixed image, unable to be personalized: Using a fixed model, all users face the same image, and it is impossible to create diverse and personalized images in real time based on the user's voice description; 2) Weak interactive expressiveness and disconnect from the scene: The voice assistant's actions and expressions rely on a limited set of pre-made animation sequences, which cannot vividly match the rich driving scenes and voice interaction content; 3) The contradiction between the generality of technology and the specificity of the scenario: General 3D generation technology has not been optimized for the core constraints of the vehicle environment and cannot meet the constraints such as low computing power, interactive safety, Q version stylization, and full automation. Summary of the Invention

[0003] The purpose of this invention is to disclose a method for generating a voice assistant image, which solves the technical problems that the virtual images of current in-vehicle voice assistants are not only static and fixed with weak scene expression, but also difficult to meet the constraints of low computing power and high security.

[0004] To achieve the above objectives, this invention discloses a method for generating a voice assistant image, comprising: The system collects user voice commands, then quantifies and analyzes the user's intent based on a natural language processing model, and outputs structured generated commands, which include a structured keyword set and a style tendency vector. Based on the structured generation instructions, the comprehensive matching degree between candidate shapes and user intent is traversed to determine the basic shape; based on the structured generation instructions, the comprehensive matching degree between feature components and user intent is traversed to determine the final set of feature components by combining mutual exclusion group filtering. Based on the basic shape and feature component set, parametric assembly is performed, and stylization processing is carried out in combination with the style tendency vector to output the target image model; A preset skinning algorithm is used to adapt the target image model to a preset skeletal template, and then associate it with motion resources based on safety level classification to output an interactive virtual image object, and perform lightweight real-time rendering and driving.

[0005] This basic solution adapts candidate shapes and feature components by analyzing user language commands to obtain a structured set of keywords and style tendency vectors representing user intent, and then calculating the comprehensive matching degree to obtain personalized adaptation. It associates action resources based on safety level classification to output interactive virtual avatar objects, and performs interactive control by integrating and analyzing the current driving situation, effectively enhancing the scene-based expressiveness and achieving high safety constraints. The lightweight real-time rendering and driving based on the Unity engine can meet the low computing power requirements. In this way, a fully automated, fully quantitative driving, and rule and parameter determination technical closed loop is constructed from "user voice input" to "screen interactive avatar output".

[0006] As an optional implementation, user voice commands are collected, and then the user's intent is quantified and parsed based on a natural language processing model, outputting structured generated commands, including: Collect user voice commands and convert them into text commands; Based on the natural language processing model, the text instructions are semantically parsed, relevant nouns are extracted from them according to the user's intent as a set of structured keywords, and style tendency vectors are determined based on relevant adjectives; The output contains a structured set of keywords and a style preference vector, which are structured generation instructions.

[0007] This solution is based on a natural language processing model. It performs semantic parsing on the text commands converted from user voice commands, extracts relevant nouns as a set of structured keywords according to the user's intent, and determines style tendency vectors based on relevant adjectives. In this way, the core objects of user needs are clearly identified through structured keywords, and the abstract description is quantified by style vectors, which enhances the voice assistant's ability to understand and respond to complex or ambiguous expressions.

[0008] As an optional implementation, based on the structured generation instructions, the comprehensive matching degree between candidate shapes and user intent is traversed to determine the basic shape, including: Traverse the preset candidate shapes in the structured component library to obtain the morphological constraints of the candidate shapes. The morphological constraints of the candidate shapes include a set of semantic tags, a preset style vector, and the number of polygon faces. Based on the morphological constraints of the candidate shapes and the structured generation instructions, a comprehensive analysis is performed to calculate the overall matching degree between the candidate shapes and the user's intent. A relative optimization strategy is adopted to select the candidate shape with the highest overall matching degree. It is then determined whether the overall matching degree meets the minimum threshold requirement. If it does, it is determined to be the basic shape; otherwise, the default shape is selected as the basic shape.

[0009] This solution calculates the comprehensive matching degree between candidate shapes and user intent based on three morphological constraints: semantic tag set, preset style vector, and polygon count. On the one hand, through quantitative processing, it transforms vague user intent into a calculable score for optimal matching, achieving precise alignment between subjective description and objective 3D model. On the other hand, it incorporates polygon count analysis to adapt to low-computing-power in-vehicle scenarios. A minimum threshold requirement for the comprehensive matching degree is set to ensure that the output results meet the basic validity requirements of user intent, thereby improving system reliability.

[0010] As an optional implementation, based on the structured generation instructions, the comprehensive matching degree between feature components and user intent is traversed, and then the final set of feature components is determined by combining the filtering within mutually exclusive groups, including: Traverse the preset feature components in the structured component library to obtain the morphological constraints of the feature components. The morphological constraints of the feature components include a semantic tag set, a preset style vector, and the number of polygon faces. Based on the morphological constraints of the feature components and the structured generation instructions, a comprehensive analysis is performed to calculate the overall matching degree between the feature components and the user intent. Based on the dynamic matching threshold, each comprehensive matching degree is traversed, and the feature components that meet the user intent benchmark are selected and saved to a temporary candidate set. Determine whether the number of feature components in the temporary candidate set is greater than a preset threshold. If so, directly output the temporary candidate set as the matching candidate set. Otherwise, perform adaptive threshold adjustment and return to the temporary candidate set for filtering. Perform mount point compatibility verification to filter out incompatible feature components from the mount points in the matching candidate set; Based on the mutual exclusion group screening, the feature components that conflict in the matching candidate set are selected and saved; then the occlusion order is determined according to the rendering order, resulting in the final feature component set.

[0011] The beneficial effects of this plan are as follows: Based on three morphological constraints—semantic tag set, preset style vector, and polygon count—the comprehensive matching degree between feature components and user intent is calculated. Through quantitative processing, the ambiguous user intent is transformed into a calculable score for optimal adaptation, achieving precise alignment between subjective description and objective 3D model. By simultaneously incorporating polygon count analysis, it adapts to low-computing-power in-vehicle scenarios. By adjusting the preset threshold and the adaptive threshold, a temporary candidate set is obtained. This is achieved by determining the number of components in the feature component, thereby improving the visual recognizability and interest of the generated assistant image. By sequentially performing mount point compatibility checks and mutual exclusion group filtering, the rationality of multi-feature component assembly is further verified to ensure the controllability and style consistency of the image generation instructions.

[0012] As an optional implementation, the adaptive threshold adjustment specifically involves: dynamically reducing the dynamic matching threshold according to a preset step size, determining whether the reduced dynamic matching threshold is greater than the lower threshold, and if so, outputting a new dynamic matching threshold, until the number of components is greater than the preset threshold, or until the dynamic matching threshold is equal to the lower threshold.

[0013] This solution sets a lower limit for the dynamic matching threshold, adjusts it according to a preset step size, and automatically controls data filtering. It gradually expands the search range while ensuring quality, and also ensures that the feature components are adapted to the user's intent.

[0014] As an optional implementation, based on screening within mutually exclusive groups, the feature components that conflict in the matching candidate set are preferentially saved, specifically: The feature components mounted on the same node in the matching candidate set are divided into several mutually exclusive groups; Iterate through each mutually exclusive group and retain the feature component with the highest overall matching degree; When the difference in the overall matching degree of the feature components within the mutually exclusive group is less than the matching threshold, the style similarity between the feature component and the style tendency vector is calculated, and the feature component with the highest style similarity is retained; wherein, when the overall matching degree of the feature components within the mutually exclusive group is the same, the feature component with higher priority is retained.

[0015] This solution employs a three-tiered approach: comprehensive matching to select the overall optimal match, style similarity to refine preference matching, and priority-based conflict resolution. This progressive approach ensures the physical rationality of feature component assembly and fully considers user intent, style consistency, and importance to accurately output a personalized language assistant image, thereby improving the overall quality of semantic matching and visual presentation.

[0016] As an optional implementation, a pre-built structured component library is also included. The structured component library includes a basic shape library and a feature component library. The basic shape library includes at least two different candidate shapes, and the feature component library includes several local feature components attached to the candidate shapes.

[0017] This solution uses a structured component library to pre-set several candidate shapes and feature components as the basis for personalized image generation. When applying the solution, the pre-set models can be directly called without the need to generate geometry in real time, resulting in fast response and controllable results.

[0018] As an optional implementation, the target image model is output by parametrically assembling the basic shape and feature component set, combining it with the style tendency vector for stylization processing, including: Using the basic shape as the supporting skeleton, the component pose of each feature component in the feature component set in the local coordinate system on the basic shape is calculated, and the assembly model is obtained by parametric assembly. Using the style tendency vector as the quantization input for style control, the corresponding set of stylization rules is mapped out, and then the assembled model is stylized according to the set of stylization rules to output the target image model.

[0019] This solution ensures the reasonable splicing of basic shapes and feature components through parametric assembly, and makes stylistic adjustments by combining style tendency vectors. The output target image model has a stable structure that adapts to the user's personalized needs and meets the requirements of Q version stylization.

[0020] As an optional implementation, by associating motion resources based on security level classification, an interactive virtual avatar object is output, including: Pre-set motion resources corresponding to different safety level categories to build a dedicated vehicle animation library; Construct a context-resource mapping rule table to define the triggering logic for vehicle status and interaction actions; The current vehicle status is obtained, and then the corresponding motion resources are obtained from the vehicle-specific animation library based on the context-resource mapping rule table. These resources are then associated with the target image model to output an interactive virtual image object.

[0021] This solution obtains the current vehicle status, then retrieves the corresponding motion resources from the vehicle-specific animation library based on the context-resource mapping rule table, associates them with the target image model, and outputs an interactive virtual image object. Based on the vehicle status, it determines the current vehicle safety requirements, forming a complete technical closed loop from user needs to safe interaction. This achieves deep adaptation between technology and scenario, and is highly practical.

[0022] As an optional implementation method, lightweight real-time rendering and driving are performed, including: Real-time rendering: The virtual avatar object is input, and the vertex positions are calculated using the Unity URP pipeline and the linear blending skin algorithm, and then rendered using GPU acceleration. Context-driven: The script queries the vehicle status every frame, matches the context-resource mapping rule table to trigger interactive actions, drives animation fusion through the Animator component, and outputs interactive screens.

[0023] This solution is based on lightweight real-time rendering and driving of the Unity engine, which can smoothly present low-polygon images on mobile and web devices. While ensuring the expressiveness of the images, it achieves low resource consumption, high real-time response and consistent experience across multiple devices, thus meeting the lightweight deployment requirements of the language assistant. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the 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.

[0025] Figure 1 This is a flowchart of a voice assistant image generation method provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] In this invention, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing the invention and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0028] Furthermore, in addition to indicating direction or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain situations to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0029] Furthermore, the terms "installation," "setup," "equipped with," "connection," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this invention based on the specific circumstances.

[0030] Furthermore, the terms "first," "second," etc., are primarily used to distinguish different devices, components, or parts (which may be the same or different in specific type and construction), and are not intended to indicate or imply the relative importance or quantity of the indicated devices, components, or parts. Unless otherwise stated, "a plurality of" means two or more.

[0031] The technical solution of the present invention will be further described below with reference to the embodiments and accompanying drawings.

[0032] Please see Figure 1 As shown in the figure, this application provides a method for generating a voice assistant image, including: S1. Collect user voice commands, then quantify and analyze the user's intent based on a natural language processing model, and output structured generated commands, including: Collect user voice commands and convert them into text commands; Based on the natural language processing model, the text instructions are semantically parsed, relevant nouns are extracted from them according to the user's intent as a set of structured keywords, and style tendency vectors are determined based on relevant adjectives; The output contains a structured set of keywords and a style preference vector, which are structured generation instructions.

[0033] Specifically: Input: User's original voice audio signal Quantification and Rules: 1) Speech to Text: Converts speech into text commands using the vehicle's ASR (Automatic Speech Recognition) engine; 2) Semantic and Style Analysis: Text is processed using a Natural Language Processing (NLP) model, outputting: Structured keyword set : Collect nouns with clear physical entities or defined identity attributes (such as animal categories, occupational categories, wearable parts), for example ; Style tendency vector S: This vector collects adjectives (e.g., cute, cool) that describe the visual presentation and emotional atmosphere. Style vector S is an n-dimensional vector, with each component... Used to quantify the intensity of a user's style preference; the specific value of n can be set according to the actual application scenario (e.g., n=2), and it supports dynamic expansion through OTA upgrades.

[0034] For example .

[0035] Output: Structured generation instructions .

[0036] This embodiment is based on a natural language processing model. It performs semantic parsing on the text commands converted from user voice commands, extracts relevant nouns as a set of structured keywords according to the user's intent, and determines style tendency vectors based on relevant adjectives. In this way, the core objects of user needs are clearly identified through structured keywords, and the abstract description is quantified by style vectors, which enhances the voice assistant's ability to understand and respond to complex or ambiguous expressions.

[0037] S2. Based on the structured generation instructions, traverse the comprehensive matching degree between candidate shapes and user intentions to determine the basic shape; based on the structured generation instructions, traverse the comprehensive matching degree between feature components and user intentions, and then combine the filtering within mutually exclusive groups to determine the final set of feature components. In this embodiment, a pre-set structured component library is also included. The structured component library includes a basic shape library and a feature component library. The basic shape library includes at least two different candidate shapes, and the feature component library includes several local feature components attached to the candidate shapes.

[0038] Specifically, the pre-built hierarchical structured component library includes two layers: a basic shape library and a feature component library, as follows: Basic Shape Library: Pre-sets a variety of complete Q-version character candidate shapes, including but not limited to human and animal shapes (such as fox, cat, and bear). Each candidate shape is pre-bound with basic skeleton and skinning data and includes predefined fields: part ID, semantic tag set. Style vectors Number of faces (number of polygons) ; Feature Component Library: Pre-defined attachable local accessories (such as masks, hats, accessories, and eye effects) as feature components. In addition to the fields mentioned above, each feature component also defines the mounting parent node ID, mutex group ID, rendering order, and component priority, as follows: Attach parent node ID (string): Specifies the name of the basic shape skeleton node to which the component should be attached, such as "Head_Face" or "Head_Top", which is used for parent positioning during subsequent spatial assembly; Mutex group ID (string): Identifies the mutex group to which the component belongs. Components with the same value cannot exist simultaneously on the same mounted parent node (e.g., face mask and sunglasses are mutually exclusive in the "face_mask" group), used for conflict resolution; Rendering order (integer): The larger the value, the earlier the rendering order (i.e., covering the components behind it), which is used to avoid visual penetration. The default value is 0. Component priority (integer): The smaller the value, the higher the priority in conflict resolution (such as when scores are the same within a mutual exclusion group). The default value is 5.

[0039] Among them, the structured component library supports OTA (over-the-air) incremental update mechanism, which can push new feature component data to the cloud regularly to continuously expand the coverage of semantic tags (such as adding feature components with styles such as "mecha", "magic", and "cyberpunk"), ensuring that the system can respond to users' increasingly diverse generation needs.

[0040] This embodiment uses a structured component library to pre-set several candidate shapes and feature components as the basis for personalized image generation. When applying, the pre-set model can be directly called without real-time geometry generation, resulting in fast response and controllable results.

[0041] In this embodiment, based on the structured generation instructions, the comprehensive matching degree between candidate shapes and user intent is traversed to determine the basic shape, including: A1. Traverse the preset candidate shapes in the structured component library and obtain the morphological constraints of the candidate shapes. The morphological constraints of the candidate shapes include a set of semantic tags, a preset style vector, and the number of polygon faces. A2. Based on the morphological constraints of the candidate shapes and the structured generation instructions, a comprehensive analysis is performed to calculate the overall matching degree between the candidate shapes and the user intent. The calculation formula is as follows:

[0042] In the formula, Representing candidate shapes A pre-defined set of semantic tags in the component library; represents the Jaccard similarity coefficient (quantifies semantic overlap, threshold [0,1]); Representing candidate shapes Preset style vectors; Cosine similarity is represented by a threshold [-1, 1], which is achieved by... Normalization maps to [0,1] to eliminate the effect of negative values. Representing candidate shapes The number of polygon faces; Represents the preset candidate shapes Allow a maximum face count threshold (e.g., 5000, used to normalize the complexity penalty term). , , Represents the corresponding candidate shape The weighting coefficients for the semantic tag set, preset style vector, and polygon count are determined using the Analytic Hierarchy Process (AHP) or multiple linear regression based on user preference questionnaires. This ensures a balance between semantic accuracy, style preference, and computational constraints, and satisfies [the requirements of the system]. + + =1.

[0043] A3. Using a relative optimization strategy, select the candidate shape with the highest overall matching degree, and determine whether its overall matching degree meets the minimum threshold requirement. If it does, determine it as the basic shape; otherwise, select the default shape as the basic shape.

[0044] Specifically, a relative optimization strategy is adopted, and a minimum threshold is set. (e.g., 0.3); If a comprehensive match exists Select comprehensive matching degree The highest-ranking candidate shape is selected; otherwise, the default shape (such as a standard chibi-style humanoid figure) is selected.

[0045] This embodiment calculates the comprehensive matching degree between candidate shapes and user intent based on three morphological constraints: semantic tag set, preset style vector, and polygon count. On the one hand, through quantitative processing, the ambiguous user intent is transformed into a calculable score for optimal matching, achieving precise alignment between subjective description and objective 3D model. On the other hand, by simultaneously incorporating polygon count analysis, it adapts to low-computing-power scenarios in vehicles. A minimum threshold requirement for the comprehensive matching degree is set to ensure that the output results meet the basic validity requirements of user intent, thereby improving system reliability.

[0046] In this embodiment, based on the structured generation instructions, the comprehensive matching degree between feature components and user intent is traversed, and then the final set of feature components is determined by combining the filtering within mutually exclusive groups, including: B1. Traverse the preset feature components in the structured component library and obtain the morphological constraints of the feature components. The morphological constraints of the feature components include a semantic tag set, a preset style vector, and the number of polygon faces. B2. Based on the morphological constraints of the feature components and the structured generation instructions, a comprehensive analysis is performed to calculate the overall matching degree between the feature components and the user intent. The calculation formula is as follows:

[0047] In the formula, Indicates feature component A pre-defined set of semantic tags in the component library; represents the Jaccard similarity coefficient (quantifies semantic overlap, threshold [0,1]); Indicates feature component Style vectors; Cosine similarity is represented by a threshold [-1, 1], which is achieved by... Normalization maps to [0,1] to eliminate the effect of negative values. Indicates feature component The number of polygon faces; Indicates the preset feature component Allow a maximum face count threshold (e.g., 5000, used to normalize the complexity penalty term). , , This represents the weighting coefficients of the semantic label set, preset style vector, and polygon count in the corresponding feature component p. These weighting coefficients can be calibrated using the Analytic Hierarchy Process (AHP) or multiple linear regression based on user preference questionnaires to ensure a balance between semantic accuracy, style preference, and computational constraints, thus satisfying... + + =1. For example , , This indicates that semantic matching has the highest weight, followed by style, while the number of faces has a relatively small impact.

[0048] B3. Based on the dynamic matching threshold, traverse each comprehensive matching degree, filter out the feature components that meet the user intent benchmark, and save them to a temporary candidate set; B4. Determine whether the number of feature components in the temporary candidate set is greater than a preset threshold. If so, directly output the temporary candidate set as the matching candidate set. Otherwise, perform adaptive threshold adjustment and return to the temporary candidate set for filtering. In this embodiment, the adaptive threshold adjustment specifically involves: dynamically reducing the dynamic matching threshold according to a preset step size, determining whether the reduced dynamic matching threshold is greater than the lower threshold, and if so, outputting a new dynamic matching threshold, until the number of components is greater than the preset threshold, or until the dynamic matching threshold is equal to the lower threshold.

[0049] Specifically: A) Set the initial threshold for dynamic matching. (e.g., 0.7), filter out all ≥ The components are denoted as the temporary candidate set. .

[0050] B) If the temporary candidate set Number of components Then the temporary candidate set will be as a matching candidate set ; If temporary candidate set Number of components ( The threshold can be dynamically adjusted based on the richness of the component library (e.g., 3 components). The threshold will then be lowered using the following formula, and the selection will be re-filtered after each lowering. ≥ The components update the temporary candidate set. until or :

[0051] In the formula, This represents the step size (e.g., 0.2), and k is the number of descents. This indicates the lower limit of the threshold (e.g., 0.3).

[0052] like hour, Then the current candidate set as a matching candidate set The threshold will not be lowered further, thus preventing the infinite lowering of the threshold from leading to the selection of low-quality parts.

[0053] This embodiment sets a lower limit for the dynamic matching threshold, adjusts it according to a preset step size, and automatically controls data filtering. It gradually expands the search range while ensuring quality, and also ensures that the feature components are adapted to the user's intent.

[0054] B5. Perform mount point compatibility verification and filter out incompatible feature components in the mount point set of the matching candidate set; Specifically, iterate through the candidate matching set. Check if the basic shape contains the mounting node required for the feature component (such as "Head_Face"). If it does not exist, remove the feature component from the set.

[0055] B6. Based on the screening within mutually exclusive groups, select and save the feature components that conflict in the matching candidate set, specifically: The feature components mounted on the same node in the matching candidate set are divided into several mutually exclusive groups; Iterate through each mutually exclusive group and retain the feature component with the highest overall matching degree; When the difference in the overall matching degree of the feature components within the mutually exclusive group is less than the matching threshold, the style similarity between the feature component and the style tendency vector is calculated, and the feature component with the highest style similarity is retained; wherein, when the overall matching degree of the feature components within the mutually exclusive group is the same, the feature component with higher priority is retained.

[0056] This embodiment ensures the physical rationality of feature component assembly by comprehensively selecting the overall optimal match, refining preference matching through style similarity, and resolving conflicts through three layers of priority. It also fully considers user intent, style consistency, and importance to accurately output a personalized language assistant image, thereby improving the overall quality of semantic matching and visual presentation.

[0057] B7. Determine the occlusion sorting based on the rendering order, resulting in the final set of feature components.

[0058] Specifically, conflict resolution is performed in the following order: first, filtering is performed within the mutual exclusion group to determine the components to be retained; then, the retained components are sorted by rendering order field for occlusion.

[0059] Mutual exclusion rule: If multiple feature components belong to the same mutual exclusion group (identified by the mutual exclusion group ID in the feature component library) and are mounted on the same node, only the one with the highest overall matching degree will be retained; If the score difference between the highest and second highest scores in a mutually exclusive group is less than the matching threshold δ (e.g., 0.1), then feature components with higher style vector cosine similarity to user style vector S are preferred. If the scores are the same within a mutually exclusive group, the component with the highest priority is retained according to the feature component priority; if the priorities are also the same, the first one is retained according to the feature component ID.

[0060] Occlusion sorting: The rendering order is determined by the rendering order field of the feature components to avoid visual penetration.

[0061] Output: A set of basic shapes and final feature components. ,like If empty, skip the feature component attachment step and only output the basic shape.

[0062] The beneficial effects of this embodiment are as follows: Based on three morphological constraints—semantic tag set, preset style vector, and polygon count—the comprehensive matching degree between feature components and user intent is calculated. Through quantitative processing, the ambiguous user intent is transformed into a calculable score for optimal adaptation, achieving precise alignment between subjective description and objective 3D model. By simultaneously incorporating polygon count analysis, it adapts to low-computing-power in-vehicle scenarios. By adjusting the preset threshold and the adaptive threshold, a temporary candidate set is obtained. This is achieved by determining the number of components in the feature component, thereby improving the visual recognizability and interest of the generated assistant image. By sequentially performing mount point compatibility checks and mutual exclusion group filtering, the rationality of multi-feature component assembly is further verified to ensure the controllability and style consistency of the image generation instructions.

[0063] S3. Parametric assembly based on the basic shape and feature component set, including: Using the basic shape as the supporting skeleton, the component pose of each feature component in the feature component set in the local coordinate system on the basic shape is calculated, and the assembly model is obtained by parametric assembly. Using the style tendency vector as the quantization input for style control, the corresponding set of stylization rules is mapped out, and then the assembled model is stylized according to the set of stylization rules to output the target image model.

[0064] Specifically: Input: Basic shape and final feature set

[0065] Quantification and Rules: 1) Spatial Assembly: Based on the feature component connection relationship rule table (fields: component ID, parent component ID, local transformation matrix) Assembly is performed, and the world matrix of the feature components is constructed. The result is calculated recursively as follows:

[0066] In the formula, Represents the world matrix of the parent node; It is represented as a local transformation matrix of the feature component (containing position, rotation, and scaling information).

[0067] This embodiment ensures the reasonable splicing of basic shapes and feature components through parametric assembly, and performs stylistic adjustments by combining style tendency vectors. The output target image model has a stable structure that adapts to the user's personalized needs and meets the requirements of Q version stylization.

[0068] S4. Perform stylization processing based on the style tendency vector to output the target image model; The Q-version stylization rule engine is as follows: Rule R1 (Head Enlargement): Apply a scaling factor to the head. =1.1+ , Indicates the amount of cute style content. This indicates the cuteness adjustment factor; Rule R2 (Exaggerated Facial Features): Apply a scaling factor to the eyes. =1.0+ ; Rule R3 (Contour Rounding): Applyes a subdivision surface algorithm to smooth the model. To control the impact of face count growth on onboard computing power, a face count pre-check is required before execution: Let the current model face count be... Detailed multiplication factor =4 (for Catmull-Clark subdivisions), if If the iteration count n is reduced to 1, then the subdivision is skipped if the limit is still exceeded after reducing it to 1. Rule R4 (Adaptive scaling) calculates the torso scaling factor. :

[0069]

[0070]

[0071] In the formula, Indicates the current head-to-body ratio. Indicates the target head-to-body ratio. Indicates head height, Indicates altitude, This indicates the default head-to-body ratio. This indicates the influence coefficient of the cute style. Scaling is applied only to the torso part along the Y-axis (in the chibi style). (The torso is shortened, resulting in a large head and small body).

[0072] Output: The final chibi-style model after stylization is used as the target character model.

[0073] S5. Using a preset skinning algorithm, the target image model is adapted to a preset bone template. The final vertex position calculation formula is as follows:

[0074] In the formula, j is the skeleton index; Indicates the first The inverse matrix of the binding pose of the root skeleton; Indicates the first The current frame transformation matrix of the root skeleton; Represents the vertex coordinates under the bound pose; Represents vertices Bone The skin weights that are affected satisfy .

[0075] The preset skinning algorithms include the linear hybrid skinning algorithm.

[0076] S6. Associate action resources based on security level classification and output interactive virtual avatar objects, including: Pre-set motion resources corresponding to different safety level categories to build a dedicated vehicle animation library; Construct a context-resource mapping rule table to define the triggering logic for vehicle status and interaction actions; The current vehicle status is obtained, and then the corresponding motion resources are obtained from the vehicle-specific animation library based on the context-resource mapping rule table. These resources are then associated with the target image model to output an interactive virtual image object.

[0077] In this embodiment, the vehicle-specific animation library includes motion resources categorized based on safety levels, such as: Level 1: Micro-action sequences, such as ACT_NOD (head nod) and ACT_IDLE (breathing standby); Level 2: Facial expression status, such as EXP_HAPPY (happy) or EXP_LISTENING (listening).

[0078] Context-resource mapping rule table: Defines the triggering logic for vehicle status and actions (fields: vehicle signal, voice status, mapped action ID), for example, IF(vehicle speed>60km / h AND voice status=end of broadcast)THEN Trigger(ACT_SAFE_IDLE).

[0079] Output: An interactive virtual avatar object.

[0080] This embodiment obtains the current vehicle status, then retrieves the corresponding motion resources from the vehicle-specific animation library based on the context-resource mapping rule table, associates them with the target image model, outputs an interactive virtual image object, determines the current vehicle safety requirements based on the vehicle status, and forms a complete technical closed loop from user needs to safe interaction, achieving deep adaptation between technology and scenario, and is highly practical.

[0081] S7 performs lightweight real-time rendering and driving, including: Real-time rendering: The virtual avatar object is input, and the vertex positions are calculated using the Unity URP pipeline and the linear blending skin algorithm, and then rendered using GPU acceleration. Context-driven: The script queries the vehicle status (such as vehicle speed, voice status and other vehicle bus signals) every frame, matches the context-resource mapping rule table to trigger interactive actions, drives the animation fusion through the Animator component, and outputs the interactive screen, that is, outputs the real-time interactive screen of the Q version voice image on the screen.

[0082] This embodiment is based on lightweight real-time rendering and driving of the Unity engine, which can smoothly present low-polygon images on mobile and web devices. While ensuring the expressiveness of the images, it achieves low resource consumption, high real-time response and consistent experience across multiple devices, thus meeting the lightweight deployment requirements of the language assistant.

[0083] Taking a user request to generate a "cute fox ninja" as an example, the steps are as follows: I. Preset Data and Parameter Definitions 1. System preset parameters Weighting coefficients: =0.5 (semantic), =0.3 (style) =0.2 (complexity); Threshold: =0.3 (as the minimum threshold for validity). =0.7 (initial threshold for dynamic matching threshold). =3 (minimum number of matching parts); Stylization parameters: =1.1 (head magnification factor). =0.3 (cuteness adjustment coefficient) =0.5 (default head-to-body ratio). =0.2 (Cute style influence coefficient).

[0084] 2. User input Keyword set K: {Animal: Fox, Occupation: Ninja} (Classification criteria: Fox is a noun) (The term "body" is a professional noun referring to ninjas.) Style vector S: In this example, n=2: {Cute: 1.0, Handsome: 0.0} (Classification basis: Cute is an adjective).

[0085] 3. Data from the structured component library Candidate shape B1 (fox base): Tag {fox}, Style vector {1.0, 0.0}, Number of faces =2000; Candidate shape B2 (wolf base): Tag {wolf}, Style vector {0.0, 1.0}, Number of faces =4500; Feature component C1 (Ninja mask): Tag {Occupation: Ninja}, Style vector {Cute: 0.9}, Attached parent node ID {Head_Face}; Feature component C2 (Ninja headband): Tag {Occupation: Ninja}, Style vector {Cute: 0.8}, Attached parent node ID {Head_Top}; Feature component C3 (Ninja Sword): Tag {Occupation: Ninja}, Style Vector {Cute: 0.7} Mount the parent node ID {Head_Sheath} (this node does not exist in Fox Base).

[0086] II. Step-by-step calculation process

[0087] 1. Basic shape selection: Candidate shape B1 (fox): Semantic matching degree: = ; Style matching: After normalization, it becomes 1; Complexity factor: 1- =0.6; Overall Score: =0.585.

[0088] Candidate shape B2 (wolf): Semantic matching degree: = ; Style matching: After normalization, it becomes 0.5; Complexity factor: 1- =0.1; Overall Score: =0.17.

[0089] Select logic verification: It meets the requirements; It was removed; Conclusion: Candidate shape B1 (fox basic) was correctly selected.

[0090] 2. Feature component selection and conflict resolution: Initial screening: Inputting the keyword "ninja", retrieved parts C1 (score 0.85), C2 (score 0.80), and C3 (score 0.75), forming a candidate matching set. ; Dynamic threshold calculation: initial threshold =0.7, number of matching parts 3 (For example, 3), there is no need to lower the threshold. .

[0091] Mount point verification: Feature component C1 needs to be mounted with "Head_Face" (which exists in the Fox base), and should be retained; Feature component C2 needs to be mounted with "Head_Top" (which exists in the Fox base), and should be retained; Feature component C3 needs to be mounted with "Back_Sheath" (this node is not available in the Fox base), so it is removed.

[0092] Conflict resolution: Feature component C1 and feature component C2 belong to different mutually exclusive groups (the mask and headband do not conflict), and there is no obstruction problem; Final component set: .

[0093] III. Parametric Assembly and Stylization Space assembly: World matrix calculation for feature component C1: Assume the world matrix of the fox head node "Head_Face" is as follows: (Identity matrix); The local transformation matrix of feature component C1 is:

[0094] (If the position offset is (0.1, 0.05, 0.02)), then:

[0095] The world matrix of feature component C2 is calculated similarly to ensure correct mounting.

[0096] Stylization calculus: R1 head magnification: =1.1+0.3 =1.4 (head magnified to 140%) R2 has exaggerated facial features: =1.0+0.5 =1.5 (eyes magnified to 150%) R3 contour rounding: Current model face count =2000, if two subdivisions are performed, the expected number of faces is 2000 × 4² = 32000. (5000), therefore it is automatically reduced to 1 subdivision, and 1 iteration is actually performed; R4 Ratio Adjustment: Current ratio: = ; Target ratio: =0.5 + 0.2 × 1.0 = 0.7; Torso scaling factor: = ; Local transformation: The torso is scaled by 0.71 along the Y-axis, resulting in a new torso height. =2.0 × 0.71 = 1.42, head Maintain height: =1.0; Final proportions: (Conforms to Q-version characteristics).

[0097] IV. Adapt the animation resources to the target character model and fit them to a standard bipedal skeleton, calculating LBS (Linear Blended Skin) weights (e.g., the head apex is affected by the "Head" skeleton, weights are calculated accordingly). Influenced by the "Neck" skeleton, weight ).

[0098] Resource association: Current vehicle status: speed 0km / h, voice interaction ended; Lookup table mapping: IF(vehicle speed=0 AND voice status=end)THEN Trigger(ACT_NOD) +EXP_HAPPY (triggers a nodding and happy emoticon).

[0099] V. Real-time rendering and driving Real-time rendering: Unity URP pipeline calculates vertex positions, GPU accelerates model rendering; Context-driven: The script queries the vehicle bus signal (vehicle speed 0), matches the rule table to trigger the ACT_NOD action, and drives the animation fusion through the Animator component; Output: The screen displays a picture of a cute chibi fox wearing a mask and headband nodding in greeting.

[0100] In this way, users can obtain a personalized image in real time through natural voice, completing the essential evolution of in-vehicle voice image from "fixed decoration" to "generative partner", and achieving a breakthrough from "one face for all" to "one face for all".

[0101] In adapting candidate shapes and feature components, this invention analyzes user language commands to obtain a structured set of keywords and style tendency vectors representing user intent, and then calculates the comprehensive matching degree to obtain personalized adaptation. It associates action resources based on safety level classification to output interactive virtual avatar objects, and performs interactive control through fusion analysis of the current driving situation, effectively enhancing the scene-based expressiveness and achieving high safety constraints. Lightweight real-time rendering and driving based on the Unity engine can meet low computing power requirements. Thus, a fully automated, fully quantitative driving, and rule- and parameter-determined technical closed loop is constructed from "user voice input" to "screen interactive avatar output."

[0102] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for generating a voice assistant avatar, characterized in that, include: The system collects user voice commands, then quantifies and analyzes the user's intent based on a natural language processing model, and outputs structured generated commands, which include a structured keyword set and a style tendency vector. Based on the structured generation instructions, the comprehensive matching degree between candidate shapes and user intent is traversed to determine the basic shape; based on the structured generation instructions, the comprehensive matching degree between feature components and user intent is traversed to determine the final set of feature components by combining mutual exclusion group filtering. Based on the basic shape and feature component set, parametric assembly is performed, and stylization processing is carried out in combination with the style tendency vector to output the target image model; A preset skinning algorithm is used to adapt the target image model to a preset skeletal template, and then associate it with motion resources based on safety level classification to output an interactive virtual image object, and perform lightweight real-time rendering and driving.

2. The method for generating a voice assistant image as described in claim 1, characterized in that... The system collects user voice commands, then quantifies and analyzes the user's intent based on a natural language processing model, and outputs structured generated commands, including: Collect user voice commands and convert them into text commands; Based on the natural language processing model, the text instructions are semantically parsed, relevant nouns are extracted from them according to the user's intent as a set of structured keywords, and style tendency vectors are determined based on relevant adjectives; The output contains a structured set of keywords and a style preference vector, which are structured generation instructions.

3. The method for generating a voice assistant image as described in claim 1, characterized in that, Based on the structured generation instructions, the comprehensive matching degree between candidate shapes and user intent is traversed to determine the basic shape, including: Traverse the preset candidate shapes in the structured component library to obtain the morphological constraints of the candidate shapes. The morphological constraints of the candidate shapes include a set of semantic tags, a preset style vector, and the number of polygon faces. Based on the morphological constraints of the candidate shapes and the structured generation instructions, a comprehensive analysis is performed to calculate the overall matching degree between the candidate shapes and the user's intent. A relative optimization strategy is adopted to select the candidate shape with the highest overall matching degree. It is then determined whether the overall matching degree meets the minimum threshold requirement. If it does, it is determined to be the basic shape; otherwise, the default shape is selected as the basic shape.

4. The method for generating a voice assistant image as described in claim 3, characterized in that, Based on the structured generation instructions, the comprehensive matching degree between feature components and user intent is traversed, and then the final set of feature components is determined by combining the filtering within mutually exclusive groups, including: Traverse the preset feature components in the structured component library to obtain the morphological constraints of the feature components. The morphological constraints of the feature components include a semantic tag set, a preset style vector, and the number of polygon faces. Based on the morphological constraints of the feature components and the structured generation instructions, a comprehensive analysis is performed to calculate the overall matching degree between the feature components and the user intent. Based on the dynamic matching threshold, each comprehensive matching degree is traversed, and the feature components that meet the user intent benchmark are selected and saved to a temporary candidate set. Determine whether the number of feature components in the temporary candidate set is greater than a preset threshold. If so, directly output the temporary candidate set as the matching candidate set. Otherwise, perform adaptive threshold adjustment and return to the temporary candidate set for filtering. Perform mount point compatibility verification to filter out incompatible feature components from the mount points in the matching candidate set; Based on the mutual exclusion group screening, the feature components that conflict in the matching candidate set are selected and saved; then the occlusion order is determined according to the rendering order, resulting in the final feature component set.

5. The method for generating a voice assistant image as described in claim 4, characterized in that, The adaptive threshold adjustment specifically involves: dynamically reducing the dynamic matching threshold according to a preset step size, determining whether the reduced dynamic matching threshold is greater than the lower threshold, and if so, outputting a new dynamic matching threshold, until the number of components is greater than the preset threshold, or until the dynamic matching threshold is equal to the lower threshold.

6. The method for generating a voice assistant image as described in claim 4, characterized in that, Based on the screening within mutually exclusive groups, the feature components that conflict in the matching candidate set are preferentially saved, specifically: The feature components mounted on the same node in the matching candidate set are divided into several mutually exclusive groups; Iterate through each mutually exclusive group and retain the feature component with the highest overall matching degree; When the difference in the overall matching degree of the feature components within the mutually exclusive group is less than the matching threshold, the style similarity between the feature component and the style tendency vector is calculated, and the feature component with the highest style similarity is retained; wherein, when the overall matching degree of the feature components within the mutually exclusive group is the same, the feature component with higher priority is retained.

7. The method for generating a voice assistant image as described in claim 4, characterized in that: It also includes a pre-built structured component library, which includes a basic shape library and a feature component library. The basic shape library includes at least two different candidate shapes, and the feature component library includes several local feature components attached to the candidate shapes.

8. The method for generating a voice assistant image as described in claim 1, characterized in that, Based on the aforementioned basic shape and feature component set, parametric assembly is performed, and stylization processing is carried out in conjunction with the aforementioned style tendency vector to output the target image model, including: Using the basic shape as the supporting skeleton, the component pose of each feature component in the feature component set in the local coordinate system on the basic shape is calculated, and the assembly model is obtained by parametric assembly. Using the style tendency vector as the quantization input for style control, the corresponding set of stylization rules is mapped out, and then the assembled model is stylized according to the set of stylization rules to output the target image model.

9. The method for generating a voice assistant image as described in claim 1, characterized in that, Associate action resources based on security level classification and output interactive virtual avatar objects, including: Pre-set motion resources corresponding to different safety level categories to build a dedicated vehicle animation library; Construct a context-resource mapping rule table to define the triggering logic for vehicle status and interaction actions; The current vehicle status is obtained, and then the corresponding motion resources are obtained from the vehicle-specific animation library based on the context-resource mapping rule table. These resources are then associated with the target image model to output an interactive virtual image object.

10. The method for generating a voice assistant image as described in claim 1, characterized in that, Perform lightweight real-time rendering and driving, including: Real-time rendering: The virtual avatar object is input, and the vertex positions are calculated using the Unity URP pipeline and the linear blending skin algorithm, and then rendered using GPU acceleration. Context-driven: The script queries the vehicle status every frame, matches the context-resource mapping rule table to trigger interactive actions, drives animation fusion through the Animator component, and outputs interactive screens.

Images