Self-adaptive voice dialogue generation system of intelligent toy and use method of self-adaptive voice dialogue generation system
By using an adaptive voice dialogue generation system for smart toys, and leveraging multi-module combinations and advanced models, the system addresses the rigidity and security issues of smart toys in open-domain dialogue. It achieves personalized and emotional adaptive interaction, adapts to children's growth, and improves the fluency and security of the interaction.
Patent Information
- Application Number
- CN202511674761.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-24
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing smart toy voice systems are unable to conduct open-domain multi-turn dialogues, lack personalized and emotional interaction, and cannot adapt to the cognitive development and interests of children, posing safety risks.
It employs a voice acquisition module, a voice recognition module, a user profiling module, a dialogue management module, a natural language generation module, and a security filtering module, combined with a microphone array, a CNN-LSTM acoustic model, a large language model, and a BERT text classification model, to achieve dynamic user profiling and adaptive dialogue generation.
It achieves highly robust, personalized, and emotional adaptive interaction, improves the fluency and security of dialogue, adapts to children's growth, and enhances educational value and safety.
Smart Images

Figure CN121565167A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent toy voice dialogue technology, and in particular to an adaptive voice dialogue generation system for intelligent toys and its usage method. Background Technology
[0002] Existing intelligent toy voice systems, such as retrieval models based on pre-set question-answer pairs or task-oriented dialogues based on simple flow trees, suffer from rigid dialogue content and are unable to conduct open-domain multi-turn dialogues.
[0003] Although some studies have attempted to introduce affective computing, such as the invention patent with publication number CN101789100A that discloses an intelligent toy based on artificial psychology, its personality growth system is mostly based on the dynamic evolution of machine learning algorithms on interactive data, which has limitations in the naturalness and depth of dialogue and the understanding of complex contexts.
[0004] Furthermore, existing solutions generally lack a refined perception and long-term memory of the personalized characteristics of children, making them unable to adapt to children of different ages, cognitive levels, and interests, let alone evolve with children's growth. Market evaluations indicate that some mainstream smart toys (such as the children's version of Amazon Alexa and Echo smart speakers) have low voice recognition accuracy, poor interaction fluency, and the risk of generating inappropriate content due to a lack of effective security filtering.
[0005] Existing technologies have not yet disclosed a technical solution for achieving highly adaptive dialogue that grows alongside children by deeply fusing dynamic quantified user profiles with large language models. Therefore, there is an urgent need in this field for an intelligent toy system that can truly understand children, conduct personalized and emotional adaptive dialogue, and ensure safety and reliability. Summary of the Invention
[0006] This invention provides an adaptive voice dialogue generation system for smart toys and its usage method, which solves the problems of environmental interference and children's voice recognition difficulties, rigid interactive content and lack of personalization, the contradiction between security and controllability of open domain dialogue, and the lack of emotional interaction in existing smart toy voice interaction systems.
[0007] The present invention provides the following solution to the above-mentioned technical problems: An adaptive voice dialogue generation system for intelligent toys, comprising a voice acquisition module, a voice recognition module, a user profiling module, a dialogue management module, a natural language generation module, a security filtering module, and a voice synthesis module. The voice acquisition module acquires user voice signals. The voice recognition module, connected to the voice acquisition module, converts the voice signals into text information. The user profiling module establishes and dynamically updates a personalized user profile corresponding to the user. The user profile includes at least a cognitive level quantification value, an interest preference vector, and an emotional pattern label calculated based on historical interaction data. The dialogue management module, connected to the voice recognition module and the user profiling module, fuses the user profile data with the current dialogue text and historical dialogue context to generate adaptive dialogue decisions. The natural language generation module, connected to the dialogue management module, generates response text that conforms to the cognitive level quantification value based on the adaptive dialogue decisions. The security filtering module, connected after the natural language generation module, performs content security checks on the generated response text. The voice synthesis module, connected to the security filtering module, converts the security-checked response text into a voice signal and plays it.
[0008] Based on the above technical solution, the present invention can be further improved as follows.
[0009] Furthermore, the voice acquisition module employs a microphone array and runs a beamforming algorithm, which can directionally enhance children's voices in noisy environments. This significantly improves the clarity and signal-to-noise ratio of children's voices captured in real, noisy home environments, providing high-quality input signals for subsequent speech recognition. This reduces the recognition error rate caused by voice quality issues from the source, thereby ensuring the smooth start and reliability of the entire dialogue process.
[0010] Furthermore, the speech recognition module employs an acoustic model optimized for children's pronunciation characteristics. This acoustic model is a CNN-LSTM deep neural network model trained using a dataset containing children's speech corpora. It addresses the industry pain point of the sharp performance drop of general speech recognition models in children's interactive scenarios, ensuring that user intentions are correctly understood. This is a key technical guarantee for achieving effective human-computer interaction and significantly improves the recognition accuracy and error tolerance for children, especially young children's ambiguous and non-standard pronunciations.
[0011] Furthermore, the user profile module dynamically updates the user profile through an incremental learning algorithm. This algorithm uses online learning technology to adjust the weights of the cognitive level quantification value, interest preference vector, and emotional pattern label based on each interaction data. The cognitive level quantification value is calculated using a linear regression model based on the average sentence length and lexical diversity index in the user's historical dialogues. This allows the toy to act like a companion, continuously adjusting its understanding of the child's cognitive abilities and interests as the child grows. This provides a personalized experience that continuously matches the child's current state, avoiding the problem of traditional systems having fixed interactive content that cannot "grow," and giving the system long-term memory and adaptive evolution capabilities.
[0012] Furthermore, the dialogue management module integrates a large language model based on the Transformer architecture, and the workflow of the dialogue management module includes:
[0013] a. Map the cognitive level quantification value in the user profile to the upper limit of lexical complexity and sentence length limit of the large language model;
[0014] b. Use the interest and preference vectors in the user profile as topic relevance weights to guide the large language model to prioritize associating high-weight topics when making generation decisions;
[0015] c. Based on the sentiment analysis results of the input text and the sentiment pattern labels in the user profile, select a sentiment response strategy from a predefined strategy library;
[0016] On the one hand, the large language model ensures the natural fluency and open-domain capability of the dialogue; on the other hand, the strong constraints of the profile conditions ensure that the generated content is always suitable for current children, personalized, and emotionally intelligent, thus achieving a qualitative leap from "being able to have a dialogue" to "being able to have a dialogue".
[0017] Furthermore, the predefined strategy library includes an "empathy-encouragement-transfer" strategy, where the "empathy" step is used to respond to the user's emotions, the "encouragement" step is used to boost the user's confidence, and the "transfer" step is used to guide the topic to high-weight topics in the user's interest preference vector. This enables the system to no longer simply retrieve content or provide empty responses when faced with children's negative emotions, but to execute a scientific, effective, and complete psychological support process, significantly improving the emotional depth and educational significance of the interaction and enhancing the companion value of toys.
[0018] Furthermore, the security filtering module employs a combination of rule matching and a neural network-based text classification model to perform content security checks on the response text. The rule matching includes filtering based on a preset sensitive keyword library, and the text classification model uses the BERT architecture. This approach balances filtering speed and certainty regarding explicit risks while also possessing the intelligent judgment capability to handle complex and ambiguous semantic scenarios, significantly reducing the probability of false positives and false negatives. This provides end-to-end, robust security protection for children.
[0019] A method for using an adaptive voice dialogue generation system for a smart toy includes the following steps:
[0020] S1, acquires user voice input through the voice acquisition module;
[0021] S2, the user's voice input is converted into input text through the speech recognition module;
[0022] S3, the dialogue management module calls the user profile module to obtain dynamic user profile data;
[0023] S4, the dialogue management module integrates the dynamic user profile data with the input text and historical dialogue context to generate adaptive dialogue decisions;
[0024] S5, the natural language generation module generates a response text that matches the cognitive level quantification value based on the dialogue decision;
[0025] S6, The response text is reviewed for compliance through the security filtering module;
[0026] S7, the speech synthesis module synthesizes the approved reply text into a speech signal and outputs it.
[0027] The beneficial effects of this invention are as follows: This invention provides an adaptive voice dialogue generation system for intelligent toys and its usage method, which has the following advantages:
[0028] 1. It achieves highly robust end-to-end interaction, from beamforming algorithm for directional sound pickup on the hardware side to CNN-LSTM acoustic model optimized for children on the software side, forming a highly fault-tolerant speech processing chain, which fundamentally ensures the smoothness and stability of the interaction process in real children's usage scenarios.
[0029] 2. It achieves deep personalization and accompanying growth, and introduces a dynamic user profile based on incremental learning algorithm. This system can not only remember the user's preferences, but also quantify the user's cognitive growth and interest changes through continuous learning, thereby dynamically adjusting the interaction strategy and content complexity, making the toy an intelligent partner that can "grow together", which greatly enhances the long-term user stickiness and educational value of the product.
[0030] 3. It breaks through the technical bottleneck of openness and security. By deeply integrating user profiles as strong constraints into the large language model generation process based on Transformer, it creatively solves the contradiction between "being open" and "being controllable". The system can generate natural and diverse open-domain dialogues using the large language model, and ensure that the generated content is highly matched with current child users in terms of vocabulary, sentence length, topics and emotions, thus achieving "constrained creativity".
[0031] 4. Enhanced emotional intelligence and interaction depth: By combining a predefined strategy library such as "empathy-encouragement-transfer" with user emotional tags, psychological principles are transformed into executable technical actions. This makes the system no longer a cold question-and-answer machine, but an intelligent agent that can perceive user emotions and provide strategic and warm emotional responses, significantly enhancing the emotional depth of interaction and companionship value.
[0032] 5. A multi-layered security system has been constructed, employing a dual filtering mechanism that combines rule matching with the BERT text classification model. This balances processing speed with semantic understanding depth. This architecture can quickly intercept explicit illegal content and intelligently identify potentially harmful deep semantics, creating a reliable and safe end-to-end environment with multiple layers of protection for children and addressing parents' core concerns about the security of intelligently generated content.
[0033] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail in the following embodiments and their accompanying drawings. Attached Figure Description
[0034] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0035] Figure 1 A system architecture diagram of an adaptive voice dialogue generation system for a smart toy and its usage method provided in an embodiment of the present invention;
[0036] Figure 2 The present invention provides an adaptive voice dialogue generation system for a smart toy and a method for using the system, which is a method flowchart of an embodiment of the present invention. Detailed Implementation
[0037] The following is in conjunction with the appendix Figure 1-2The principles and features of the present invention are described below. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. The advantages and features of the invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the invention.
[0038] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.
[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0040] like Figure 1 As shown, this invention provides an adaptive voice dialogue generation system for smart toys, including a voice acquisition module, a voice recognition module, a user profiling module, a dialogue management module, a natural language generation module, a security filtering module, and a voice synthesis module. The system is characterized by: the voice acquisition module acquiring user voice signals; the voice recognition module, connected to the voice acquisition module, converting the voice signals into text information; the user profiling module establishing and dynamically updating a personalized user profile corresponding to the user, the user profile including at least a cognitive level quantification value calculated based on historical interaction data, an interest preference vector, and an emotional pattern label; the dialogue management module, connected to the voice recognition module and the user profiling module, fusing user profile data with the current dialogue text and historical dialogue context to generate adaptive dialogue decisions; the natural language generation module, connected to the dialogue management module, generating response text conforming to the cognitive level quantification value based on the adaptive dialogue decisions; the security filtering module, connected after the natural language generation module, performing content security checks on the generated response text; and the voice synthesis module, connected to the security filtering module, converting the security-approved response text into a voice signal and playing it.
[0041] The voice acquisition module uses a circular 6-microphone array (e.g., the ReSpeaker series), with each microphone spaced 2-5 cm apart to cover a 360-degree sound field. The microphone array is connected to the processor's audio interface (e.g., I2S bus) and runs a real-time beamforming algorithm. The specific steps are as follows:
[0042] Step 1: Acquire multiple audio signals through a microphone array. Preprocessing includes pre-emphasis, framing (frame length 25ms, frame shift 10ms), and windowing (Hamming window).
[0043] Step 2: Calculate the Time Difference of Arrival (TDOA) and estimate the source direction using the Generalized Cross-Correlation Function (GCC-PHAT). The formula is as follows:
[0044]
[0045] Where Xi(f) and Xj(f) are the frequency domain signals of microphones i and j, and τ is the time difference;
[0046] Step 3: Based on the sound source direction, generate an enhanced speech signal using the Minimum Variance Distortionless Response (MVDR) beamforming algorithm. The beamforming weight vector w is obtained by solving the following optimization problem:
[0047]
[0048] Where Φnn is the noise covariance matrix and a(θ) is the steering vector of the sound source direction;
[0049] Results: In noisy environments with a signal-to-noise ratio below 5dB (such as a family living room), this algorithm can improve the clarity of children's speech by approximately 30%.
[0050] The speech recognition module utilizes the main processor (such as a quad-core ARM Cortex-A72) to run the speech recognition model, supports real-time inference, and adopts a CNN-LSTM hybrid acoustic model. The specific structure is as follows:
[0051] Input: Audio features are Mel spectrograms (80-dimensional Mel filter bank, 25ms per frame, 10ms step).
[0052] The CNN part consists of three convolutional layers, each using a 3x3 convolutional kernel with 32, 64, and 128 channels respectively. Each convolutional layer is followed by ReLU activation and 2x2 max pooling.
[0053] LSTM section: 2-layer bidirectional LSTM, 128 hidden units per layer, output connected to a fully connected layer;
[0054] Output: Character-level probability distribution (e.g., 29 classes: 26 letters + space + punctuation);
[0055] Training data: A children's speech corpus (such as CMUKidsCorpus) containing 500 hours of speech from children aged 3-12 years was used. Data augmentation included adding background noise and velocity perturbations.
[0056] Decoding: Bundle search decoding is performed using a language model (a 3-gram model trained on children's text corpus);
[0057] Results: On the children's speech test set, the word error rate (WER) was reduced from 25% for the general model to 12%;
[0058] The user profiling module runs on the main processor, and user data is stored in a local encrypted database (such as SQLite). It dynamically updates user profile parameters using an online gradient descent algorithm. The specific steps are as follows:
[0059] Calculation of cognitive level quantification:
[0060] Based on historical dialogue data, features were extracted, including average sentence length (L) and lexical diversity index (D calculated using Type-TokenRatio).
[0061] Using a linear regression model:
[0062]
[0063] The coefficient αi is updated through online learning, and the update rule is as follows:
[0064]
[0065] Where η is the learning rate (set to 0.01), Ctrue(t) is supervised by expert-annotated cognitive level labels (such as based on age criteria), and Xi(t) is the feature value;
[0066] Interest preference vector update: The vector dimension is 100, corresponding to predefined topics (such as animals, science, music), and is initialized as a zero vector;
[0067] After each interaction, use TF-IDF to extract keywords from the input text and update the vector:
[0068]
[0069] Where f(t) is the topic frequency vector of this interaction, and β=0.1 is the forgetting factor;
[0070] Sentiment pattern label update: Analyze the sentiment score (positive, negative, neutral) of each conversation using a pre-trained sentiment classifier (based on BERT fine-tuning);
[0071] Based on the sentiment score distribution of the sliding window (the most recent 50 interactions), assign labels (such as "optimistic" and "sensitive").
[0072] Effect: User profiles are updated every 24 hours, enabling the system to adapt to children's cognitive growth (such as vocabulary expansion).
[0073] The dialogue management module uses an integrated GPU accelerator (such as NVIDIA Jetson Nano) to run large language models, and a pre-trained model with the Transformer architecture (such as a mini version of GPT-2 with 150 million parameters), and is fine-tuned for children's dialogue.
[0074] Workflow: a. Vocabulary and sentence length control: The cognitive level quantification value C is mapped to the upper limit of vocabulary complexity (e.g., when C<5, only the basic 1000 words are used) and the sentence length limit (e.g., when C<5, the sentence length is ≤10 words). This is achieved by limiting the sampling space during generation.
[0075] b. Topic Guidance: The interest preference vector v is used as the bias weight of the attention mechanism. Specifically, in the cross-attention layer of the Transformer, the attention score is modified as follows:
[0076]
[0077] Where λ=0.5 is the weight coefficient, and Q, K, and V are the query, key, and value matrices, respectively.
[0078] c. Choice of Emotional Strategies:
[0079] The sentiment analysis model (based on fine-tuning of RoBERTa) is used to analyze the sentiment of the input text and output the sentiment label Einput. Combined with the sentiment pattern label Euser in the user profile, a strategy is selected from the strategy library. For example, if Einput is "negative" and Euser is "sensitive", then the "empathy-encouragement-displacement" strategy is selected.
[0080] A predefined strategy library, with an example of the "empathy-encouragement-transfer" strategy:
[0081] Empathy: Generate responses such as "I know you're feeling sad right now, that's okay."
[0082] Encouragement: Generate responses such as, "You were very clever in solving that problem before!"
[0083] Shift the focus: Guide the conversation to high-priority topics, such as, "Shall we talk about your favorite dinosaurs?"
[0084] Results: In open-domain dialogue testing, the user satisfaction rate for personalized responses reached 85%.
[0085] The natural language generation module and the dialogue management module share GPU resources. Based on the dialogue decisions (such as policies and constraints) output by the dialogue management module, the same Transformer model is used to generate text. By adjusting the generation parameters (such as top-k sampling, k=50), the text is ensured to be consistent with the cognitive level.
[0086] The security filtering module runs on the main processor and supports real-time inference.
[0087] A pre-set sensitive keyword library (such as 1000 words like violence and inappropriate advertising) is used for fast matching, and if a match is found, the reply is directly rejected.
[0088] Neural network model: The BERT mini-model (4 layers, 512 hidden dimensions) was used, and the training data included 100,000 child safety annotation texts (positive and negative examples).
[0089] Input: The reply text, converted into WordPiece word vectors;
[0090] Output: Binary classification (safe / unsafe), threshold 0.8;
[0091] Training: Using cross-entropy loss and the Adam optimizer;
[0092] Results: False positive rate less than 5%, processing latency <100ms;
[0093] The speech synthesis module integrates a hardware audio encoder (such as ES8316Codec) and uses WaveNet or Tacotron2 models (optimized for children's speech) to convert text into speech. Parameters: sampling rate 16kHz, bit rate 128kbps.
[0094] Detailed usage steps Figure 2 The flowchart and steps for using the method are as follows:
[0095] S1: Voice capture. The user (child) speaks to the toy, the microphone array captures the raw audio, and the beamforming algorithm enhances the child's voice in real time while suppressing background noise (such as television sound).
[0096] S2: Speech recognition. The enhanced audio is fed into the CNN-LSTM acoustic model, which outputs a text sequence. For example, if a child inputs "I am not happy today", it will be recognized as the text "I am not happy today".
[0097] S3: Call the user profile. The dialogue management module loads user profile data from the database based on the user ID (via voiceprint recognition or manual setting). For example, it obtains the user's cognitive level quantitative value C=4.2 (medium level), and the interest preference vector shows that the high-weight topic is "dinosaur" and the emotion pattern label is "sensitive".
[0098] S4: Generate dialogue decisions by integrating input text, historical context (the last 10 rounds of dialogue), and user profile. Sentiment analysis identifies "unhappy" as a negative emotion. Based on the sentiment strategy library, select the "empathy-encouragement-transfer" strategy. Based on interest vectors, guide the topic to "dinosaurs" and output the dialogue decision: empathy response + encouragement statement + transfer to dinosaur topic.
[0099] S5: Generate response text. The natural language generation module generates text based on the decision, while being constrained by cognitive level (sentence length ≤ 12 words, simple vocabulary). For example, the generated response might be: "It sounds like you're in a bad mood today, but you're a very brave child! Would you like to hear a dinosaur story?"
[0100] S6: Security filtering, rule matching checks for sensitive words, BERT model classifies the response text, outputs a security probability of 0.95, passes review;
[0101] S7: Speech Synthesis and Output. The speech synthesis module converts text into speech, which is then played through a speaker using a child-friendly tone (such as high-frequency tuning) and automatically adjusts the volume.
[0102] like Figure 2 As shown, the specific working principle and usage method of this invention are as follows:
[0103] Example 1: System Hardware Configuration and Initialization;
[0104] The adaptive voice dialogue generation system for smart toys of the present invention can be integrated into a physical smart toy, such as a toy dog or a smart story machine, and its hardware core is shown below.
[0105] Main control unit: Employs an embedded processor with a certain computing power, such as ARM Cortex-A53, responsible for the scheduling, calculation and logic control of the entire system;
[0106] Voice acquisition module: It adopts a microphone array consisting of 4 MEMS microphones, which is installed in a suitable position on the toy's head to collect user voice from all directions;
[0107] Audio playback module: Equipped with a 2W speaker for playing synthesized speech signals;
[0108] Storage module: Equipped with 4GB RAM and 16GB Flash storage, used to store system programs, user profile data and cached interaction records;
[0109] Communication module: Integrated Wi-Fi module for connecting to cloud services;
[0110] Power management module: Built-in 3000mAh lithium battery and charge / discharge management circuit to power the entire system;
[0111] After the system is powered on, it completes a hardware self-test and connects to the cloud server via Wi-Fi. On the first use, the system initializes the user profile through guided dialogue. The initial cognitive level is set according to the user's input age, the interest preference vector is initialized to a zero vector, and the emotional mode label is set to the default "neutral" state.
[0112] Example 2: Dynamic Update Process of User Profile
[0113] The user profiling module dynamically updates the user profile after each interaction using an incremental learning algorithm. The specific implementation process is as follows:
[0114] Assuming the current user profile has an interest weight of 0.6 for "dinosaurs," and in a new interaction, the user mentions "Tyrannosaurus Rex" and "fossils" multiple times, the incremental weight ΔD for "dinosaur"-related topics in this interaction is calculated to be 0.3 using the TF-IDF algorithm. The system's forgetting factor α is set to 0.9. Then, according to the weight update formula W_new=α×W_old+(1-α)×ΔD, the new weight for "dinosaurs" is calculated to be: W_new=0.9×0.6+(1-0.9)×0.3=0.57;
[0115] Meanwhile, the system analysis showed that the average sentence length of the user's statements was 8 words and the vocabulary diversity index was 0.7. These two indicators were input into the pre-trained linear regression model, and the cognitive level quantification value was calculated to be 0.1 higher than the previous interaction. The sentiment analysis model judged the user's emotion to be "excited", so the weight of the "cheerful" sentiment label was increased by 0.05 accordingly.
[0116] Example 3: A Complete Adaptive Dialogue Flow Example
[0117] To illustrate the system's workflow, let's take a specific example: User "Lele" (5 years old) says, "My drawings were laughed at by other children..."
[0118] S1 Voice Acquisition: The voice acquisition module captures Lele's crying voice through a microphone array. The beamforming algorithm accurately locates the direction of the sound source, improving the signal-to-noise ratio by about 15dB and effectively suppressing background TV noise.
[0119] S2 Speech Recognition: The speech recognition module uses a CNN-LSTM acoustic model optimized for children to accurately convert speech into text. This model achieves a recognition accuracy of 94.5% on a children's speech test set, significantly higher than the 82.3% of the general model.
[0120] S3 obtains user profile: The dialogue management module calls Lele's user profile, the data is: {cognitive level: Lv2, interest preference: [dinosaur: 0.9, princess: 0.2], emotional pattern: [sensitivity: 0.7]};
[0121] S4 generates dialogue decisions: The dialogue management module performs the following operations;
[0122] Sentiment analysis identifies the input text as expressing the emotion of "sadness";
[0123] Cognitive level Lv2 is mapped to the restrictions of "using basic vocabulary" and "sentence length ≤ 15 words";
[0124] Combining real-time emotions and profile-based emotion patterns, the "empathy-encouragement-transfer" strategy is selected from a predefined strategy library;
[0125] Guide the large language model to generate dialogue decisions: "First express empathy → encourage through dinosaur analogy → guide to dinosaur drawing activity";
[0126] S5 Natural Language Generation: Generates response text consistent with Lv2 cognitive level: "Don't be sad, Lele. Did you know? Even the mighty Tyrannosaurus Rex fell down when it first learned to walk, but we didn't give up, and in the end, it became the most powerful dinosaur! How about we draw your Tyrannosaurus Rex?"
[0127] S6 Security Filtering: The security filtering module performs a double review. Rule matching scans did not find any sensitive words, and the BERT model determined that the content was safe with a confidence level of 0.95, which is greater than the threshold of 0.9. The review was passed.
[0128] S7 Speech Synthesis and Playback: The speech synthesis module synthesizes text into speech signals in a gentle and encouraging tone. This interaction data is fed back to the user profile module for dynamic updates in Implementation Example 2.
[0129] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Content not described in detail in this specification is prior art known to those skilled in the art.
[0130] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.
Claims
1. An adaptive voice dialogue generation system for an intelligent toy, comprising a voice acquisition module, a voice recognition module, a user profiling module, a dialogue management module, a natural language generation module, a security filtering module, and a voice synthesis module, characterized in that: The voice acquisition module acquires the user's voice signal, and the voice recognition module is connected to the voice acquisition module to convert the voice signal into text information. The user profile module establishes and dynamically updates a personalized user profile corresponding to the user. The user profile includes at least a cognitive level quantification value, interest preference vector, and emotional pattern label calculated based on interaction history data. The dialogue management module is connected to the speech recognition module and the user profiling module, and integrates user profiling data with the current dialogue text and historical dialogue context to generate adaptive dialogue decisions. The natural language generation module is connected to the dialogue management module and generates response text that conforms to the cognitive level quantification value based on the adaptive dialogue decisions. The security filtering module is connected after the natural language generation module and performs content security checks on the generated response text. The speech synthesis module is connected to the security filtering module and converts the security-approved response text into a speech signal and plays it.
2. The adaptive voice dialogue generation system for an intelligent toy according to claim 1, characterized in that, The voice acquisition module uses a microphone array and runs a beamforming algorithm, which can directionally enhance children's voices in noisy environments.
3. The adaptive voice dialogue generation system for an intelligent toy according to claim 1, characterized in that, The speech recognition module employs an acoustic model optimized for children's pronunciation characteristics. This acoustic model is a CNN-LSTM deep neural network model trained using a dataset containing children's speech corpus.
4. The adaptive voice dialogue generation system for an intelligent toy according to claim 1, characterized in that, The user profile module dynamically updates the user profile through an incremental learning algorithm, wherein the incremental learning algorithm uses online learning technology to adjust the weights of the cognitive level quantification value, interest preference vector, and emotional pattern label based on each interaction data. The cognitive level quantification value is calculated using a linear regression model based on the average sentence length and lexical diversity index in the user's historical dialogues.
5. The adaptive voice dialogue generation system for an intelligent toy according to claim 1, characterized in that, The dialogue management module integrates a large language model based on the Transformer architecture. The workflow of the dialogue management module includes: a. Map the cognitive level quantification value in the user profile to the upper limit of lexical complexity and sentence length limit of the large language model; b. Use the interest and preference vectors in the user profile as topic relevance weights to guide the large language model to prioritize associating high-weight topics when making generation decisions; c. Based on the sentiment analysis results of the input text and the sentiment pattern labels in the user profile, select a sentiment response strategy from the predefined strategy library.
6. The adaptive voice dialogue generation system for an intelligent toy according to claim 5, characterized in that, The predefined strategy library includes the "empathy-encouragement-transfer" strategy, where the "empathy" step is used to respond to the user's emotions, the "encouragement" step is used to boost the user's confidence, and the "transfer" step is used to guide the topic to a high-weight topic in the user's interest preference vector.
7. The adaptive voice dialogue generation system for an intelligent toy according to claim 1, characterized in that, The security filtering module uses a combination of rule matching and a neural network-based text classification model to perform content security checks on the response text; wherein, the rule matching includes filtering based on a preset sensitive keyword library, and the text classification model uses the BERT architecture.
8. A method of using an adaptive voice dialogue generation system based on any one of the intelligent toys described in claims 1-7, characterized in that, Includes the following steps: S1, acquires user voice input through the voice acquisition module; S2, the user's voice input is converted into input text through the speech recognition module; S3, the dialogue management module calls the user profile module to obtain dynamic user profile data; S4, the dialogue management module integrates the dynamic user profile data with the input text and historical dialogue context to generate adaptive dialogue decisions; S5, the natural language generation module generates a response text that matches the cognitive level quantification value based on the dialogue decision; S6, The response text is reviewed for compliance through the security filtering module; S7, the speech synthesis module synthesizes the approved reply text into a speech signal and outputs it.
Citation Information
Patent Citations
Processing method and system of accounting data
CN101789100A
