Whole-house intelligent voice control method based on semantic compression and dynamic priority

Abstract

The application belongs to the technical field of speech recognition and processing, and relates to a whole-house intelligent speech control method based on semantic compression and dynamic priority. A large language model is used to analyze user speech instructions to obtain a control intention, a target device and an adjustment parameter. The control intention and the adjustment parameter are analyzed to determine an instruction priority. The control intention and the adjustment parameter are encoded to generate a compressed code. When the whole-house network load exceeds a set threshold, the instruction priority is mapped to a channel access parameter of a medium access control layer, and the channel competition priority of a compressed data packet is dynamically adjusted. The transmission timing is determined according to the adjusted priority, and the data packet is sent to the target device based on a 2.4G wireless protocol according to the timing. If the response instruction data packet is lost, the instruction priority is improved, and the channel access parameter is remapped. Through the cooperation of semantic compression and dynamic priority, the transmission is lightweight and real-time while the semantic understanding depth is maintained.

Description

Technical Field

[0001] This invention relates to the field of speech recognition and processing technology, and in particular to a whole-house intelligent voice control method based on semantic compression and dynamic prioritization. Background Technology

[0002] With the rapid development of smart home technology, whole-house intelligent voice control has become a core interaction method for enhancing the living experience. Users expect to have seamless conversations with home devices through natural language, which requires the system to not only recognize the literal meaning of commands but also to deeply understand the true intent behind complex statements. However, most current mainstream voice control solutions rely on preset fixed command words or simple semantic templates. When faced with diverse user expressions and complex commands containing implicit conditions, they often appear rigid and clumsy, failing to accurately capture the core objectives and specific parameters of the commands. This leads to control deviations or requires repeated corrections by the user, severely affecting the natural fluency of the interaction.

[0003] The underlying reason for this limitation lies in the significant technical gap between adapting the results of deep semantic understanding to the actual home network environment. Specifically, while large language models can meticulously parse user commands and output structured content containing rich semantic information, the bandwidth of the widely used 2.4G wireless networks in home environments is limited and susceptible to interference. Directly transmitting the model's massive output of understanding results would consume excessive channel resources, causing transmission delays or even packet loss, resulting in the loss or distortion of accurate semantics before it even reaches the device.

[0004] To address the aforementioned issues, an attempt was made to deploy a lightweight semantic understanding model on the terminal device side to complete instruction parsing locally in order to reduce data transmission volume. However, due to limitations in the computing and storage resources of smart home devices, the understanding accuracy of the lightweight model is far inferior to that of the large language model in the cloud, making it difficult to handle complex instructions. Summary of the Invention

[0005] Therefore, the purpose of this invention is to overcome the problems that deep semantic understanding results are difficult to transmit efficiently in bandwidth-limited 2.4G home networks due to the large amount of data, and that key instructions are easily blocked by non-key data, leading to control delays. This invention provides a whole-house smart voice control method based on semantic compression and dynamic priority. By coordinating semantic compression and dynamic priority, while maintaining the depth of semantic instruction understanding, it achieves lightweight and real-time wireless transmission, thereby realizing efficient and reliable whole-house smart voice control in a 2.4G network environment.

[0006] To address the aforementioned technical problems, this invention provides a whole-house intelligent voice control method based on semantic compression and dynamic prioritization, comprising:

[0007] The system parses user voice commands using a large language model to obtain control intent, target device, and adjustment parameters.

[0008] By analyzing the control intent and adjustment parameters, the urgency characteristics of the voice command are obtained to determine the command priority;

[0009] The control intent and adjustment parameters are encoded to obtain a compressed code; the compressed code and the target device identifier are encapsulated to obtain a compressed data packet; when the whole-house network load exceeds a threshold, the instruction priority is mapped to the channel access parameters of the media access control layer to adjust the channel contention priority of the compressed data packet.

[0010] The transmission timing is determined according to the adjusted channel contention priority, and the compressed data packet is sent to the target device according to the transmission timing based on the 2.4G wireless protocol;

[0011] Upon receiving the acknowledgment response from the target device, if the acknowledgment response indicates that a data packet has been lost, the instruction priority is increased, and the channel access parameters are remapped according to the increased instruction priority.

[0012] Preferably, analyzing the control intent and adjustment parameters to obtain the urgency features of the voice command to determine the command priority includes: extracting the command type features from the control intent and the change amplitude features from the adjustment parameters; inputting the command type features and the change amplitude features into a decision tree model to obtain the corresponding command priority; wherein, the decision tree model is constructed using historical voice commands and their corresponding actual transmission delays as training samples through a gradient boosting algorithm, so that the command priority can reflect the real-time requirements of the voice command in the historical network environment.

[0013] Preferably, after sending the compressed data packet to the target device via the 2.4G wireless protocol, the method further includes incremental learning: acquiring transmission logs recorded within a preset period; the transmission logs include the command priority, transmission time, retransmission count, and transmission delay for each transmission; extracting training samples from the transmission logs, each training sample containing input features and a label, the input features being the command priority and network load at the time of transmission, and the label being the actual transmission delay; merging the training samples in the current period with historical training samples to form an incremental training set; using a gradient boosting algorithm to incrementally train the decision tree model on the incremental training set, updating the parameters of the decision tree model; and deploying the updated decision tree model to a priority evaluation module to determine the command priority, enabling the decision tree model to adapt to changes in different home network environments and user habits.

[0014] Preferably, encoding the control intent and adjustment parameters to obtain a compressed code includes: inputting the control intent into a semantic embedding model to obtain a semantic feature vector; performing a hash mapping on the semantic feature vector based on a locality-sensitive hashing algorithm to obtain an intent hash value; performing a hash mapping on the adjustment parameters based on a cyclic redundancy check algorithm to obtain a parameter hash value; and concatenating the intent hash value and the parameter hash value to obtain the compressed code.

[0015] Preferably, the Locality Sensitive Hash algorithm ensures that semantic feature vectors corresponding to semantically similar control intentions fall into the same hash bucket with a high probability, thereby generating the same intention hash value.

[0016] Preferably, mapping the instruction priority to channel access parameters of the medium access control layer includes: mapping the instruction priority to the number of arbitration inter-frame intervals and the minimum contention window number of the medium access control layer; the higher the instruction priority, the smaller the number of arbitration inter-frame intervals and the smaller the minimum contention window number.

[0017] Preferably, determining the transmission timing based on the adjusted channel contention priority includes: waiting for the channel to become idle according to the arbitration inter-frame interval, randomly generating a backoff count value within the minimum contention window number, and determining the channel access time of the compressed data packet according to the backoff count mechanism as the transmission timing.

[0018] Preferably, parsing user voice commands based on a large language model to obtain control intent, target device, and adjustment parameters includes: an encoder based on the large language model converting the user voice command into a semantic vector sequence; the semantic vector sequence being input into an intent classification head, and the control intent being calculated and output using a softmax function; the semantic vector sequence being input into an entity recognition head, and the target device and adjustment parameters being obtained by performing label decoding on each position in the semantic vector sequence through a conditional random field layer.

[0019] Preferably, after parsing the user's voice command based on a large language model to obtain the control intent, target device, and adjustment parameters, the method further includes: performing a validity check on the control intent, target device, and adjustment parameters; if the check indicates that the target device is not in the whole-house device list or the adjustment parameters exceed the device's supported range, then generating an error message; determining the feedback method for the error message based on the command priority, using real-time voice broadcast feedback for high-priority commands and mobile push feedback for low-priority commands.

[0020] Preferably, it further includes: monitoring the channel occupancy rate and collision count of the whole-house network, and determining that the network load exceeds the threshold when the channel occupancy rate exceeds the threshold or the collision count exceeds the threshold.

[0021] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0022] First, the deep parsing results of the large language model are compressed into a concise data packet containing only the core semantic elements. This significantly reduces the amount of data transmitted while ensuring semantic integrity, enabling deep semantic understanding to truly adapt to the bandwidth-constrained 2.4G home network environment.

[0023] Secondly, by directly determining the urgency of instructions at the semantic level and mapping it to MAC layer channel access parameters, the abstract semantic priority is transformed into a specific transmission priority. This enables critical instructions such as security alarms and emergency control to actively preempt the channel when the network is congested, fundamentally avoiding control delays caused by the blocking of non-critical data such as device status reporting.

[0024] Furthermore, the dynamic priority enhancement mechanism during retransmission further enhances the success rate of emergency command transmission under extreme network conditions, ensuring the reliability of whole-house smart control.

[0025] In summary, the whole-house smart voice control method based on semantic compression and dynamic priority described in this invention achieves lightweight and real-time wireless transmission while maintaining the depth of semantic command understanding through the synergy of semantic compression and dynamic priority, thereby realizing efficient and reliable whole-house smart voice control in a 2.4G network environment. Attached Figure Description

[0026] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0027] Figure 1 This is a flowchart of a whole-house smart voice control method based on semantic compression and dynamic priority in a preferred embodiment of the present invention;

[0028] Figure 2 This is a flowchart of the large language model parsing user voice commands in a preferred embodiment of the present invention;

[0029] Figure 3 This is a flowchart illustrating the determination of instruction priority in a preferred embodiment of the present invention;

[0030] Figure 4 This is a flowchart of the incremental learning and updating decision tree model in a preferred embodiment of the present invention;

[0031] Figure 5 This is a flowchart of the encoding process for obtaining compressed codes in a preferred embodiment of the present invention. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0033] This invention relates to a whole-house intelligent voice control system employing a master-slave architecture, comprising a main controller and several peripheral controllers. The main controller is installed on core furniture such as electric sofas, uses an Android operating system, and possesses network connectivity and 2.4G wireless communication capabilities. The main controller is responsible for voice acquisition, local preprocessing, cloud communication, global control logic execution, and user feedback. Peripheral controllers are installed on other furniture and are used to control individual furniture devices, including but not limited to: electric bed control systems, single-chair sofa control systems, electric curtain control systems, and lighting control systems. Each peripheral controller has 2.4G wireless communication capabilities and can interact with the main controller.

[0034] The main controller consists of the following functional modules:

[0035] Voice acquisition module: Acquires user voice commands in real time through a microphone array, and performs preprocessing for sound source localization and noise suppression;

[0036] Local speech processing module: Starts the local speech monitoring service and detects whether the speech is a preset wake word.

[0037] Communication module: It interacts with the peripheral controller via 2.4G wireless protocol, and establishes a connection with the cloud server via Wi-Fi or cellular network.

[0038] Cloud-based interaction module: Used to send voice commands to the cloud server, request intent recognition services, and receive control intents returned from the cloud.

[0039] Control execution module: Based on the control intent returned from the cloud, it parses the target device, control action and adjustment parameters, and sends control commands to the corresponding peripheral controllers through the communication module.

[0040] Voice feedback module: Outputs the control execution results through speech synthesis to provide real-time feedback to the user.

[0041] The system workflow includes local hardware processing and cloud-based intelligent processing, forming a closed loop from voice acquisition to execution feedback.

[0042] The local hardware processing flow includes: starting the local voice monitoring service and collecting audio information in real time through the microphone; detecting whether the voice is a wake word, and if so, converting the voice into text; sending the converted text to the cloud server and requesting intent parsing; receiving the control intent returned from the cloud, finding and executing the corresponding furniture control operation; and feeding back the execution result to the user through voice output.

[0043] The cloud server process includes: receiving text input from local hardware; combining a predefined intent library with a large language model to analyze the text and generate structured control intents; and returning the generated control intents to the local hardware.

[0044] Dynamic Addition of New Devices: This system supports the dynamic addition of new furniture and appliances and online updates to voice capabilities, enabling the system to expand its control range as more home appliances are added. The process for adding new devices includes:

[0045] Triggering and Discovery Phase: After the cloud server system adds a new voice capability, the main control system initiates the device addition process; the main control system sends out a search broadcast to scan for nearby peripheral controller devices; other devices respond to the broadcast, providing their device model and a list of voice control capabilities.

[0046] User confirmation and addition phase: The system prompts the user via voice to discover a new device and asks the user to confirm whether to add it; if the user agrees to add it, the system accepts the request to add the device's voice capabilities; the system adds the new device to the list of nearby devices and obtains the device's complete control capability information.

[0047] Capability assessment and registration phase: The system searches its internal voice control capability list to determine if the new device brings new voice control capabilities; if new control capabilities exist, it registers the new capabilities with the cloud server and updates the local control intent list; after registration, the capability addition is complete, and the device addition process ends.

[0048] Through the above process, newly purchased furniture only needs to be installed with the corresponding peripheral controller to be automatically detected by the main controller and incorporated into the whole-house control system, without requiring users to perform complicated configuration operations.

[0049] The cloud server of the above system processes voice commands based on a large language model to obtain control intentions; however, the following technical problems exist in its implementation:

[0050] In a 2.4G wireless network environment, when multiple furniture devices communicate with the main controller simultaneously, channel resources are limited and susceptible to interference, leading to the following problems: First, the amount of semantic data returned after parsing the large language model is large, and direct transmission will occupy too many channel resources, causing transmission delays; Second, voice commands of different urgency levels cannot be prioritized when the network is congested. For example, emergency control commands and information query commands may be treated equally when the network is congested, resulting in emergency commands not being delivered in a timely manner; Third, data packets are simply retransmitted after loss without considering changes in network congestion, resulting in a low retransmission success rate.

[0051] The purpose of this invention is to overcome the problem that deep semantic understanding results are difficult to transmit efficiently in a bandwidth-limited 2.4G home network due to the large amount of data, and that key commands are easily blocked by non-key data, resulting in control delays. This invention provides a whole-house smart voice control method based on semantic compression and dynamic priority.

[0052] The following combination Figure 1 The technical solutions of the present invention will be described in detail with specific embodiments. It should be noted that the embodiments are only used to explain the present invention and not to limit the present invention.

[0053] Step 1, Voice Command Parsing:

[0054] When a user issues a voice command, the main controller's voice acquisition module captures the audio in real time. The local voice processing module detects the wake word and converts the subsequent speech into text. This text is then sent to the cloud server via the communication module.

[0055] After receiving the text, the cloud server calls a large language model for semantic parsing, and analyzes the user intent, target device, and control parameters based on the large language model. The cloud then returns the parsing results to the main controller in structured data form, obtaining the control intent, target device, and control parameters as core semantic elements.

[0056] Step 2, Determine instruction priority:

[0057] After receiving the core semantic elements, the main controller enters the priority determination stage. It extracts instruction type features from the control intent to identify whether the instruction belongs to the immediate control category, scene switching category, or information query category; it also extracts the change amplitude features from the adjustment parameters to determine whether they are abrupt, gradual, or unchanged. The instruction type features and change amplitude features are then input into a pre-trained decision tree model, which calculates and outputs the corresponding instruction priority based on historical training data.

[0058] Command priority is determined by assigning different transmission weights to different types of commands. Immediate control commands receive high priority, scene switching commands receive medium priority, and information query commands receive low priority. This semantic content-based priority division ensures that emergency commands receive priority processing during network congestion, avoiding the unreasonable situation of treating different types of commands equally and guaranteeing the real-time nature of emergency control commands.

[0059] Step 3, Semantic Compression and Data Encapsulation:

[0060] After obtaining the instruction priority, the control intent and adjustment parameters are encoded and compressed. The control intent is input into a semantic embedding model to obtain a semantic feature vector; this semantic feature vector is then hashed using a locality-sensitive hashing algorithm to obtain an intent hash value; the adjustment parameters are standardized, and a cyclic redundancy check algorithm is used to generate parameter hash values; the intent hash value and the parameter hash value are concatenated to obtain a compressed code. Subsequently, this compressed code is encapsulated with the target device identifier to form a compressed data packet.

[0061] Semantic compression coding compresses variable-length text returned by a large language model into fixed-length binary code, significantly reducing the amount of data transmitted. Under the limited bandwidth of 2.4G networks, this reduction in data volume directly lowers the channel occupancy time for each data packet, creating space for more devices to communicate simultaneously and reducing transmission delays and packet loss caused by excessive data volume.

[0062] Step 4, Network Load Monitoring and Priority Mapping:

[0063] The system monitors channel occupancy and collision count throughout the home network in real time. When either channel occupancy or collision count exceeds a threshold, the network load is deemed to be exceeding a threshold. Instruction priorities are extracted from compressed data packets and mapped to channel access parameters of the Media Access Control (MAC) layer, including the arbitration inter-frame interval (ACI) and the minimum contention window size. Higher priorities correspond to smaller ACIs and smaller minimum contention windows. These two parameters are written to the MAC layer's channel access parameter configuration register to adjust the channel contention priority of compressed data packets.

[0064] When network load exceeds a threshold, priority is mapped to MAC layer channel access parameters, transforming abstract priority into specific physical layer contention parameters. Higher priority commands receive smaller arbitration inter-frame intervals and smaller contention windows, resulting in shorter waiting times and a higher probability of winning in channel contention.

[0065] Step 5, Determine transmission timing and send:

[0066] The transmission timing is determined based on the adjusted channel contention priority. First, the system waits for the channel to become idle based on the number of arbitration inter-frame intervals. Then, a backoff count is randomly generated within the minimum contention window. Each time slot becomes idle, the backoff count is decremented by 1. When the backoff count reaches 0, the current moment is determined as the transmission timing, and the channel is immediately occupied to send compressed data packets to the target device via the 2.4G wireless protocol.

[0067] The backoff counting mechanism, combined with priority mapping, forms a complete channel contention solution. Random backoff avoids conflicts caused by multiple devices transmitting simultaneously, while freezing the backoff count value preserves the waiting time, ensuring that high-priority commands maintain an advantage in multiple rounds of channel contention, thus improving the determinism and fairness of channel access.

[0068] Step 6: Confirm Response and Retransmission Processing:

[0069] After receiving the compressed data packet, the peripheral controller of the target device decompresses it to obtain the control intent and adjustment parameters, executes the corresponding control operations, and returns an acknowledgment response. Upon receiving the acknowledgment response, the main controller indicates that the data packet was successfully delivered, and the current control flow ends.

[0070] If the acknowledgment response indicates that the data packet was lost, the system enters the retransmission process. The retransmission count for the current data packet is obtained, and the command priority is increased. The increased priority is re-encapsulated into the retransmitted data packet, and the system returns to step 4 to remap the channel access parameters and execute the retransmission until an acknowledgment response is received or the retransmission count limit is reached.

[0071] The whole-house smart voice control method based on semantic compression and dynamic priority in this invention achieves lightweight and real-time wireless transmission while maintaining the depth of semantic command understanding through the synergy of semantic compression and dynamic priority, thereby realizing efficient and reliable whole-house smart voice control in a 2.4G network environment.

[0072] In the above embodiments, large language models are used to parse voice commands. During the parsing process, intent recognition and entity recognition are typically processed in a pipeline manner, i.e., intent is first classified, and then entity extraction is performed based on the classification results. This serial processing method has the following drawbacks: First, intent classification errors will propagate to entity recognition, leading to error accumulation; second, entity information cannot be used to assist intent judgment. For example, the presence of the entity "lamp" itself can enhance the confidence of the intent "adjustment device".

[0073] To address the error accumulation caused by the sequential processing of intent recognition and entity recognition during voice command parsing, this preferred embodiment provides a whole-house intelligent voice control method based on semantic compression and dynamic prioritization, employing a voice command parsing and differentiated feedback approach based on multi-task joint training. The following describes the method in conjunction with... Figure 2 The technical solution of the present invention will be described in detail with reference to specific embodiments.

[0074] Step 1: Generating semantic vector sequences:

[0075] The user issues a voice command, which the main controller converts into text and inputs into the large language model. The encoder of the large language model performs word segmentation and embedding processing on the text, converting each word into a corresponding semantic vector, forming a sequence of semantic vectors. This sequence maintains the sequential structure of the input text, with each vector corresponding to a word in the original text.

[0076] The second step involves the intent classification head outputting the control intent, and the entity recognition head outputting the target device and adjustment parameters.

[0077] The semantic vector sequence is simultaneously input into the intent classification head. The intent classification head first performs a pooling operation on the vector sequence to obtain the global semantic vector of the entire sentence; then it inputs the global semantic vector into a fully connected layer, which maps it to a score vector of a preset intent category; finally, the softmax function is used to convert the score vector into a probability distribution, and the category with the highest probability is taken as the control intent output.

[0078] Meanwhile, the semantic vector sequence is input into the entity recognition head. The entity recognition head adopts a structure combining a bidirectional long short-term memory network and a conditional random field. The bidirectional long short-term memory network performs context encoding on the semantic vector at each position and outputs the hidden state corresponding to each position. The conditional random field layer performs sequence decoding on these hidden states, considers the transition probability between labels, predicts the optimal entity label for each position, extracts continuous segments based on the label sequence, and obtains the target device and adjustment parameters.

[0079] The intent classification head is a task-specific module connected to the encoder, used to extract global semantics from the entire semantic vector sequence and perform classification. Its internal structure includes: a pooling layer, which converts the variable-length vector sequence into a fixed-length global vector; a fully connected layer, which maps the global vector to the intent category space; and a softmax function, which converts the scores into a probability distribution and selects the category with the highest probability as the output. The intent classification head shares encoder parameters with the large language model, requiring only fine-tuning of the fully connected layer parameters during training.

[0080] The Conditional Random Field (CRF) layer is used for global optimization of sequence labels. Unlike softmax, which predicts labels independently for each position, CRF considers the transition probabilities between labels. For example, a start label for an entity is more likely to be followed by an internal label of that entity than a start label for another entity type. The CRF layer maintains a label transition matrix and searches for the globally optimal label sequence during decoding using the Viterbi algorithm. This mechanism ensures the validity of the output label sequence, avoids invalid label sequences, and effectively improves the accuracy of entity recognition.

[0081] In this scheme, the intent classification and entity recognition tasks share encoder parameters, and the total loss function is a weighted sum of the intent classification loss and the entity recognition loss. During backpropagation, the gradients of both tasks simultaneously update the encoder parameters, enabling the encoder to learn a general semantic representation that benefits both tasks. This training method offers the following advantages: intent information assists entity recognition, and entity information assists intent judgment, with both mutually reinforcing each other.

[0082] This parallel processing architecture, which simultaneously inputs semantic vector sequences into both the intent classification head and the entity recognition head, breaks the limitations of pipelined processing. Intent classification and entity recognition are performed concurrently, with intent classification results assisting entity recognition and entity recognition results assisting intent judgment, creating a synergistic effect of mutual promotion. This effectively improves intent classification accuracy and entity recognition F1 score while reducing inference latency.

[0083] The third step is a multi-task joint training mechanism:

[0084] The intent classification head and entity recognition head share underlying parameters with the encoder of the large language model, and the three are optimized through multi-task joint training. The total loss function is a weighted sum of the intent classification loss and the entity recognition loss. During training, backpropagation updates the encoder parameters simultaneously. This ensures that the semantic representation learned by the encoder is beneficial to both intent classification and entity recognition tasks. Intent information assists entity recognition, and entity information assists intent judgment, forming a mutually reinforcing synergistic effect.

[0085] Step 4, Legality Verification:

[0086] After obtaining the control intent, target device, and adjustment parameters, the system enters the validity verification phase. It queries the whole-house device list to confirm the existence of the target device. Simultaneously, it verifies whether the adjustment parameters are within the device's supported parameter range. If the verification indicates that the target device is not in the whole-house device list or the adjustment parameters exceed the device's supported range, an error message is generated.

[0087] Fifth, determine the feedback method based on the instruction priority:

[0088] For high-priority commands, real-time voice broadcast feedback is used. The main controller immediately broadcasts the error message through the speech synthesis module, providing immediate feedback to the user. For low-priority commands, mobile push notifications are used to send error messages to the user's mobile app without occupying the voice channel, thus avoiding interference with ongoing conversations.

[0089] The preferred embodiment of this invention employs a multi-task joint training architecture that combines parallel processing of the intent classification head and the entity recognition head with shared encoder parameters. This architecture enables intent recognition and entity recognition to mutually reinforce each other, effectively improving parsing accuracy and reducing inference latency. A conditional random field layer is used to globally optimize the label sequence, ensuring the legality and accuracy of entity recognition. A legality verification mechanism intercepts invalid commands after parsing, avoiding waste of channel resources and device security risks. A priority-driven differentiated feedback method provides immediate voice feedback for high-priority commands and push notifications to the mobile device for low-priority commands, achieving refined management of the interactive experience.

[0090] In whole-house smart voice control systems, different voice commands have significantly different real-time requirements. For example, commands like "Emergency stop the sofa" and "Query current time" require millisecond-level response, while the latter can tolerate second-level latency. Using fixed priorities or simple rules (such as preset priorities based on command type) cannot reflect the real-time requirements of commands in a real network environment. Furthermore, home network environments are dynamically changing; different households have different network load patterns, and the network conditions of the same household vary at different times (e.g., day and night). If priority determination rules cannot be adaptively optimized based on actual transmission latency data, the priorities may deviate from the actual real-time requirements, resulting in delays in emergency commands or wasted channel resources.

[0091] To address the issues of fixed voice command priorities, inability to reflect real-time requirements, and inability to adapt to dynamic network environment changes, a preferred embodiment of this invention provides a whole-house smart voice control method based on semantic compression and dynamic prioritization, which adaptively optimizes command priorities based on feature extraction and incremental learning. The following is combined with... Figure 3 and Figure 4 The technical solution of the present invention will be described in detail with reference to specific embodiments.

[0092] Step 1, Instruction type feature extraction:

[0093] After receiving semantic elements, the main controller extracts instruction type features from the control intent. Instruction type features characterize the macro-category of the instruction, reflecting its inherent real-time requirements. The system pre-defines three instruction types: immediate control, scene switching, and information query. Immediate control includes instructions requiring immediate execution, such as "stop movement" and "emergency shutdown"; scene switching includes instructions requiring multi-device collaboration, such as "switch viewing mode" and "activate sleep mode"; and information query includes instructions with lower real-time requirements, such as "query time" and "get battery level." Through keyword matching or semantic similarity calculation, the current control intent is categorized into the corresponding type and represented by a numerical code, such as 3 for immediate control, 2 for scene switching, and 1 for information query.

[0094] The second step is to extract the variation amplitude feature:

[0095] The variation amplitude feature is used to characterize the drastic change in parameters required by the instruction, reflecting its implicit requirement for timely execution. The system presets three variation amplitudes: abrupt change, gradual change, and no change. An abrupt change indicates a rapid and significant change in the parameter, such as "stop immediately" or "quickly brighten"; a gradual change indicates a smooth change, such as "slightly darken"; and no change indicates that the instruction does not involve parameter change, such as "query status". By detecting whether the adjustment parameter contains keywords indicating an abrupt change, or by the magnitude of the parameter's numerical change, the current adjustment parameter is categorized into the corresponding amplitude and represented by a numerical code, such as 3 for abrupt change, 2 for gradual change, and 1 for no change.

[0096] The third step, feature input and priority output:

[0097] The instruction type feature encoding and the change magnitude feature encoding are combined into a two-dimensional feature vector, which is then input into a pre-trained decision tree model. The decision tree model consists of multiple decision trees, each performing a binary tree judgment on the input features and ultimately outputting a priority value. For example, feature vector [3,3] (immediate control + rapid change) might be calculated by the model to output a priority of 7, while feature vector [1,1] (information query + no change) might output a priority of 1. This priority value is an integer in the range of 0-7. The larger the value, the higher the real-time requirement of the instruction, and the more priority is needed to obtain channel access rights.

[0098] The decision tree model uses historical voice commands and their corresponding actual transmission delays as training samples, and is constructed using a gradient boosting algorithm. Specifically, it collects voice command data over a past period, recording the command type features, variation amplitude features, and actual transmission delay for each command. The command type features and variation amplitude features are used as input features X, and the actual transmission delay is used as the label y to train the gradient boosting decision tree model. The model's learning objective is to output a priority value P for each input feature X, such that commands with the same or similar features in historical data have priorities that reflect the magnitude of the actual transmission delay—commands with larger actual transmission delays should have lower priorities, because high-priority commands should receive lower latency.

[0099] This scheme uses Gradient Boosting Decision Tree (GBDT), which consists of multiple decision trees. Each tree is a CART regression tree, and each CART regression tree has the following components:

[0100] The root node contains the entire training dataset and is the starting point of the tree;

[0101] Internal nodes: Each internal node corresponds to a test condition on a feature. In this scheme, the feature is a two-dimensional vector composed of instruction type feature and change range feature. The question form of each node is "whether the instruction type feature is equal to a certain value" or "whether the change range feature is greater than a certain threshold".

[0102] Branching: Each test result corresponds to a branch, dividing the data into child nodes;

[0103] Leaf nodes: Each leaf node stores an output value (i.e., a priority value), and all samples falling into that leaf node receive the same predicted value.

[0104] After sending compressed data packets to the target device via the 2.4G wireless protocol, detailed information for each transmission is automatically recorded to create a transmission log. Each log entry includes: command priority (the actual priority value used during transmission), transmission time (a timestamp down to milliseconds), retransmission count (the number of retransmissions required for the data packet to eventually succeed), and transmission delay (the total time from the initial transmission to receiving the acknowledgment response).

[0105] Training samples are extracted from the transmission logs at preset intervals (e.g., weekly). Each training sample contains two parts: input features and a label. The input features are the command priority and network load at the time of transmission. Network load refers to the channel occupancy rate at the time of transmission, which can be obtained from the network monitoring module. The label is the actual transmission delay. In this way, "the priority used at the time of transmission + the network load at that time" is associated with "the actual transmission delay achieved".

[0106] The training samples extracted in the current period are merged with historical training samples to form an incremental training set. All historical data is retained during the merging process, allowing the model to leverage long-term accumulated experience while incorporating the latest network state information.

[0107] During incremental training, the original decision tree model structure is preserved, and only the weights and split thresholds of the leaf nodes are adjusted. This training method avoids the high computational cost of complete retraining while maintaining model stability. Knowledge from new data is incorporated into the model by adjusting the weights of the leaf nodes of the existing trees, and new decision trees can be added as needed to capture emerging patterns.

[0108] The updated decision tree model is deployed to the priority evaluation module for determining the priority of subsequent voice commands. Through multiple rounds of incremental learning, the model gradually adapts to the characteristics of the current home network environment and user habits. For example, for the "movie mode" command, which is frequently used during peak evening hours, the model learns that the network load is high during this period and appropriately increases its priority to ensure a smooth experience when switching between scenes.

[0109] This preferred solution quantifies the real-time requirements at the semantic level into a computable feature vector by extracting instruction type features and variation amplitude features; it uses a decision tree model labeled with historical transmission delay to enable priority determination rules to learn from real network experience; it employs an incremental learning mechanism to enable the model to adapt to changes in the network environment characteristics of different households and user habits; and it uses network load as an input feature to enable the model to understand the actual effect of priority under different load conditions.

[0110] In the above embodiment, the compressed code is obtained by encoding, referring to... Figure 5 As shown, it specifically includes the following steps.

[0111] The first step is to generate the semantic feature vector of the control intent:

[0112] Semantic embedding models are pre-trained neural networks capable of converting variable-length text into fixed-dimensional semantic feature vectors. This approach uses a pre-trained Sentence-BERT model, whose architecture includes: an input layer that receives the segmented text sequence, mapping each word to an initial word vector; an encoder layer based on the Transformer architecture, which captures the contextual relationships between words through a self-attention mechanism and outputs a contextual representation for each position; a pooling layer that performs average pooling on all positional representations output by the encoder to obtain a fixed-length sentence vector; and an output layer that, after linear transformation and normalization, outputs the final semantic feature vector. The semantic embedding model is pre-trained on a corpus, ensuring that semantically similar texts are positioned close together in the vector space. For example, the Euclidean distance between "brighten" and "increase brightness" is small, while the distance between "brighten" and "turn off" is large.

[0113] The second step is to generate the intent hash value:

[0114] After obtaining the semantic feature vectors, a Locality Sensitive Hash (LSH) algorithm is used to perform hash mapping on them. LSH divides the high-dimensional vector space into multiple hash buckets using a set of random hyperplanes. Each vector is mapped to a corresponding hash bucket based on its relative position to the hyperplane. Specifically, k random hyperplanes are preset, and each hyperplane corresponds to a hash function h. k =sign(v·w k ), h k Represents the k-th hash function; sign() represents the sign function; w k 'v' represents the hyperplane normal vector; '·' represents the dot product operation. For a semantic feature vector v, calculate its dot product with each hyperplane, and obtain k binary bits according to the sign of the dot product. Combine these k binary bits to obtain the intent hash value.

[0115] Assuming k=16, the intended hash value is a 16-bit binary number, such as "1011001011100101" or "0xB2E5" in hexadecimal. The core characteristic of the Locality Sensitive Hashing (LSH) algorithm is that two semantically similar vectors have a high probability of having the same relative position with respect to the random hyperplane, thus generating the same hash value; semantically dissimilar vectors have an extremely low probability of generating the same hash value.

[0116] Locality-Sensitive Hash (LSH) algorithms are designed with completely different goals than ordinary hash functions. Ordinary hash functions (such as MD5 and SHA) aim for the "avalanche effect," meaning that a small change in the input leads to a large change in the output, in order to avoid collisions. LSH, on the other hand, aims for the "distance-preserving property," meaning that similar inputs have a high probability of producing the same output (falling into the same hash bucket), while dissimilar inputs have a high probability of producing different outputs.

[0117] This scheme employs Random Hyperplane Hashing as a specific implementation of Locality Sensitive Hashing. For two semantic feature vectors u and v, with an angle θ between them, the probability that they are assigned to the same side by the same random hyperplane is 1-θ / π. After passing through k independent random hyperplanes, the probability that they generate the exact same hash value is positively correlated with their cosine similarity. The more semantically similar the vectors, the higher the cosine similarity, and the greater the probability of generating the same hash value.

[0118] For example, the control intents "turn the lights up" and "increase brightness" are highly similar semantically, with a small angle between their semantic feature vectors. After passing through 16 random hyperplanes, the probability of generating the exact same intent hash value "0xB2E5" exceeds 95%. On the other hand, "turn the lights up" and "close the curtains" have significant semantic differences, and the probability of generating the same hash value is close to 0.

[0119] Step 3: Parameter hash value generation:

[0120] After standardizing the adjustment parameters, a cyclic redundancy check algorithm is used to generate parameter hash values. Adjustment parameters may take various forms, such as the numerical parameter "50%", the descriptive parameter "soft mode", or the composite parameter "both brightness and color temperature are adjusted to be soft". First, standardization is performed: for numerical parameters, they are converted to integers with uniform dimensions, such as converting "50%" to 50; for descriptive parameters, they are mapped to preset enumerated values, such as mapping "soft mode" to 1 and "bright mode" to 2; for composite parameters, they are either decomposed and processed separately or merged according to preset rules.

[0121] After standardization, the resulting integer or byte sequence is input into a cyclic redundancy check (CRC) algorithm. This scheme uses the CRC32 algorithm, outputting a 32-bit parameter hash value, such as "0x7A3F9B2C". The CRC32 algorithm is deterministic, efficient, and has good anti-collision properties, ensuring that the same input produces the same output, and that different inputs produce different outputs with a very high probability.

[0122] Step 4, concatenating the compression code:

[0123] The intent hash value generated in the second step and the parameter hash value generated in the third step are concatenated in a preset order, and a parameter count identifier is added to the concatenation result to form the final compressed code. For example, if the intent hash value is "0xB2E5", the parameter hash value is "0x7A3F9B2C", and the parameter count is 1, then the compressed code is "01B2E57A3F9B2C", with a total length of 48 bits. Compared to the tens or even hundreds of bytes that the original text may occupy, the compressed code significantly reduces the amount of data.

[0124] This scheme converts control intent into semantic feature vectors through a semantic embedding model, preserving semantic information in the vector space; it compresses high-dimensional vectors into fixed-length intent hash values ​​using a locality-sensitive hashing algorithm, achieving a data compression rate of over 100 times, while ensuring that semantically similar instructions generate the same hash value with a high probability; it compresses adjustment parameters into parameter hash values ​​using a cyclic redundancy check algorithm, unifying the encoding methods for numerical and descriptive parameters; and it concatenates these hash values ​​to form the final compressed code, significantly reducing the amount of transmitted data while maintaining semantic integrity.

[0125] In a 2.4G wireless network environment, multiple devices share the same channel. When multiple compressed data packets are waiting to be transmitted simultaneously, a mechanism is needed to determine which device gains access to the channel first. While the existing IEEE 802.11 standard's distributed coordination function provides a basic backoff mechanism, it has the following drawbacks: First, all data packets use the same channel access parameters, making it impossible to distinguish between commands of different urgency levels; second, the backoff mechanism only provides randomness and cannot guarantee that higher-priority commands will receive earlier transmission; and third, it lacks explicit rules for mapping semantic priority to physical layer parameters, resulting in priority not being implemented at the MAC layer.

[0126] To address this problem, a preferred embodiment of the present invention provides a differentiated configuration scheme for MAC layer channel access parameters based on priority mapping, specifically including:

[0127] When the whole-house network load exceeds a threshold, the system extracts the instruction priority P from the compressed data packet. The instruction priority P is an integer in the range of 0-7, with a higher value indicating a higher requirement for real-time performance. For example, the priority of an emergency stop instruction is P=7, the priority of a scene switching instruction is P=5, and the priority of an information query instruction is P=1.

[0128] The instruction priority P is mapped to the arbitration inter-frame interval A of the media access control layer. The arbitration inter-frame interval determines the length of time the device needs to wait after detecting that the channel is idle; its calculation formula is as follows:

[0129] A = AIFSN[P] × aSlotTime; AIFSN[P] is the number of arbitration inter-frame intervals; aSlotTime is the slot duration. This scheme uses a non-linear mapping function so that the higher the priority, the smaller AIFSN[P] is; the preset mapping table is as follows:

[0130] .

[0131] Priority 7 and 6 are mapped to AIFSN=2 (minimum), priority 5 to 3, priority 4 to 4, and so on, with priority 0 mapped to 9 (maximum). This non-linear design amplifies the priority differences within high-priority regions, ensuring that the most urgent instructions receive the minimum waiting time.

[0132] Simultaneously, the instruction priority P is mapped to the minimum contention window size CWmin; the minimum contention window size determines the upper limit of the random range of the backoff count value. The preset mapping table for this scheme is as follows:

[0133] .

[0134] Priorities 7 and 6 are mapped to CWmin=3, priorities 5 and 4 to 7, priorities 3 and 2 to 15, and priorities 1 and 0 to 31. Higher priorities result in a smaller contention window and a smaller range for random backoff, thus increasing the probability of obtaining earlier transmission timings.

[0135] The mapped arbitration inter-frame interval number A and minimum contention window number CWmin are written into the channel access parameter configuration register of the medium access control layer for subsequent channel contention processes.

[0136] The device continuously monitors the physical channel and detects its status. When the channel transitions from busy to idle, it enters a waiting phase. Based on the mapped arbitration inter-frame interval number A, the device waits for the corresponding duration. Specifically, the device starts an AIFS timer, decrementing by 1 for each idle time slot. When the timer reaches 0, it indicates that the arbitration inter-frame interval waiting is complete, and the device enters the backoff phase.

[0137] A backoff count value B is randomly generated within the minimum competition window size CWmin. The backoff count value B is a random integer between 0 and CWmin, generated using a uniformly distributed random number generator. For example, if CWmin=7, then B is randomly selected with equal probability within the range of 0-7; if CWmin=3, then B is randomly selected within the range of 0-3.

[0138] The device starts a backoff counter with an initial value of B. The backoff counter decrements by 1 after each idle time slot. During the backoff counting process, the device continuously listens to the channel.

[0139] If the channel remains idle, the backoff counter continues to decrease until it reaches 0;

[0140] If the channel is detected to be occupied by other devices during the backoff process, the backoff count is paused and the current backoff count value is frozen.

[0141] Once the channel transitions from busy to idle again and the arbitration inter-frame interval wait is completed, the backoff count resumes and continues to decrease from the frozen value.

[0142] When the backoff counter decreases to 0, the current moment is determined as the transmission timing of the compressed data packet, and the device immediately occupies the channel to start sending the compressed data packet.

[0143] This scheme maps instruction priorities to the number of arbitration inter-frame intervals and the minimum contention window number in the medium access control layer, enabling high-priority instructions to obtain shorter channel idle waiting times and smaller random backoff ranges. Through nonlinear mapping, the priority differences within the high-priority region are amplified, ensuring that the most urgent instructions obtain the minimum contention parameters when the channel is extremely congested. Through the freezing function in the backoff counting mechanism, high-priority instructions that have been partially backed up retain their existing progress after the channel is preempted, thus maintaining a cumulative advantage in multiple rounds of channel contention.

[0144] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A whole-house intelligent voice control method based on semantic compression and dynamic prioritization, characterized in that... ,include: The system parses user voice commands using a large language model to obtain control intent, target device, and adjustment parameters. By analyzing the control intent and adjustment parameters, the urgency characteristics of the voice command are obtained to determine the command priority; The control intent and adjustment parameters are encoded to obtain a compressed code; the compressed code and the target device identifier are encapsulated to obtain a compressed data packet; when the whole-house network load exceeds a threshold, the instruction priority is mapped to the channel access parameters of the media access control layer to adjust the channel contention priority of the compressed data packet. The transmission timing is determined according to the adjusted channel contention priority, and the compressed data packet is sent to the target device according to the transmission timing based on the 2.4G wireless protocol; Receive the confirmation response from the target device. If the confirmation response indicates that the data packet was lost, increase the priority of the instruction and remap the channel access parameters according to the increased instruction priority. Analyzing the control intent and adjustment parameters to obtain the urgency features of the voice command to determine the command priority includes: extracting the command type features from the control intent and the change amplitude features from the adjustment parameters; inputting the command type features and the change amplitude features into a decision tree model to obtain the corresponding command priority; wherein, the decision tree model is constructed using historical voice commands and their corresponding actual transmission delays as training samples through a gradient boosting algorithm, so that the command priority can reflect the real-time requirements of the voice command in the historical network environment; Encoding the control intent and adjustment parameters to obtain a compressed code includes: inputting the control intent into a semantic embedding model to obtain a semantic feature vector; performing a hash mapping on the semantic feature vector based on a locality-sensitive hashing algorithm to obtain an intent hash value; performing a hash mapping on the adjustment parameters based on a cyclic redundancy check algorithm to obtain a parameter hash value; and concatenating the intent hash value and the parameter hash value to obtain the compressed code.

2. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 1, characterized in that, After sending the compressed data packet to the target device via the 2.4G wireless protocol, incremental learning is also included: Obtain the transmission logs recorded within a preset period; the transmission logs include the instruction priority, transmission time, retransmission count, and transmission delay for each transmission. Training samples are extracted from the sending logs. Each training sample contains input features and a label. The input features are the command priority and network load at the time of sending, and the label is the actual transmission delay. Merge the training samples in the current period with the historical training samples to form an incremental training set; The decision tree model is incrementally trained on the incremental training set using the gradient boosting algorithm, and the parameters of the decision tree model are updated accordingly. The updated decision tree model is deployed to the priority evaluation module to determine the priority of the instructions, so that the decision tree model can adapt to changes in different home network environments and user habits.

3. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 1, characterized in that, The Locality Sensitive Hash algorithm ensures that semantic feature vectors corresponding to semantically similar control intentions fall into the same hash bucket with a high probability, thereby generating the same intention hash value.

4. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 1, characterized in that, Mapping the instruction priority to channel access parameters of the medium access control layer includes: mapping the instruction priority to the number of arbitration inter-frame intervals and the minimum contention window number of the medium access control layer; the higher the instruction priority, the smaller the number of arbitration inter-frame intervals and the smaller the minimum contention window number.

5. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 4, characterized in that, Determining the transmission timing based on the adjusted channel contention priority includes: waiting for the channel to become idle according to the arbitration inter-frame interval, randomly generating a backoff count value within the minimum contention window number, and determining the channel access time of the compressed data packet according to the backoff count mechanism, as the transmission timing.

6. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 1, characterized in that, Based on a large language model, user voice commands are parsed to obtain control intent, target device, and adjustment parameters, including: The encoder based on the large language model converts the user's voice commands into a sequence of semantic vectors; The semantic vector sequence is input into the intent classification head, and the control intent is calculated and output by the softmax function. The semantic vector sequence is input to the entity recognition head, and the label is decoded at each position in the semantic vector sequence through a conditional random field layer to obtain the target device and adjustment parameters.

7. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 6, characterized in that, After parsing user voice commands using a large language model to obtain control intent, target device, and adjustment parameters, the following steps are also included: The legality of the control intent, target device, and adjustment parameters is verified; if the verification indicates that the target device is not in the whole-house device list or the adjustment parameters exceed the device's supported range, an error message is generated. The feedback method for the error message is determined based on the priority of the instruction. For high-priority instructions, real-time voice broadcast feedback is used, while for low-priority instructions, mobile push feedback is used.

8. The whole-house intelligent voice control method based on semantic compression and dynamic prioritization according to claim 1, characterized in that, It also includes: monitoring the channel occupancy rate and collision count of the whole-house network, and determining that the network load exceeds the threshold when the channel occupancy rate or the collision count exceeds the threshold.

Images