Tag generation method, tag generation device and system based on speech recognition

Abstract

The application provides a label generation method, a label generation device and a system based on speech recognition, and belongs to the technical field of label generation. The method can effectively purify the speech signal by collecting the user's voice and matching the noise reduction model corresponding to the target use scene to filter out the scene-specific frequency band noise. The method is compatible with network recognition mode and local recognition mode, can adapt to different network environments, and widens the application range. In the network recognition mode, the scene label keywords are stored in the cloud, the use frequency, the latest use time and the matching priority of the keywords are dynamically updated, the keyword frequency table vocabulary is replaced and updated according to the priority, the scene vocabulary is optimized, and the scene vocabulary characteristics are matched. At the same time, the user's personalized keyword frequency table optimized in the cloud is associated with the target use scene, and is called by the local recognition, the traditional fixed general keyword frequency table is abandoned, the offline speech recognition accuracy can be effectively improved, and the recognition accuracy and generation efficiency of the label generation in multiple scenes are improved.

Description

Technical Field

[0001] This application relates to the field of tag generation technology, and in particular to a tag generation method, tag generation device and system based on speech recognition. Background Technology

[0002] In daily life, when faced with a wide variety of complex things, people often use various labels to identify or classify different items. For example, in the kitchen, it is often necessary to label various packaged meats with information such as their purchase date or expiration date. To meet the need for rapid label generation, there are now some label generation devices on the market based on voice recognition. These devices collect users' voice commands, extract key information, and then send the key information back to the label generation device for printing.

[0003] However, while existing speech recognition-based tag generation methods offer both local and cloud-based recognition modes, the local word frequency tables used in local speech recognition are often general, preset ones. These cannot adapt to the specific word characteristics and noise features of different usage scenarios. As usage time increases, these general offline word frequency tables lack scenario differentiation, contain mixed vocabulary, and have fixed word matching priorities. They cannot be dynamically updated based on the actual word usage in different scenarios, resulting in low accuracy of local speech recognition for commonly used tag words in various scenarios. The word frequency tables cannot dynamically adapt to scenario usage preferences, further reducing the efficiency of tag generation in different scenarios. Summary of the Invention

[0004] To address at least one of the aforementioned issues, this application provides a speech recognition-based tag generation method, tag generation device, and system. The system aims to match a dedicated noise reduction model to filter out scene-specific noise based on the current target usage scenario, and to construct a user-personalized word frequency table bound to the scenario in the cloud. It dynamically updates vocabulary usage data and matching priorities based on tag keywords generated by speech recognition, adaptively replaces words in the word frequency table, and synchronizes them to local storage associated with the scenario. This allows local recognition to access a dedicated word frequency table adapted to the word characteristics of the current scenario, improving speech recognition accuracy in the corresponding scenario and optimizing tag generation efficiency.

[0005] According to a first aspect of the embodiments of this application, a tag generation method based on speech recognition is provided, comprising:

[0006] Collect user voice; Identify the current target use case; The noise reduction model corresponding to the current target usage scenario is invoked to perform noise reduction processing on the user's voice, so as to filter out noise in the preset frequency band under the current target usage scenario; Determine whether the current speech recognition mode is network-based or local. When the network recognition mode is triggered, the noise-reduced user voice is sent to the cloud platform for speech recognition, the tag keywords are extracted, and the tag keyword information is stored in the user's personalized word frequency table on the cloud platform; Update the usage frequency and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user-personalized word frequency table are replaced according to the priority of each keyword, and the user-personalized word frequency table is updated after the word replacement. The system receives tag keyword information and the user-personalized word frequency table from the cloud platform, performs tag printing operation based on the tag keywords, and establishes an association between the received user-personalized word frequency table and the current target usage scenario so that the tag keywords can be identified from the user-personalized word frequency table when the local recognition mode is triggered.

[0007] According to a second aspect of the embodiments of this application, a label generation device is provided, comprising: The voice acquisition module is used to collect user voice data. The scenario confirmation module is used to confirm the current target usage scenario; The noise reduction processing module is used to call the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's voice, so as to filter out noise in the preset frequency band under the current target usage scenario; A wireless communication module is used to establish a communication connection between the tag generation device and the cloud platform; The recognition mode selection module is used to determine whether the current speech recognition mode is network recognition mode or local recognition mode; The network recognition processing module is used to send the noise-reduced user voice to the cloud platform when the network recognition mode is triggered, so that the cloud platform performs the following operations: The user's voice is subjected to speech recognition, and the tag keywords are extracted and stored in the user's personalized word frequency table on the cloud platform; Update the usage frequency and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user-personalized word frequency table are replaced according to the priority of each keyword, and the user-personalized word frequency table is updated after the word replacement. The data receiving and association module is used to receive the tag keyword information and the user personalized word frequency table sent by the cloud platform, and to establish an association between the received user personalized word frequency table and the current target usage scenario, so that the tag keywords can be identified from the user personalized word frequency table when the local recognition mode is triggered in the future. The label printing module is used to perform label printing operations based on the label keywords.

[0008] According to a second aspect of the embodiments of this application, a label generation system is provided, including the label generation device described in the second aspect of the embodiments of this application and a cloud platform, wherein the label generation device is communicatively connected to the cloud platform; The cloud platform is used for: The system receives the noise-reduced user voice sent by the tag generation device, performs speech recognition on the user voice, extracts tag keywords, and stores the tag keyword information in the user's personalized word frequency table on the cloud platform. Update the usage frequency and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user-personalized word frequency table are replaced according to the priority of each keyword, and the user-personalized word frequency table is updated after the word replacement. The tag keyword information and the user-personalized word frequency table are sent to the tag generation device.

[0009] The technical solutions provided by the embodiments of this application have at least the following beneficial effects: The scheme disclosed in this application first collects user voice and confirms the current target usage scenario. Then, it calls the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user voice to filter out noise in the preset frequency band of the current target usage scenario. It can perform targeted filtering for noise specific to different scenarios, effectively eliminating the interference of noise in specific frequency bands of the scenario on the voice signal, improving the purity of the voice signal, and providing a foundation for accurate extraction of tag keywords. This method further determines whether the current speech recognition mode is network recognition mode or local recognition mode, is compatible with both online and offline working modes, and can adapt to various usage environments without network, thus broadening the applicable scenarios of the method. When the network recognition mode is triggered, the noise-reduced user voice is uploaded to the cloud platform to complete speech recognition and extract tag keywords. At the same time, the information of the tag keywords is stored in the user's personalized word frequency table on the cloud platform, relying on the cloud to complete the long-term accumulation of scenario vocabulary. This method updates the usage frequency and recent usage time of each keyword in the user's personalized word frequency table according to the information of the tag keywords, and simultaneously adjusts the matching priority of each keyword in the user's personalized word frequency table. It can quantify the usage popularity of each keyword in the corresponding scenario in real time, overcoming the defects of existing general word frequency tables with fixed matching priorities and unable to adaptively adjust with changes in scenario vocabulary. Based on this, this method replaces words in the user's personalized word frequency table according to the priority of each keyword, and updates the user's personalized word frequency table after word replacement. It performs a process of word selection and elimination within the word frequency table according to the actual usage of words in the scenario, ensuring that the user's personalized word frequency table always matches the word usage characteristics of the current target usage scenario and reducing the accumulation of invalid and redundant words. After receiving the tag keyword information and user's personalized word frequency table from the cloud platform, the device can directly perform tag printing operations based on the tag keywords to quickly complete tag output. It can also establish a connection between the received user's personalized word frequency table and the current target usage scenario. This allows the device to directly call the user's personalized word frequency table, which is bound to the scenario and continuously iterated and optimized in the cloud, to recognize tag keywords when the local recognition mode is triggered. This overcomes the limitation of local offline recognition being limited to fixed general word frequency tables, significantly improving the speech recognition accuracy of the corresponding scenario in offline mode, and overall improving the recognition accuracy and generation efficiency of tag generation in multiple scenarios.

[0010] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit this application. Attached Figure Description

[0011] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the principles of this application.

[0012] Figure 1 This is a first flowchart of a tag generation method based on speech recognition provided in an embodiment of this application.

[0013] Figure 2 This is a second flowchart of a speech recognition-based tag generation method provided in an embodiment of this application.

[0014] Figure 3 This is a flowchart illustrating the steps of user-personalized word frequency table partitioning management provided in one embodiment of this application.

[0015] Figure 4 This is a flowchart illustrating the steps for updating a user-personalized word frequency list according to an embodiment of this application.

[0016] Figure 5 This is a third flowchart of a speech recognition-based tag generation method provided in an embodiment of this application.

[0017] Figure 6 This is a flowchart of the steps performed when storing tag keyword information to a user-personalized word frequency table on a cloud platform, according to an embodiment of this application.

[0018] Figure 7 This is a flowchart of the steps performed before performing a label printing operation based on label keywords, according to an embodiment of this application.

[0019] Figure 8 This is a structural block diagram of a label generation device provided in one embodiment of this application.

[0020] Figure 9 This is a schematic diagram of a label generation system provided in an embodiment of this application.

[0021] Explanation of reference numerals in the attached figures: 800. Label generation device; 810. Voice acquisition module; 820. Scene confirmation module; 830. Noise reduction processing module; 840. Wireless communication module; 850. Recognition mode selection module; 860. Network recognition processing module; 870. Data receiving and association module; 880. Label printing module; 890. Local recognition processing module; 8100. User identity verification module. Detailed Implementation

[0022] To make the objectives, implementation methods, and advantages of this application clearer, exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these exemplary embodiments are provided to make the description of this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. It should be noted that the brief descriptions of terminology in this application are merely for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0023] In the description of this application, it should be understood that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include one or more features.

[0024] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Existing speech recognition-based tag generation solutions generally employ a dual-recognition mechanism, combining cloud and local methods. However, the local recognition stage largely relies on fixed, universally preset word frequency lists, making it difficult to adapt to the specific word usage patterns of different scenarios and to match the unique noise characteristics of each scenario. In long-term use, these universal offline word frequency lists lack a refined scenario differentiation mechanism, resulting in a chaotic and redundant vocabulary. Furthermore, the word matching priority is fixed and cannot be dynamically iterated and optimized based on actual word usage behavior in different scenarios. Consequently, the local recognition accuracy of commonly used tag words is poor across various scenarios, and the word frequency list struggles to adaptively match scenario usage preferences, ultimately leading to poor stability and overall efficiency in tag generation across different application scenarios.

[0027] Based on this, this application provides a method, device, and system for generating tags based on speech recognition. The aim is to match a dedicated noise reduction model to filter out scene-specific noise according to the current target usage scenario, and to build a user-personalized word frequency table bound to the scenario in the cloud. The vocabulary usage data and matching priority are dynamically updated based on the tag keywords generated by speech recognition. After adaptively replacing the words in the word frequency table, the data is synchronized to the local storage associated with the scenario. This allows local recognition to call a dedicated word frequency table adapted to the word characteristics of the current scenario, thereby improving the accuracy of speech recognition in the corresponding scenario and optimizing the tag generation efficiency.

[0028] Reference Figure 1 , Figure 1This is a first flowchart of a tag generation method based on speech recognition provided in an embodiment of this application, including but not limited to steps S110 to S180.

[0029] Step S110: Collect user voice.

[0030] In this step, the tag generation device is normally in standby listening mode. The microphone module on the device continuously collects ambient audio signals and detects preset wake words or trigger commands in real time. When a valid wake signal is captured, the device immediately enters voice acquisition mode, completely recording the entire voice content used by the user to edit tags, forming raw audio data. This raw audio data is temporarily stored in the device's local cache as the basic data source for subsequent scene noise reduction, decoding, word segmentation, and keyword matching, ensuring the integrity and traceability of the voice data chain.

[0031] Step S120: Confirm the current target usage scenario.

[0032] In this step, after completing the voice acquisition in step S110, the device immediately performs the recognition operation of the current target usage scenario.

[0033] This device can be pre-configured for various common usage scenarios, such as kitchen, cable storage, office, organization, and custom scenarios. Each scenario corresponds to an independent scenario library, noise reduction model, and user-personalized word frequency list. This step provides three scenario recognition methods that can be used independently or in combination: The first is manual scenario selection, where the device interface displays all preset scenario options and responds to the user's click to directly determine the target usage scenario for this task; the second is speech feature analysis and recognition, where the device performs semantic and lexical feature analysis on the collected raw speech data, combining it with scenario-specific keyword features to automatically match the corresponding usage scenario; the third is behavioral habit recognition, where the device retrieves the current user's historical label printing records and long-term word preferences, relying on the user's frequently used scenario data to intelligently determine the target usage scenario for this task.

[0034] After scene confirmation, the device will bind a scene identifier to this task. This identifier serves as the prerequisite for subsequent calls to the corresponding scene noise reduction model, loading the scene-specific user-personalized word frequency table, partitioning and managing the user-personalized word frequency table, and executing word frequency table quota allocation, thus achieving scene-based adaptation throughout the entire process.

[0035] For example, taking a real-world multi-scenario application as an example, a certain park is equipped with a kitchen, material storage, and comprehensive office area, and the equipment is pre-stored with corresponding scenario categories and supporting resources.

[0036] Scenario 1: Cable storage operation. The operator directly clicks the "Cable Storage" option on the equipment screen, and the equipment immediately locks the current target scenario as cable storage, subsequently calling the scenario-specific noise reduction model and keyword frequency list.

[0037] Scenario 2: Kitchen ingredient management. Operators directly input information such as "beef shank," "shelf life," and "braised" via voice. The equipment analyzes the typical vocabulary features in the voice and automatically identifies it as a kitchen scenario.

[0038] Scenario 3: Routine Office Scenarios. A staff member frequently uses the device in the office area to create labels for documents and consumables. The device retrieves the user's historical printing data and, combined with their long-standing word usage habits, can automatically determine that the current situation is an office scenario without manual operation or semantic analysis.

[0039] For speech words that cannot be categorized into existing standard scenarios, the device will uniformly classify them into "custom" scenarios. The recognition results of different scenarios are independent of each other, ensuring that subsequent noise reduction, word frequency table retrieval, word frequency statistics, and other processes are all matched to the current environment and usage requirements.

[0040] Step S130: Call the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's voice in order to filter out noise in the preset frequency band under the current target usage scenario.

[0041] In this step, after confirming the target usage scenario in step S120, the device retrieves the noise reduction model specific to that scenario based on the determined scenario identifier. Different usage scenarios have significantly different environmental sound sources and noise frequency bands. In this embodiment, typical noise frequency ranges are pre-defined for various scenarios such as kitchens, cable storage, offices, organization, and custom settings, and independent noise reduction models are trained based on the scenario noise characteristics. The device inputs the raw speech audio data collected in step S110 into the matched noise reduction model, specifically filtering out environmental noise and background interference in the preset frequency bands of that scenario, eliminating invalid noise signals, and retaining clear and effective human voice signals to complete speech noise reduction optimization. The processed clean speech data will be used for subsequent decoding, word segmentation, and keyword recognition, effectively reducing the interference of environmental noise on the recognition results and improving the overall accuracy of speech recognition.

[0042] For example, in a kitchen setting, the environment is often accompanied by continuous low-frequency noise such as the operation of the range hood and the collision of kitchen utensils. The device uses a kitchen-specific noise reduction model to selectively filter out the preset low-frequency noise bands for this scenario. In cable storage and warehouse settings, the open space is prone to echoes and the friction noise of goods being handled. The corresponding noise reduction model will filter out these mid-to-high frequency interference bands. In an office setting, the main noises are conversations and the operation of office equipment. The device uses an office-specific noise reduction model to block the corresponding noise bands. For instance, when a kitchen worker voice-records "beef shank, shelf life one week" to create a food label, the original speech is mixed with range hood noise. After processing by the kitchen noise reduction model, the interference noise is filtered out, leaving only clear human voice content, providing a reliable voice data foundation for subsequent accurate extraction of label keywords.

[0043] Step S140: Determine whether the current speech recognition mode is network recognition mode or local recognition mode.

[0044] In this step, after completing the voice noise reduction processing in step S130, the device performs a voice recognition mode determination operation to distinguish between two working states: network recognition mode and local recognition mode, adapting to different usage environments without network access.

[0045] This step provides two types of independently effective mode determination trigger logic, which can be used alone or in combination: The first type is manual selection determination. The device's touch interface displays two mode options: network recognition and local recognition. Responding to the user's manual click selection, it directly locks the currently running voice recognition mode. The second type is automatic network connection status determination. The device monitors the network connection status and signal strength between the device and the cloud platform in real time. If the current network is detected to be disconnected, or the network signal strength is lower than a preset threshold, the local recognition mode is automatically triggered; if the network is detected to be normally connected and the signal strength meets the transmission standard, it automatically switches to network recognition mode.

[0046] After completing the mode determination, the device uses the recognition mode result as the basis for distinguishing subsequent branch processes: when it is determined to be the network recognition mode, it executes the cloud speech recognition and cloud word frequency table iterative update process; when it is determined to be the local recognition mode, it directly starts the local offline speech recognition process. The two processes are independent of each other and do not interfere with each other, realizing dual-mode adaptive switching and broadening the applicable scenarios of the device.

[0047] For example, taking a multi-area operation scenario in a park as an example, there are three actual working conditions: Working condition 1: Cable storage warehouse operation. The warehouse walls are thick and there is no network coverage. The equipment continuously cannot detect a valid network signal, so it automatically determines and triggers the local recognition mode. The entire process relies on the local storage of the scenario-specific user-personalized word frequency list to complete offline recognition, without uploading voice data to the cloud.

[0048] Operating mode 2: Kitchen operations. The kitchen Wi-Fi signal is stable, the equipment detects that the network meets the requirements, and automatically switches to network recognition mode by default. The user's voice is processed by noise reduction and then uploaded to the cloud for recognition and word frequency table updates.

[0049] Scenario 3: Operation in the integrated office area. Occasionally, the office network may experience fluctuations and disconnections. Users can manually click the "Local Recognition" option on the device screen to force a switch to local recognition mode, unaffected by network fluctuations, ensuring an uninterrupted label production process.

[0050] Regardless of whether automatic judgment or manual selection is used, the pattern judgment result only applies to the current tag generation task. The pattern detection will be re-executed after the next voice acquisition to ensure that each recognition can match the current network environment and user operation needs.

[0051] Step S150: When the network recognition mode is triggered, the noise-reduced user voice is sent to the cloud platform for voice recognition, the tag keywords are extracted, and the information of the tag keywords is stored in the user's personalized word frequency table on the cloud platform.

[0052] In this step, after completing the speech recognition mode determination in step S140 and confirming that the current mode is network-connected, the device initiates the cloud-based collaborative recognition process. The device simultaneously uploads the user's voice, which has been processed by noise reduction in step S130 and is free of background noise, along with the current target usage scenario identifier obtained in step S120, to the cloud platform via the wireless communication module. The cloud platform, equipped with a speech recognition model, transcribes, segments, and semantically analyzes the received speech, removes colloquial and redundant content, and extracts tag keywords for generating tags.

[0053] The cloud platform pre-maintains independent user-personalized word frequency tables for various usage scenarios, with the word frequency table data for different scenarios being isolated from each other. Based on the scenario identifier uploaded by the device, the cloud stores all identified tag keyword information into the cloud-based user-personalized word frequency table corresponding to the current target usage scenario, continuously collecting usage records specific to that scenario. Based on newly added keywords, the cloud updates the usage frequency, most recent usage time, and time decay weight of each keyword only within the independent word frequency table for that scenario, simultaneously adjusting the matching priority of each keyword within the word, and performing word replacement according to the priority, completing a full iterative update of the scenario-specific word frequency table. The cloud only distributes the updated single-scenario-specific user-personalized word frequency table to the device, without transmitting word frequency data from other scenarios.

[0054] The device has independent storage partitions for various usage scenarios, persistently storing user-personalized word frequency lists for each scenario. After receiving the word frequency list specific to the current scenario from the cloud, the device binds the word frequency list to the target usage scenario for this identification and stores it in the local partition of the corresponding scenario. This achieves partition management of independent word frequency lists corresponding to each usage scenario on the device. When the local identification mode is triggered subsequently, after the device identifies the target scenario, it retrieves the specific word frequency list bound in the corresponding local partition to complete the tag keyword identification.

[0055] For example, in a kitchen scenario, the device connects to the network and triggers cloud recognition, simultaneously carrying a "kitchen" scenario identifier when uploading voice. The cloud extracts keywords such as "beef shank, shelf life one week," stores the data only in a cloud-based independent word frequency table for the kitchen scenario, updates the parameters within the table, and only sends the kitchen scenario word frequency table to the device; the device stores this word frequency table locally in its kitchen partition and binds it to the kitchen scenario. Subsequently, when switching to a cable storage scenario for network use, the device uploads voice with a "cable storage" identifier. The cloud only operates on and sends a storage scenario-specific word frequency table, while the device stores it separately in its local storage partition. The two scenario word frequency tables are independent of each other in the cloud and on both ends of the device, without interfering with each other, achieving independent iteration and reuse of local storage for different scenarios.

[0056] Step S160: Update the usage frequency and recent usage time of each keyword in the user-personalized word frequency table according to the information of the tag keywords, and adjust the matching priority of each keyword in the user-personalized word frequency table.

[0057] This step is performed on the cloud platform, immediately following step S150 where tag keyword extraction and scene-specific word frequency table data writing are completed. It iterates only on the cloud-based user-personalized word frequency table corresponding to the current target usage scenario. The cloud reads all newly added tag keyword information and performs basic field update operations on each matched keyword in the word frequency table: incrementing the cumulative usage count of the keyword by one, and updating the most recent usage time of the keyword to the system time corresponding to this speech recognition. Simultaneously, the matching priority of each keyword in the user-personalized word frequency table is adjusted, thereby enabling real-time quantification of the usage popularity of each keyword in the corresponding scenario. This overcomes the shortcomings of existing general word frequency tables, which have fixed matching priorities and cannot adaptively adjust to changes in scenario-specific vocabulary.

[0058] Specifically, the cloud can recalculate the time decay weight of each keyword based on the updated usage frequency, the latest usage interval, and a preset decay coefficient. This weight comprehensively reflects the long-term usage frequency and recent usage activity of the keyword. The cloud uses the real-time calculated time decay weight as the evaluation criterion to dynamically rearrange and adjust the matching priority of all keywords in the same scenario-specific user-personalized keyword frequency table; the more times a word has been used cumulatively, the shorter the interval between uses, and the higher the time decay weight value, the higher the matching priority is assigned, and vice versa.

[0059] The time decay weight can be calculated based on the cumulative number of times a keyword has been used, the time interval since its last use, and the decay coefficient. Specifically, it can be based on the cumulative number of times a keyword has been used and the time since its last use, as recorded in the user's personalized keyword frequency table, and according to the formula... The calculated time decay weights are used to dynamically reorder and adjust the matching priority of all keywords in the user-personalized keyword frequency table corresponding to the target use scenario. The cumulative number of times representative words are used. The time interval between the last use of a word. As an adjustable decay coefficient, this weight can comprehensively reflect the overall usage frequency and recent activity level of the vocabulary.

[0060] This application abandons the traditional fixed word matching rules. Priority adjustment follows a unified logic: words with more cumulative usage, closer to their last usage time, and higher time decay weight values ​​represent words that better match the user's current word usage habits, and the device will assign them a higher matching priority. Conversely, words with fewer cumulative usage, long periods of inactivity, and low time decay weights will have their matching priority lowered accordingly. This priority adjustment only applies to the user's personalized word frequency list corresponding to the current usage scenario, and different scenarios are independent of each other. The adjustment results take effect immediately, providing dynamically adapted retrieval rules for subsequent speech decoding and word matching stages, which can significantly improve the recognition success rate of commonly used words.

[0061] For example, the original keyword frequency table for the kitchen scenario included two categories: "beef shank" and "steamer." This time, the online recognition extracted the keywords "beef shank, shelf life one week." The cloud system incremented the usage count of "beef shank" by one, refreshed the most recent usage time, recalculated it to obtain a higher time decay weight, and simultaneously increased its matching priority. Meanwhile, "steamer," which hadn't been used for a long time, didn't undergo this update, and its time decay weight continued to decrease, resulting in a corresponding drop in its matching priority. In the cable storage scenario, the high-frequency keyword "electrical wire" continuously accumulated usage records, maintaining a stable high priority, while the priority of low-frequency, less common material keywords gradually decreased, continuously adapting to the word usage patterns of daily tag creation in this scenario.

[0062] Step S170: Replace the words in the user's personalized word frequency table according to the priority of each keyword, and update the user's personalized word frequency table after word replacement.

[0063] This step is also performed on the cloud platform. Following step S160, which dynamically adjusts the matching priority of all keywords, a word selection and replacement operation is performed only on the cloud-based user-specific personalized word frequency table corresponding to the current target usage scenario. The cloud uses the keyword matching priority calculated in step S160 as the screening standard to distinguish between high-frequency, highly compatible words in the word frequency table and long-term idle, low-priority invalid words. Under the constraint of word frequency table storage capacity, low-priority, long-term infrequently used words are replaced with newly identified scenario tag words with higher matching priority, completing the iterative replacement of words within the word table and avoiding redundant and obscure words occupying word frequency table storage space for a long time.

[0064] After the word replacement operation is completed, the cloud synchronously refreshes all the stored content of the scene-specific user-personalized word frequency table, synchronously retaining the usage count, recent usage time, time decay weight, matching priority, and complete word list of each keyword after the update, forming a user-personalized word frequency table with the updated version. This provides a complete and up-to-date word frequency data carrier for subsequent distribution to the device, for local binding of the scene on the device, and offline recognition.

[0065] For example, the cloud-based word frequency table for the cable storage scenario has limited storage capacity. The table contains both high-priority terms like "electric wire" and "three-meter cable," and low-priority, long-unused terms like "insulating pad." This online recognition adds the high-priority term "flame-retardant corrugated pipe." Based on priority rules, the cloud replaces the low-priority "insulating pad" with the newly added "flame-retardant corrugated pipe." After the replacement, the cloud updates the entire user-personalized word frequency table for this storage scenario, fully storing all updated terms and corresponding statistical parameters for later distribution to the device's local storage and binding. Similarly, in the kitchen scenario, the long-unused, low-priority term "casserole" will be replaced by the newly added highly compatible food term. After the replacement, the entire word frequency table for the kitchen scenario is updated synchronously.

[0066] Step S180: Receive the tag keyword information and user-personalized word frequency table sent by the cloud platform, perform tag printing operation based on the tag keywords, and establish an association between the received user-personalized word frequency table and the current target usage scenario so that the tag keywords can be identified from the user-personalized word frequency table when the local recognition mode is triggered later.

[0067] This step is performed locally on the device and is the final step in the network recognition mode. The device continuously monitors downlink data from the cloud via the wireless communication module, receiving two types of data sent down from the cloud after iterative updates: the tag keyword information obtained in this recognition and the user-personalized word frequency table specific to the current target scenario.

[0068] The device first extracts the received label keyword information, and then drives the label printing module to complete the physical label printing according to the preset label layout, font, and size format, quickly outputting the item identification label and completing the label generation task.

[0069] After printing, the device establishes a one-to-one binding association between the scenario-specific user-personalized word frequency list downloaded from the cloud and the current target usage scenario identified in step S120, and stores it locally on the device. Through scenario binding, independent word frequency list partition management for each usage scenario is achieved locally on the device. If the network is subsequently disconnected and the device triggers local recognition mode, the device can first identify the target usage scenario, directly retrieve the corresponding scenario-specific user-personalized word frequency list already bound locally to perform word matching, extract tag keywords, and achieve accurate offline scenario-based recognition.

[0070] For example, after the kitchen scene completes cloud-based recognition, the device receives keyword information such as "beef shank, shelf life one week" and an updated kitchen scene word frequency list from the cloud. The device first prints food labels based on the keywords, then binds the word frequency list to the kitchen scene and stores it locally. Subsequently, when the warehouse switches to local recognition mode after a network outage, the device recognizes the target scene as the kitchen and directly retrieves the locally bound kitchen-specific word frequency list to complete voice recognition, without needing to reconnect to the network to synchronize data. The process is the same for the cable storage scene; the storage-specific word frequency list issued by the cloud is only bound to the cable storage scene, and the word frequency lists for different scenes are independent of each other and cannot be used interchangeably.

[0071] In this embodiment, user voice is first collected and the current target usage scenario is identified. A scenario-specific noise reduction model is then called to filter environmental noise in the corresponding frequency band, purifying the voice signal and reducing noise interference with the recognition results from the source. Simultaneously, it supports switching between online and local dual recognition modes in both manual and automatic network modes, adapting to various working environments without network coverage and thus having a wider range of applications. In network mode, the device uploads the noise-reduced voice with a scenario identifier. The cloud collects scenario-specific personalized word frequency lists by setting up independent isolation for each user. Based on the identified keywords, the frequency of word usage, recent usage time, and time decay weight are updated to dynamically adjust the matching priority of each word. Based on the priority, words within the word are replaced through a process of natural selection to continuously match the actual word usage characteristics of the corresponding scenario. The cloud only sends the updated personalized word frequency list for the current scenario to the device. The device directly uses the keywords to print labels and binds the word frequency list to the local target scenario for partitioned storage, enabling independent management of the word frequency list by scenario on the device side. When the local recognition mode is activated, the device can directly retrieve the local scene-specific word frequency table to complete keyword recognition offline. This overcomes the shortcomings of traditional offline recognition, which relies on a general fixed word library and cannot adapt to scene-specific words, effectively improving the accuracy of speech recognition and the overall efficiency of tag generation in various scenarios.

[0072] In some embodiments, the user-personalized word frequency table is managed in partitions according to different usage scenarios. Each usage scenario corresponds to an independent user-personalized word frequency table. The user-personalized word frequency table includes word ID, word text, usage scenario tag, number of times used, recent usage time, and time decay weight field.

[0073] Specifically, at the partition management level, the system implements isolated storage for various usage scenarios such as kitchens, cable storage, offices, and custom scenarios. Each target usage scenario is configured with a dedicated and independent user-personalized word frequency table. The word frequency table data corresponding to different scenarios are isolated from each other and are not mixed. Only the usage records of tagged words in the corresponding scenario are collected, which can accurately distinguish the specific words of each scenario and avoid cross-scenario word mixing from interfering with the recognition and matching effect. At the data table structure level, each independent user-personalized word frequency table uniformly sets six core fields: word ID serves as a unique identifier for each word entry, enabling quick word retrieval and differentiation; word text stores the original text content of the tagged word; usage scenario tag marks the usage scenario to which the current entry belongs, matching the partition management logic; usage count records the total number of times the word has been recognized and retrieved in the corresponding scenario; last usage time records the last time the word participated in speech recognition in the system; and time decay weight is calculated by combining the cumulative usage frequency of the word and the idle interval to quantify the recent activity level of the word. The entire set of fields retains all usage behavior data of the scenario-specific vocabulary, providing a complete and unified data basis for dynamically adjusting the vocabulary matching priority and performing word replacement within the word frequency table based on the priority.

[0074] Correspondingly, refer to Figure 2 , Figure 2 This is a second flowchart of a speech recognition-based tag generation method provided in an embodiment of this application. The method includes, but is not limited to, steps S210 to S260.

[0075] Step S210: Collect user voice.

[0076] Step S220: Confirm the current target usage scenario.

[0077] Step S230: Call the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's voice in order to filter out noise in the preset frequency band under the current target usage scenario.

[0078] Step S240: Determine whether the current speech recognition mode is network recognition mode or local recognition mode.

[0079] Step S250: When the local recognition mode is triggered, the noise-reduced user speech is decoded and segmented, and the user-personalized word frequency table corresponding to the current target usage scenario is called for matching and recognition to extract tag keywords.

[0080] Step S260: Perform label printing operation based on label keywords.

[0081] In this embodiment, after determining the speech recognition mode and confirming that the current mode is local, the device first decodes and segments the denoised speech signal, converting continuous audio into a text sequence. Based on Chinese semantic rules, it separates the audio into independent words and phrases, filtering out modal particles and meaningless colloquial redundancy to obtain matching segmentation units. Subsequently, the device reads the locally bound and stored scene identifiers, retrieves the user-specific word frequency table corresponding to the current target usage scene, and compares the segmentation results sequentially based on the preset matching priority of each word in the word frequency table. It prioritizes matching frequently used words with high activity and high weight, while filtering out irrelevant words not belonging to the current scene. Finally, it selects effective content with identifying functions, completing the tag keyword extraction. The entire operation does not rely on the cloud network, achieving accurate recognition based on the locally updated scene-specific word frequency table, adapting to offline operation scenarios such as warehouses and kitchens.

[0082] For example, when the cable storage warehouse loses network access, the local recognition mode is triggered. The noise-reduced speech is decoded and segmented to obtain "three meters, wire, spare". The device retrieves the local cable storage-specific word frequency table, prioritizes matching the high-priority words "three meters" and "wire", and removes the worthless word "spare" to extract the tag keywords. The same applies when the kitchen is operating offline. The device retrieves the local kitchen scene word frequency table, prioritizes matching high-frequency words such as "beef shank" and "shelf life" to complete the keyword extraction.

[0083] After keyword extraction, the device's local printing module completes the output. The device reads the label keywords obtained from offline recognition, generates printing data according to the device's preset label size, layout format, and font specifications, and drives the built-in printing mechanism to output physical label paper, printing the keywords onto the label medium, thus completing the production of identification labels required for item classification and material control. This step is consistent with the printing logic in network recognition mode. Regardless of whether the keywords come from cloud recognition or local offline recognition, labels can be output directly and quickly, ensuring a complete closed loop from voice input to physical label production even in environments without a network connection.

[0084] For example, after the local offline recognition extracts the keywords "3-meter cable, storage area 1", the equipment automatically typeset and prints cable material labels; when the kitchen offline recognition obtains "beef shank, shelf life one week", the equipment immediately prints fresh food storage labels to meet the on-site instant labeling needs.

[0085] The offline recognition process provided in this application first completes user voice acquisition, target usage scenario confirmation, and scenario-differentiated noise reduction in sequence. It filters out environmental noise in the corresponding scenario-specific frequency bands, improving the purity of the voice signal and laying the foundation for accurate offline recognition. After switching to local offline recognition through mode judgment, the device does not rely on cloud computing power and network transmission. It directly decodes and segments the denoised voice locally, retrieves the independent user-personalized word frequency table corresponding to the current scenario for word matching, and prioritizes the recognition of high-frequency words in the scenario based on the dynamically updated word priority within the word frequency table. Simultaneously, it filters out irrelevant words across scenarios, significantly improving the accuracy of keyword extraction in offline environments. After recognition, it can directly drive the printing module locally to output entity labels. The entire local recognition and printing process operates independently in a closed loop, unaffected by network disconnection or signal fluctuations. This adapts to offline operation scenarios such as warehousing and kitchens, broadening the device's applicable working conditions. Simultaneously, it reuses the scenario-specific word frequency table iteratively distributed from the cloud, eliminating the need to repeatedly build an offline dictionary. This balances offline recognition stability and scenario-specific word adaptability, effectively improving the continuity and efficiency of label generation in weak or offline network scenarios.

[0086] In some embodiments, refer to Figure 3 , Figure 3 This is a flowchart illustrating the steps of user-personalized word frequency table partitioning management according to an embodiment of this application. When user-personalized word frequency tables are partitioned according to different usage scenarios, storage capacity can be further allocated through steps S310 to S330.

[0087] Step S310: Obtain the usage count of the user-personalized word frequency table corresponding to the current target usage scenario based on the user-personalized word frequency table, and record and accumulate it.

[0088] In this step, after the device receives the user-personalized word frequency table from the cloud and binds it to the local target scenario, the device uses the current target usage scenario identifier as the retrieval index to retrieve the user-personalized word frequency table for that scenario stored in the local partition, reads and obtains the historical usage count value recorded in the word frequency table; whenever the device completes a complete tag generation process for that scenario (including voice acquisition, noise reduction, recognition, and printing), it is determined as a valid use, and the device automatically increments the usage count corresponding to the word frequency table for that scenario by one, synchronously updates the count record stored locally, and continuously accumulates the actual call frequency of the word frequency table for that scenario.

[0089] The usage counts for each scenario are independent and stored in separate partitions. Only the number of times the word frequency table is used in the corresponding scenario is counted, and there is no confusion with the counts of other scenarios. The usage counts accumulated over a long period of time can objectively reflect the actual usage frequency of the word frequency table in different scenarios, and serve as the core basis for determining the dynamic allocation of storage capacity quotas among various scenarios in the future.

[0090] For example, in a cable storage scenario, material labels are frequently produced daily. Each time a local or cloud-based recognition and printing process is completed, the usage count of the term frequency table for the cable storage scenario is incremented, and the value continues to rise. In contrast, custom scenarios have a lower usage frequency, and the count increases slowly. By recording and accumulating the independent usage counts for each scenario over a long period, the device can accurately distinguish between high-frequency and idle scenarios, providing quantitative data support for the balanced allocation of storage capacity resources.

[0091] Step S320: Based on the cumulative usage count of the user-personalized word frequency table corresponding to each usage scenario, calculate the usage frequency of the user-personalized word frequency table corresponding to each usage scenario.

[0092] This step follows the execution of the count accumulation result in step S310 and is used to quantify the actual activity level of the word frequency table in different scenarios. The device retrieves the cumulative usage count of all scenario word frequency tables stored in the local partition, combines it with the total number of uses within a preset statistical period, and calculates the usage frequency of the personalized word frequency table for each usage scenario independently to obtain the usage frequency of the user-specific word frequency table for the corresponding scenario.

[0093] Usage frequency is represented by the ratio of the cumulative usage count of a particular scenario to the total usage count of all scenarios. A higher value indicates that the scenario is frequently used to access the word frequency table and create tags, while a lower value indicates that the scenario is underutilized. The usage frequency of each scenario is calculated and stored independently, ensuring that the data does not interfere with each other.

[0094] The usage frequency calculated is the core quantitative indicator for dynamically allocating storage capacity. It can intuitively distinguish between high-frequency usage scenarios and low-frequency idle scenarios, providing an objective judgment standard for subsequent cross-scenario allocation of storage space.

[0095] For example, during the statistical period, the cable storage scenario was used 90 times, the kitchen scenario 60 times, the office scenario 30 times, and the total number of all scenarios was 180 times. It was calculated that the cable storage scenario had the highest usage frequency and the office scenario had the lowest usage frequency. Based on this frequency result, the equipment can allocate some of the storage quota of the idle scenario to the high-frequency operation scenario.

[0096] Step S330: Based on the usage frequency of the user-personalized word frequency table corresponding to each usage scenario, allocate part of the storage capacity of the user-personalized word frequency table with a usage frequency lower than the first predetermined threshold to the user-personalized word frequency table with a usage frequency higher than the second predetermined threshold.

[0097] This step follows the calculation results of the word frequency table usage in each scenario from step S320, and is used to dynamically and evenly allocate the total local storage space of the device. The device is pre-configured with two levels of thresholds: a first predetermined threshold is used as the criterion for determining low-frequency idleness, and a second predetermined threshold is used as the criterion for determining high-frequency capacity shortage.

[0098] The device iterates through the usage frequency of all scenarios, filters out the word frequency lists of idle scenarios whose usage frequency is below the first predetermined threshold and which are rarely used in daily life, and allocates a portion of the reserved storage capacity from them; then it filters out the word frequency lists of high-frequency scenarios whose usage frequency is above the second predetermined threshold, whose vocabulary is continuously expanding and whose storage space is tight, and allocates the allocated free capacity to the dedicated storage partition of this type of scenario.

[0099] The storage capacity of the word frequency tables for each scenario is managed independently. Allocation only transfers idle quotas and does not delete existing word data in low-frequency scenarios. It only reduces the reserved expansion space and expands the storage limit for adding new words in high-frequency scenarios, thus balancing the basic storage needs of low-frequency scenarios with the expansion needs of high-frequency scenarios. Through dynamic allocation, it avoids low-activity scenarios from occupying a large amount of idle storage for a long time, while alleviating the problem of insufficient capacity and inability to add new words in high-frequency scenarios, thereby improving the overall storage resource utilization of the device.

[0100] For example, the device sets a first predetermined threshold and a second predetermined threshold. The usage frequency of the custom scenario is lower than the first predetermined threshold, which is considered a low-frequency idle scenario. The usage frequency of the cable storage scenario is higher than the second predetermined threshold, with a long-term increase in material terms and insufficient storage capacity. The device allocates a portion of the reserved storage capacity from the custom scenario term frequency table to the dedicated partition for the cable storage scenario, expanding the upper limit of new terms that the storage scenario can accommodate. The original terms in the custom scenario are fully retained, with only the redundant reserved space reduced, thus realizing the dynamic allocation of storage resources on demand.

[0101] In this embodiment, adaptive allocation of word frequency table storage resources for each scenario is achieved through step-by-step statistics, calculation, and dynamic allocation. First, the usage count of the word frequency table for each scenario is continuously recorded and accumulated to objectively quantify the actual call frequency of each scenario. Then, the usage frequency of each scenario is calculated based on the cumulative usage count across all scenarios, forming a comparable quantitative evaluation index. Finally, low-frequency idle scenarios and high-frequency scarce scenarios are distinguished according to preset two-level thresholds, and the idle reserved storage capacity of low-frequency scenarios is allocated to high-frequency scenarios. This solution only adjusts the expansion quota for each scenario and does not delete the words already stored in low-frequency scenarios. This ensures that the basic recognition function of low-frequency scenarios is not affected, while expanding the upper limit of new word capacity for high-frequency usage scenarios. It effectively solves the problem of insufficient storage space and inability to add new words due to the continuous addition of words in high-frequency scenarios. At the same time, it releases long-term idle storage resources in low-frequency scenarios, avoids waste of storage space resources, significantly improves the overall local storage utilization rate of the device, and allows limited hardware storage resources to be tilted towards high-frequency operation scenarios, continuously ensuring the stable operation of multi-scenario and differentiated label production needs.

[0102] In some embodiments, each user-personalized word frequency table corresponding to each use scenario is allocated dedicated storage capacity. (See reference...) Figure 4 , Figure 4This is a flowchart of the steps for updating a user's personalized word frequency table according to an embodiment of this application, including but not limited to steps S410 to S430.

[0103] Step S410: Obtain the matching priority of each keyword in the user's personalized word frequency table, and sort all keywords in the user's personalized word frequency table according to the matching priority.

[0104] This step is performed before adding new keywords. The device reads the matching priority values ​​of all entries in the user's personalized word frequency table for the current scenario. This priority is calculated by combining the word's usage frequency, recent usage time, and time decay weight. The higher the value, the more suitable the word is for the daily word usage habits in this scenario. The device sorts all keywords uniformly from high to low according to their matching priority, with highly active and frequently used words at the top and long-term inactive and low-weight words at the bottom, providing a clear elimination order for word deletion after the capacity limit is exceeded.

[0105] For example, the frequency list of kitchen scene words includes "beef shank, shelf life, casserole, steamer". Among them, "beef shank" and "shelf life" have higher matching priority and are ranked higher; "casserole" and "steamer" are rarely used and have lower priority, so they are ranked lower.

[0106] Step S420: Determine whether the storage space occupied by all keywords exceeds the storage capacity allocated to the user's personalized keyword frequency table.

[0107] In this step, each time a new tag keyword needs to be written to the current scene's word frequency table, the device calculates the total storage space occupied by all existing words and the newly added words, and then compares it with the pre-allocated dedicated storage capacity limit for this scene. If the total occupied space is less than or equal to the dedicated capacity, it is determined that there is sufficient space; if the total occupied space exceeds the preset limit, it is determined that there is a storage overflow, and a word deletion operation needs to be performed.

[0108] For example, the dedicated storage capacity of the frequency list for the cable storage scenario can accommodate 20 sets of words. The original words occupied a total of 16 sets of space. This time, 6 new words were added, totaling 22 sets, which exceeded the capacity limit and triggered the elimination process.

[0109] Step S430: If the storage space occupied by all keywords exceeds the storage capacity allocated to the user's personalized keyword frequency table, then delete them sequentially starting from the keyword with the lowest matching priority, until the storage space occupied does not exceed the storage capacity allocated to the user's personalized keyword frequency table.

[0110] In this step, when storage capacity is determined to be overflowing, the device uses the priority sorting result from step S410 as a basis to delete complete word entries one by one from the lowest priority matching words at the end of the sort. Each deleted word simultaneously deducts the corresponding occupied storage space. This deletion operation is repeated until the total occupied space of all remaining words (including the new words to be stored this time) is less than or equal to the dedicated storage capacity for this scenario. The deletion operation only applies to the word frequency table of the current scenario and will not modify word frequency data for other scenarios; it prioritizes eliminating long-term idle and low-matching unpopular words, retaining high-frequency, high-priority words that fit the usage habits of the scenario.

[0111] For example, in the above cable storage scenario, a total of 22 sets of terms exceeded the limit of 20 sets. The device started deleting from the two least popular terms with the lowest priority. After the deletion, 20 sets of terms remained, and the space met the storage requirements. High-frequency material terms were completely retained.

[0112] Step S440: If the storage space occupied by all keywords does not exceed the storage capacity allocated to the user-personalized keyword frequency table, then all keywords will be directly stored in the user-personalized keyword frequency table.

[0113] If the total space occupied by the counted words does not exceed the scene-specific storage capacity, there is no need to perform word elimination. The device directly writes the original words and the newly added keywords into the user's personalized word frequency table corresponding to the current scene, and updates fields such as the number of times each word is used, the most recent use time, and the time decay weight, thus completing the word frequency table update. This simplifies the storage and writing process and eliminates the need to delete any words.

[0114] For example, the original vocabulary in the kitchen scene occupied 12 sets of space. This time, 3 more sets were added, for a total of 15 sets, which is less than the exclusive capacity limit of 20 sets. All vocabulary was directly stored in the word frequency table for complete retention.

[0115] In this embodiment, each scenario's word frequency table is independently configured with dedicated storage capacity, achieving storage isolation between scenarios. Before writing new words, words are uniformly sorted based on matching priority, with word scenario adaptability as the core criterion for elimination. When capacity overflows, only inactive and unpopular words are deleted, maximizing the retention of scenario-specific words frequently used by users and preventing the cleanup of key recognition words. Simultaneously, a two-pronged approach is implemented through pre-emptive capacity judgment: when space is sufficient, words are stored completely; when space is insufficient, redundant and low-value entries are automatically cleaned up. Under conditions of limited hardware storage resources, this ensures that the word frequency table continuously accommodates highly adaptable keywords, steadily improving the accuracy of local offline speech recognition.

[0116] In some embodiments, the user-personalized word frequency table is managed in partitions according to different users and different usage scenarios. Each user and each usage scenario corresponds to an independent user-personalized word frequency table. The user-personalized word frequency table includes word ID, word text, usage scenario tag, user ID, number of uses, most recent use time, and time decay weight field.

[0117] Specifically, at the partitioned management mechanism level, the word frequency table is bound to both user identifiers and usage scenario identifiers, employing a two-layer independent partitioning logic: first, different users are distinguished, and then various usage scenarios under a single user are further subdivided. Each unique combination of user and each usage scenario is configured with a completely independent, data-isolated, user-personalized word frequency table. Word frequency data between different users are not shared, and word records for different scenarios under the same user are also independent. When multiple operators such as warehouse managers, sorters, and logistics personnel share a single tag device, each person's word records for their own kitchen, cable storage, and office scenarios are collected separately, preventing situations where user A's warehouse vocabulary is mixed into user B's kitchen vocabulary. This completely isolates cross-data from multiple users and multiple scenarios, accurately matching the unique word usage habits of each individual.

[0118] At the data table field structure level, each independent word frequency table uniformly sets seven types of standardized core fields. Compared with the single-scenario-dimensional word frequency table, a user ID field is added. The functions of each field are clear: the word ID is a globally unique number for a single word, used for quick retrieval and location of the word; the word text stores the original text content of the tag keywords; the usage scenario tag marks the operation scenario to which this word belongs; the user ID marks the identity of the operator to which this word belongs; the usage count counts the total number of times the user has recognized and called this word in the corresponding scenario; the most recent usage time records the system time of the last speech recognition of the word; and the time decay weight is calculated by combining the usage frequency and idle interval to quantify the recent activity of the word.

[0119] The entire set of fields retains comprehensive usage behavior data across all dimensions of "personnel-scenario-vocabulary". It can distinguish the differences in word usage among different operators, as well as the word usage preferences of the same person in different scenarios. It provides a complete and distinguishable quantitative data source for dynamically calculating word matching priority by user and scenario, eliminating words with capacity limits, and dynamically allocating storage quotas.

[0120] For example, warehouse worker A and sorting worker B share equipment within a comprehensive warehousing park. Both use the cable storage and kitchen scenarios. The system creates two independent word frequency tables for worker A: "A-Cable Storage" and "A-Kitchen," and two independent word frequency tables for worker B: "B-Cable Storage" and "B-Kitchen." When warehouse worker A enters "3-meter flame-retardant wire" in the cable scenario, this term is only stored in A's corresponding warehouse word frequency table, and is simultaneously populated with fields such as exclusive user ID, scenario tag, usage frequency, and time decay weight. When sorting worker B enters "food storage box, refrigerated storage" in the kitchen, this term is only collected in B's kitchen-specific word frequency table. The four types of word frequency table data are completely isolated. Subsequent word frequency weight calculation, word priority sorting, and storage eviction only read exclusive entries under the corresponding user + scenario combination, avoiding cross-person and cross-scenario word interference in recognition and matching.

[0121] Correspondingly, after receiving the user-personalized word frequency list from the cloud platform, the device also determines the current user's identity and associates the received user-personalized word frequency list with the current target usage scenario and the current user's identity, so that when the local recognition mode is triggered later, the device can call the user-personalized word frequency list of the corresponding user identity in the corresponding target usage scenario to identify the tag keywords.

[0122] Specifically, after the device completes the cloud-based keyword recognition and printing operations, it receives a personalized word frequency list for the user, carrying the target scenario identifier, sent from the cloud. The device simultaneously retrieves the current user identity identifier and the current target usage scenario identifier obtained from the recognition, establishes a unique binding relationship between the personalized word frequency list sent from the cloud and the current user identity and the current target usage scenario, and then stores it in the independent storage partition of the corresponding user and scenario on the device.

[0123] This binding relationship is stored locally on the device as a local retrieval index. When the local recognition mode is triggered by a subsequent network outage, the device will simultaneously identify the identity of the operator and the target usage scenario to be identified. Using the dual identifiers of user identity and usage scenario as search conditions, it will accurately retrieve the user's personalized word frequency table that is pre-bound locally and exclusive to this combination to perform word segmentation and matching, and extract tag keywords.

[0124] The word frequency lists for different operators and different scenarios for the same operator are independently bound and stored, preventing issues such as cross-user mixing of word frequency lists or incorrect vocabulary databases across scenarios. Even if multiple staff members alternately share the same tagging device, offline recognition will only load the exclusive vocabulary data corresponding to the current operator's scenario. The matching logic aligns with the individual's unique word usage habits, avoiding recognition biases caused by the mixing of vocabulary from different personnel and scenarios.

[0125] For example, after warehouse worker A completes network-based identification in a cable storage scenario, the device receives a dedicated word frequency table for A-cable storage from the cloud and binds and stores this word frequency table with user A and the cable storage scenario. When there is no network connection, warehouse worker A uses the local cable storage for identification again. The device reads the dual identifiers of user A and the cable storage, retrieves the bound dedicated word frequency table, and completes the identification. If the system switches to a sorting worker B operating in a kitchen scenario, the device retrieves the previously bound user B-kitchen word frequency table. The two sets of bound data are independent and do not interfere with each other.

[0126] Correspondingly, refer to Figure 5 , Figure 5 This is a third flowchart of a speech recognition-based tag generation method provided in an embodiment of this application. The method includes, but is not limited to, steps S510 to S560.

[0127] Step S510: Collect the user's voice and confirm the current user's identity.

[0128] Step S520: Confirm the current target usage scenario.

[0129] Step S530: Call the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's voice in order to filter out noise in the preset frequency band under the current target usage scenario.

[0130] Step S540: Determine whether the current speech recognition mode is network recognition mode or local recognition mode.

[0131] Step S550: When the local recognition mode is triggered, the noise-reduced user speech is decoded and segmented, and the user's personalized word frequency table under the current target usage scenario is called for matching and recognition to extract tag keywords.

[0132] Step S560: Perform label printing operation based on label keywords.

[0133] In this embodiment, during the synchronous period of voice acquisition, the device can initiate the user identity recognition process in parallel. This embodiment can set three independently usable identity recognition methods to adapt to different operating habits and usage scenarios: The first is account login recognition, where the device receives the user's manually entered account, password, and other identity information, compares and verifies it with a locally pre-stored account database, and then determines the current user's identity; the second is local identifier selection recognition, where the device's touch interface displays all recorded local identity tags, responds to the user's click selection operation, and directly locks the current user; the third is voiceprint comparison recognition, where the device extracts voiceprint feature parameters from the acquired voice, compares them one by one with a locally pre-stored voiceprint sample library of each registered user, and automatically completes identity determination based on the feature matching results.

[0134] After identity verification is completed, the device will bind the current voice data with the current user's identity identifier. Since the user-personalized word frequency table, historical print records, and word time decay weights in this application are all stored and managed in partitions according to different user identities and different usage scenarios, the identity verification result is the core basis for subsequent retrieval of the dedicated word frequency table, dynamic adjustment of word matching priority, and iterative update of word frequency table data. This can completely isolate the usage data of different users and prevent the problem of cross-interference of word data when multiple people share the device.

[0135] For example, within a comprehensive warehousing park, multiple staff members, including warehouse managers, sorters, and logistics personnel, share a single label generation device. All personnel can pre-register their information: registering an independent account on a companion mobile app and syncing it to the device, while simultaneously recording their personal voiceprint characteristics. The device also creates a quick identity option for each individual locally. When creating material labels, staff can choose their authentication method as needed: those accustomed to account login first enter their personal account on the device interface to log in, then speak the label content; those seeking convenient operation directly click on their local identity identifier corresponding to their name on the screen, followed by voice recording; and those who use the device frequently simply speak the label, and the device automatically matches their identity using voiceprint recognition. Assuming warehouse staff member A completes identity verification and voice-records "electric wire, 3 meters, storage area 1," the device will collect all the voice data, subsequent recognition results, and label printing records into staff member A's dedicated data partition. Subsequent operations such as retrieving word frequency tables, calculating word usage frequency and time decay weights, and optimizing word matching priorities are all based solely on the historical usage data of person A, accurately matching their personal word usage habits; the usage data of other staff members are stored independently and do not affect each other, ensuring personalized recognition results for each user.

[0136] Therefore, after the device determines that it has entered the local recognition mode and completed the speech noise reduction process, the device performs decoding and word segmentation processing on the clean speech audio signal after scene noise reduction, converting the continuous speech waveform into a parsable text character sequence; then, according to Chinese semantic rules, it splits out independent words and phrases, filters out interjections and meaningless verbal redundancy, and outputs standardized word segmentation units as the basic material for word matching.

[0137] Combining the aforementioned user- and scenario-based dual-layer partitioning management mechanism, the device simultaneously reads the current user's identity and the current target usage scenario identifier corresponding to this identification. Using the user ID and scenario tag as a dual retrieval index, it accurately retrieves the uniquely bound, personalized word frequency table for each user from the local storage partition. Each word in the word frequency table has a pre-stored matching priority. The device compares and matches the word segmentation unit with the word text in the word frequency table in descending order of priority; it prioritizes matching frequently used words with high frequency and weight in this scenario for the user, filtering out irrelevant words that are cross-scenario or not in the user's habit, and finally selecting effective content that serves as a material / item identifier, completing the tag keyword extraction. After keyword extraction is completed, the device's local printing module completes the output.

[0138] This application embodiment relies on a dedicated word frequency table that is iterated in the cloud during the network phase and bound to the device for local storage to achieve offline recognition. It does not require real-time uploading of voice and is suitable for working environments with no or weak network, such as warehouses and kitchens. At the same time, it relies on a dual-dimensional index of user and scenario to accurately retrieve the dedicated word frequency table, which can avoid the use of other people's or other scenario words when multiple people share the device, and reduce the probability of recognition mismatch.

[0139] For example, in a cable storage scenario with no internet connection, a warehouse worker initiates voice input. The noise-reduced speech is decoded and segmented to obtain the words "flame-retardant corrugated pipe, three meters, spare". The device reads the current user A and the cable storage scenario identifier, retrieves the locally bound A-cable storage-specific word frequency table, prioritizes matching the high-priority words "flame-retardant corrugated pipe" and "three meters", and removes the word "spare" which has no recognition value. Finally, the effective tag keywords are extracted for subsequent printing. If the system switches to offline recognition in a kitchen scenario for sorting personnel B, the device retrieves the independent word frequency table corresponding to B-kitchen to complete the matching. The two sets of word frequency data do not interfere with each other.

[0140] In some embodiments, refer to Figure 6 , Figure 6 This is a flowchart of the steps performed when storing tag keyword information to a user-personalized word frequency table on a cloud platform, as provided in an embodiment of this application, including but not limited to steps S610 to S630.

[0141] Step S610: Determine whether the remaining storage space of the user's personalized word frequency table can accommodate new words.

[0142] In this step, the cloud first locates the user ID and usage scenario tags corresponding to the uploaded data, retrieves the user-specific personalized word frequency table under this matching relationship, reads the preset total storage capacity and currently occupied storage space of the word frequency table, and calculates the remaining available storage space; then, it calculates the total storage size required after all the tag keywords to be added are written, and compares the two. If the remaining space is greater than or equal to the space required for the new words, the capacity is considered sufficient; otherwise, the space is considered insufficient, and the word filtering and interception logic needs to be executed.

[0143] Step S620: If the remaining storage space of the user-personalized word frequency table cannot accommodate all the tag keywords, then the target words with time decay weights less than or equal to the preset threshold and the complete entries corresponding to the target words will not be written into the user-personalized word frequency table.

[0144] In this step, when the cloud determines that the storage capacity is insufficient to store all the keywords to be uploaded, the cloud extracts the time decay weight values ​​of all the entries to be written this time, and uses the pre-set weight threshold as the filtering boundary; any word whose weight value is lower than or equal to the threshold means that the word has low overall usage frequency, has been idle for a long time, and has poor scene adaptability, and is directly blocked and not stored in the cloud word frequency table; only words with time decay weight higher than the preset threshold and higher activity are retained and written into the table.

[0145] The selection criteria use quantitative indicators of long-term word usage behavior as the selection standard. Given the limited cloud space, it prioritizes the retention of high-frequency core words that fit the user's and scenario's usage habits, while discarding unpopular and low-value words to avoid invalid entries occupying cloud storage resources.

[0146] Step S630: If the remaining storage capacity of the user-personalized word frequency table can accommodate all the tag keywords, then directly write the tag keywords into the user-personalized word frequency table and create new entries corresponding to the tag keywords.

[0147] In this step, if the remaining storage space in the cloud-based word frequency table is sufficient after verification, there is no need to delete or intercept any new words. The cloud will create an independent and complete word entry for each tag keyword to be written, and fill in all fields in sequence, including word ID, word text, usage scenario tag, user ID, number of uses, most recent use time, and time decay weight, to complete the word entry into the database. If the keyword already exists in the word frequency table, the statistical fields of the original word entry will be updated directly, and there is no need to create a new word entry.

[0148] For example, warehouse worker A identifies three new words in a cable storage scenario: "insulating tape," "small clips," and "spare gaskets." The cloud retrieves user A's dedicated word frequency table for cable storage and calculates the remaining space. It determines that there is insufficient space to store all three. The system compares the time decay weights of the three word groups. Since the weight of "spare gaskets" is below a preset threshold, this word is blocked and not written to the cloud. Only the two high-weight words, "insulating tape" and "small clips," are stored in the word frequency table. If the user adds "flame-retardant conduit" in the next identification in the same scenario, and the cloud word frequency table has sufficient remaining capacity, a complete word entry for "flame-retardant conduit" is created directly, and all fields are synchronously entered into the cloud word frequency table.

[0149] In this embodiment, the cloud adds a storage space prediction mechanism before writing new words to avoid single-user, single-scenario word frequency table storage overflow failure in advance; when space is insufficient, low-activity words are filtered and intercepted for storage based on time decay weight, rather than being randomly deleted, so as to retain the most frequently used and scenario-adaptable tag words; when space is sufficient, all new words are fully entered, and the user's full word usage behavior data is fully recorded, taking into account both cloud storage resource management and word frequency table data validity, and ensuring the accuracy of subsequent cloud computing matching priority and local offline recognition.

[0150] In some embodiments, the user-personalized word frequency table on the device adopts a three-level storage structure: a core word area, a high-frequency word area, and a cached word area. The core word area stores control instruction words, numeric words, and date words; words stored in the core word area cannot be deleted. The high-frequency word area stores the top preset number of words sorted by time decay weight from highest to lowest. The cached word area stores newly added words whose frequency has not yet been established; the cached word area is managed using a first-in, first-out (FIFO) storage strategy.

[0151] Specifically, in this embodiment of the application, the local user-personalized word frequency table of the device is divided into three-level isolated storage partitions. Storage rules are differentiated according to word function, activity level, and usage stage. Different partitions have independent storage limits, write, eviction, and deletion constraint logic to achieve hierarchical and refined management of words, taking into account both the stability of basic recognition and the needs of dynamic new word iteration.

[0152] The core term area is specifically designed to store essential basic and general vocabulary for label recognition, including equipment control instruction terms, various numbers, and date terms such as year, month, and day. It is an indispensable basic text unit for generating standard material labels.

[0153] This partition is equipped with a rigid protection mechanism, permanently locking all words within it. Deletion will not be performed regardless of any scenario, such as exceeding storage capacity limits, word obsolescence, or cache clearing. Core vocabulary forms the underlying support for speech-to-text transcription and tag formatting; prohibiting deletion prevents recognition failures and tag formatting errors caused by the loss of fundamental vocabulary, ensuring the stability and reliability of basic recognition functions.

[0154] For example, words such as "print", "cancel", "1 meter", "2026" and "May 17" are uniformly stored in the core word area, and no storage cleanup process will remove such entries.

[0155] The high-frequency word section stores highly relevant business terms for corresponding users and scenarios. The selection criteria are based on time decay weight, sorted from high to low, with a preset number of terms at the top. The time decay weight is calculated by combining cumulative usage frequency, idle interval, and decay coefficient. A higher value indicates that the term has been used more frequently recently and has a stronger scenario matching degree.

[0156] This partition stores business-specific vocabulary used by users to create tags daily. When matching and recognizing, it will prioritize searching for words in the high-frequency word area, greatly improving keyword hit efficiency. When the cloud sends an updated word frequency table and reorders the weights, the system will simultaneously refresh the words in the high-frequency word area, removing words whose weights have fallen below a preset number and moving in new words whose weights have risen, dynamically synchronizing with users' real-time word preferences.

[0157] For example, in the cable storage scenario, "flame-retardant corrugated pipe" and "3-meter cable" are used frequently over a long period of time and have a higher weighting over time, so they are stored in the high-frequency word area; while unpopular material words that are idle for a long time and have a continuously declining weighting will be removed from this area.

[0158] The cached word area is specifically designed to accommodate brand-new business words that have just been synchronized from the cloud, are initially identified locally, have not yet accumulated usage frequency, and whose time decay weight has not been stably established. It is a temporary transitional storage area.

[0159] The cached word area adopts a first-in, first-out (FIFO) management strategy and has an independent storage capacity limit. When new words are continuously added and the cache space is full, the oldest word with the fewest cumulative usages is automatically removed to make room for the latest recognized words. After a new word is recognized multiple times in the cache area, its usage frequency and weight continue to increase. Once it reaches the admission criteria, it will be migrated to the high-frequency word area. If it remains idle for a long time without being called, it will be cleared in the cache eviction process.

[0160] For example, when a newly added material "refractory wire trough" is first identified and synchronized locally, it is first stored in the cache word area; after multiple printings and uses, its weight increases and it is automatically transferred to the high-frequency word area; if it is stored and not used for a long time, the word will be cleared first when the cache space is insufficient.

[0161] Reference Figure 7 , Figure 7 This is a flowchart of the steps performed before performing a label printing operation based on label keywords, provided in an embodiment of this application, including but not limited to steps S710 to S730.

[0162] Step S710: Display the tag keywords and determine whether the user confirms or deletes the recognition result.

[0163] In this step, after the device extracts the tag keywords from the speech recognition output, it first displays the fully recognized text on the touch screen, awaiting user verification. The device listens to user commands in real time, distinguishing between two types of interactions: a confirmation operation, indicating that the recognized keywords are accurate and the user can proceed directly to the printing process; and a deletion operation, indicating that the recognition result has errors and the user needs to clear the existing text and re-enter the speech. This step serves as the entry point for human-machine verification, capturing the user's direct feedback on recognition accuracy, and forming the basis for negative feedback statistics.

[0164] For example, the speech recognition outputs "insulating pad" and displays it in a pop-up window. After the user checks the wording, if it is correct, they can click to confirm; if the recognition is incorrect, they can click to delete and clear the text.

[0165] Step S720: If it is detected that the number of times the user performs deletion and re-voice input operations on the same speech recognition result within a preset time period is greater than or equal to the first preset number, then a negative feedback event is recorded, and the association relationship between the original recognition word corresponding to this speech recognition and the label keyword finally confirmed and printed by the user is established.

[0166] In this step, the device has a built-in timing window and operation counter, which limit a preset duration as the statistical interval. For the same round of speech recognition task, if the user repeatedly deletes the incorrect recognition results and re-records the recognition, and the number of such loop operations reaches or exceeds the first preset number, the system determines that there is a significant recognition error in the current original recognition word, and automatically generates and stores a negative feedback event record.

[0167] At the same time, a mapping relationship is established: the original recognized words (incorrect words) output by this algorithm are bound to the standard keywords that are finally confirmed by the user after multiple corrections and used for printing, and the comparison data of incorrect words and correct words is retained to provide a comparison sample for subsequent correction.

[0168] For example, the first preset number of attempts is set to 2. After the user identifies "insulating gasket", they delete and re-enter the word more than 2 times in a row. Finally, the correct word "insulating gasket" is confirmed. The system records a negative feedback and binds the mapping relationship: original identified word = insulating gasket, standard keyword = insulating gasket.

[0169] Step S730: Accumulate statistics on negative feedback events for the same original recognition word. When the cumulative number of negative feedback events corresponding to the original recognition word is greater than or equal to the second preset number, add an optimization mark to the original recognition word for subsequent adjustment of the matching priority of the original recognition word, correction of the speech recognition model and updating of the user's personalized word frequency table.

[0170] In this step, the device categorizes and collects all negative feedback events according to the original recognized words, and continuously accumulates the total number of negative feedback events for the corresponding erroneous words. When the cumulative number of negative feedback events for a certain original recognized word reaches the second preset threshold, it indicates that the word has long-term recognition confusion and matching deviation problems, and the word is automatically marked as needing optimization.

[0171] This tag can be used in multiple subsequent stages: in the cloud, the matching priority of the word can be reduced to decrease the probability of priority hit during recognition; during the training stage, it can be used as a sample to correct the basic speech recognition model and improve the problem of confusion in the recognition of homophones and near-homophones; when updating the user's personalized word frequency list, the time decay weight of the word can be reduced, and it can even be cleaned up first when storage capacity is insufficient, thereby reducing the probability of recognition error from multiple levels.

[0172] For example, if the second preset number of times is set to 3, and the word "insulating pad" generates negative feedback 3 or more times, the system will add an optimization mark to it; its priority will be reduced during subsequent recognition and matching, and the recognition model will be optimized synchronously in the cloud to reduce the recurrence of similar recognition errors.

[0173] In this embodiment, a manual verification step is added before printing to promptly intercept and identify incorrect labels, preventing the direct printing of incorrect labels from causing confusion in material labeling and wasting consumables. Recognition defects are quantitatively determined by user behavior of repeatedly deleting and re-entering data, objectively collecting negative feedback from real-world usage scenarios. This eliminates the need for users to manually submit feedback, reducing manual operation costs. By establishing a binding mapping between incorrect words and standard keywords, correction samples are accumulated in batches; high-frequency incorrect words are automatically marked and optimized, and a three-layer mechanism—matching priority, speech model, and user word frequency table—is simultaneously optimized, forming a closed loop of "recognition error—user feedback—automatic marking—global optimization," continuously reducing the probability of confusion with homophones and near-homophones, and improving speech recognition accuracy in the long term.

[0174] Reference Figure 8 , Figure 8 This is a structural block diagram of a label generating device provided in one embodiment of this application. This application also provides a label generating device 800, which includes: The voice acquisition module 810 is used to acquire user voice.

[0175] The scenario confirmation module 820 is used to confirm the current target usage scenario.

[0176] The noise reduction processing module 830 is used to call the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's voice, so as to filter out noise in the preset frequency band under the current target usage scenario.

[0177] The wireless communication module 840 is used to establish a communication connection between the tag generation device and the cloud platform.

[0178] The recognition mode selection module 850 is used to determine whether the current speech recognition mode is network recognition mode or local recognition mode.

[0179] The network recognition and processing module 860 is used to send the noise-reduced user voice to the cloud platform when the network recognition mode is triggered, so that the cloud platform performs the following operations: Speech recognition is performed on the user's voice, tags and keywords are extracted, and the information of the tags and keywords is stored in the user's personalized word frequency table on the cloud platform; Update the frequency of use and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user's personalized word frequency table are replaced according to the priority of each keyword, and the user's personalized word frequency table is updated after the word replacement.

[0180] The data receiving and association module 870 is used to receive the information of tag keywords and user-personalized word frequency table sent by the cloud platform, and to establish an association between the received user-personalized word frequency table and the current target usage scenario, so that the tag keywords can be identified from the user-personalized word frequency table when the local recognition mode is triggered.

[0181] The label printing module 880 is used to perform label printing operations based on label keywords.

[0182] The specific implementation of the tag generation device 800 is basically the same as the specific implementation of the tag generation method based on speech recognition described above, and will not be repeated here.

[0183] In this embodiment, the label generation device 800 relies on the collaborative work of various functional modules. First, it improves the quality of the voice signal through voice acquisition, scene recognition, and scene-based noise reduction. Then, it uses a wireless communication module to achieve data interaction with the cloud. The recognition mode selection module 850 distinguishes between networked and local working modes. In networked mode, complex recognition and iterative word frequency table calculations are handled by the cloud, reducing the local computing power burden and lowering hardware configuration costs. The device receives updated word frequency tables from the cloud and stores them in conjunction with corresponding scenes. When the network is offline, it can independently complete recognition and printing based on the local word frequency table, balancing the cloud's adaptive learning optimization capabilities with offline operation stability. By storing scene-specific word frequency tables locally, it can match the word usage habits of different scenes, improving the accuracy of offline keyword recognition. The entire device has a clear structure and well-defined division of labor, adaptable to various operating environments with different network conditions, such as warehouses and kitchens, improving label generation efficiency and the device's applicability.

[0184] Furthermore, the recognition mode selection module 850 is specifically used to identify whether the current voice recognition mode is a network recognition mode or a local recognition mode in response to the user's manual selection of the voice recognition mode. Alternatively, the recognition mode selection module 850 is used to identify whether the current voice recognition mode is a network recognition mode or a local recognition mode based on the current network connection status. If the current network connection is disconnected or the network signal strength is lower than a preset threshold, the local recognition mode is triggered; if the current network connection is normal, the network recognition mode is triggered.

[0185] Furthermore, continue to refer to Figure 8 The tag generation device 800 also includes a local recognition processing module 890. When the local recognition mode is triggered, the local recognition processing module 890 decodes and segments the denoised user speech, and calls the user-personalized word frequency table corresponding to the current target usage scenario for matching and recognition to extract tag keywords.

[0186] Furthermore, continue to refer to Figure 8 The tag generation device 800 also includes a user identity verification module 8100. The user identity verification module 8100 is used to verify the identity of the current user. Correspondingly, the local recognition processing module 890 is also used to decode and segment the noise-reduced user speech when the local recognition mode is triggered, and to call the user's personalized word frequency table in the current target usage scenario for matching and recognition to extract tag keywords.

[0187] Reference Figure 9 , Figure 9This is a schematic diagram of a tag generation system provided in one embodiment of this application. This application also provides a tag generation system comprising a tag generation device 800 as described in any of the foregoing embodiments and a cloud platform 100. A communication connection is established between the tag generation device 800 and the cloud platform 100 to achieve bidirectional data interaction. The cloud platform 100 is configured with user-personalized word frequency table storage resources. It can receive noise-reduced user speech uploaded by the tag generation device 800, perform speech recognition processing on the speech to extract tag keywords, and simultaneously store the extracted tag keyword information into the corresponding user-personalized word frequency table on the cloud side. The cloud platform 100 can also update the usage frequency, recent usage time, and time decay weight of each keyword in the user-personalized word frequency table based on the identified tag keyword information, and adjust the matching priority of all keywords based on the updated weight parameters. The cloud platform 100 completes the replacement and elimination operation of words within the word frequency table according to the matching priority order from high to low, and completes the overall data update of the user-personalized word frequency table. After completing all the above calculation and storage update processes, the cloud platform 100 will send the identified tag keyword information and the updated user-personalized word frequency table to the tag generation device 800 for local storage and binding, and to support subsequent business processes such as offline local recognition and tag printing.

[0188] In this embodiment, the label generation system delegates complex calculations such as speech recognition, vocabulary weight updates, matching priority adjustments, and dynamic iteration of the word frequency table to the cloud platform 100. This effectively reduces the computing and storage pressure on the label generation device 800, lowering hardware and maintenance costs. The cloud platform 100 can update the usage parameters and matching priorities of keywords in the personalized word frequency table in real time based on each user's speech recognition behavior, automatically eliminating low-priority invalid words. It continuously learns the personalized word usage habits of different users and in different scenarios, dynamically optimizing speech recognition accuracy. This solves the problems of poor adaptability and inability to iterate autonomously in traditional fixed word libraries. Simultaneously, the cloud platform 100 synchronously distributes the latest iterated word frequency table to the label generation device 800 and stores it locally, enabling the device to independently complete offline speech recognition and label printing in offline or weak network environments. This balances network-based iterative optimization capabilities with offline operational stability, adapting to complex offline operation scenarios such as warehousing and kitchens. The cloud platform's centralized optimization and simultaneous terminal deployment working mode can complete the upgrade of word frequency tables for multiple terminals in batches without the need for manual maintenance on each terminal. This can effectively improve the system's iteration efficiency and scenario adaptability, forming a continuously optimized intelligent tag generation closed-loop system.

[0189] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A tag generation method based on speech recognition, characterized in that, include: Collect user voice; Identify the current target use case; The noise reduction model corresponding to the current target usage scenario is invoked to perform noise reduction processing on the user's voice, so as to filter out noise in the preset frequency band under the current target usage scenario; Determine whether the current speech recognition mode is network-based or local. When the network recognition mode is triggered, the noise-reduced user voice is sent to the cloud platform for speech recognition, the tag keywords are extracted, and the tag keyword information is stored in the user's personalized word frequency table on the cloud platform; Update the usage frequency and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user-personalized word frequency table are replaced according to the priority of each keyword, and the user-personalized word frequency table is updated after the word replacement. The system receives tag keyword information and the user-personalized word frequency table from the cloud platform, performs tag printing operation based on the tag keywords, and establishes an association between the received user-personalized word frequency table and the current target usage scenario so that the tag keywords can be identified from the user-personalized word frequency table when the local recognition mode is triggered.

2. The method according to claim 1, characterized in that, Determining the current target use case includes: In response to the user's manually selected scenario options, determine the current target usage scenario; Alternatively, feature analysis can be performed on the collected user voice data to identify the current target usage scenario; Alternatively, based on the current user's historical usage records and word usage habits, the current target usage scenario can be identified.

3. The method according to claim 1, characterized in that, The determination of whether the current speech recognition mode is network recognition mode or local recognition mode includes: In response to the user's manual selection of a voice recognition mode, identify whether the current voice recognition mode is a network recognition mode or a local recognition mode; Alternatively, the system can identify whether the current voice recognition mode is network recognition mode or local recognition mode based on the current network connection status. If the current network connection is disconnected or the network signal strength is lower than a preset threshold, the local recognition mode is triggered; if the current network connection is normal, the network recognition mode is triggered.

4. The method according to claim 1, characterized in that, The user-personalized word frequency table is managed in different usage scenarios. Each usage scenario corresponds to an independent user-personalized word frequency table. The user-personalized word frequency table includes word ID, word text, usage scenario tag, number of times used, recent usage time, and time decay weight field.

5. The method according to claim 4, characterized in that, After invoking the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's speech to filter out noise in the preset frequency band under the current target usage scenario, the method further includes: When the local recognition mode is triggered, the user's voice after noise reduction is decoded and segmented, and the user's personalized word frequency table corresponding to the current target usage scenario is called for matching and recognition to extract tag keywords; Perform a label printing operation based on the label keywords.

6. The method according to claim 4, characterized in that, After receiving the tag keyword information and the user-personalized word frequency table issued by the cloud platform, the method further includes: Based on the user-personalized word frequency table, obtain the usage count of the user-personalized word frequency table corresponding to the current target usage scenario, and record and accumulate it; Based on the cumulative usage count of the user-personalized word frequency table corresponding to each usage scenario, calculate the usage frequency of the user-personalized word frequency table corresponding to each usage scenario. Based on the usage frequency of the user-personalized word frequency table corresponding to each usage scenario, a portion of the storage capacity of the user-personalized word frequency table with a usage frequency lower than the first predetermined threshold is allocated to the user-personalized word frequency table with a usage frequency higher than the second predetermined threshold.

7. The method according to claim 4, characterized in that, Each user-personalized word frequency table corresponding to each use scenario is allocated dedicated storage capacity, and the update methods for the user-personalized word frequency table include: Obtain the matching priority of each keyword in the user-personalized word frequency table, and sort all keywords in the user-personalized word frequency table according to the matching priority; If the storage space occupied by all the keywords exceeds the storage capacity allocated to the user-personalized word frequency table, then the keywords will be deleted sequentially from the lowest matching priority according to the sorting order until the storage space occupied does not exceed the storage capacity allocated to the user-personalized word frequency table.

8. The method according to claim 1, characterized in that, The user-personalized word frequency table is managed in partitions according to different users and different usage scenarios. Each user and each usage scenario corresponds to an independent user-personalized word frequency table. The user-personalized word frequency table includes word ID, word text, usage scenario tag, user ID, usage frequency, recent usage time, and time decay weight field. Correspondingly, after receiving the user-personalized word frequency table issued by the cloud platform, the method further includes: Determine the current user's identity; The received user-personalized word frequency table is associated with the current target usage scenario and the current user identity, so that when the local recognition mode is triggered, the user-personalized word frequency table of the corresponding user identity under the corresponding target usage scenario can be called to identify the tag keywords.

9. The method according to claim 8, characterized in that, The confirmation of the current user's identity includes: Receive account information entered by the user to complete account login and confirm the current user's identity; Alternatively, it can identify the current user's identity in response to a local identity manually selected by the user. Alternatively, the collected user voice can be compared with pre-stored user voiceprint samples to identify the current user's identity.

10. The method according to claim 1, characterized in that, The step of updating the usage frequency and recent usage time of each keyword in the user-personalized keyword frequency table based on the information of the tag keywords, and adjusting the matching priority of each keyword in the user-personalized keyword frequency table, includes: The frequency of use, recent use time and time decay weight of each keyword in the user's personalized keyword frequency table are updated according to the information of the tag keywords. The time decay weight is calculated based on the cumulative number of times the keyword is used, the time interval since the last use, and the decay coefficient. Based on the updated time decay weight, the matching priority of each keyword in the user-personalized word frequency table is adjusted.

11. The method according to claim 10, characterized in that, When storing tag keyword information into a user-personalized word frequency table on a cloud platform, the method further includes: Determine whether the remaining storage space of the user-personalized word frequency table can accommodate all the tag keywords; If the remaining storage space of the user-personalized word frequency table cannot accommodate all the tag keywords, then the target words with time decay weight less than or equal to the preset threshold and the complete entries corresponding to the target words will not be written into the user-personalized word frequency table. If the remaining storage capacity of the user-personalized word frequency table can accommodate all the tag keywords, then the tag keywords are directly written into the user-personalized word frequency table, and a new entry corresponding to the tag keywords is created.

12. The method according to claim 10 or 11, characterized in that, The user-personalized word frequency table adopts a three-level storage structure, namely the core word area, the high-frequency word area, and the cache word area; The core word area is used to store control instruction words, number words, and date words. The words stored in the core word area cannot be deleted. The high-frequency word area is used to store a preset number of words that are sorted from high to low according to the time decay weight; The cache word area is used to store newly added words whose word frequencies have not yet been established, and the cache word area is managed using a first-in-first-out storage strategy.

13. The method according to claim 1, characterized in that, Before performing the label printing operation based on the label keywords, the method further includes: The tag keywords are displayed, and it is determined whether the user confirms or deletes the recognition results; If it is detected that the number of times the user performs deletion and re-voice input operations on the same speech recognition result within a preset time period is greater than or equal to the first preset number, a negative feedback event is recorded, and an association relationship is established between the original recognized word corresponding to this speech recognition and the label keyword finally confirmed and printed by the user. The negative feedback events of the same original recognition word are cumulatively counted. When the cumulative number of negative feedback events corresponding to the original recognition word is greater than or equal to the second preset number, the original recognition word is marked as needing optimization, so as to adjust the matching priority of the original recognition word, correct the speech recognition model and update the user's personalized word frequency table.

14. A label generating device, characterized in that, include: The voice acquisition module is used to collect user voice data. The scenario confirmation module is used to confirm the current target usage scenario; The noise reduction processing module is used to call the noise reduction model corresponding to the current target usage scenario to perform noise reduction processing on the user's voice, so as to filter out noise in the preset frequency band under the current target usage scenario; A wireless communication module is used to establish a communication connection between the tag generation device and the cloud platform; The recognition mode selection module is used to determine whether the current speech recognition mode is network recognition mode or local recognition mode; The network recognition processing module is used to send the noise-reduced user voice to the cloud platform when the network recognition mode is triggered, so that the cloud platform performs the following operations: The user's voice is subjected to speech recognition, and the tag keywords are extracted and stored in the user's personalized word frequency table on the cloud platform; Update the usage frequency and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user-personalized word frequency table are replaced according to the priority of each keyword, and the user-personalized word frequency table is updated after the word replacement. The data receiving and association module is used to receive the tag keyword information and the user personalized word frequency table sent by the cloud platform, and to establish an association between the received user personalized word frequency table and the current target usage scenario, so that the tag keywords can be identified from the user personalized word frequency table when the local recognition mode is triggered in the future. The label printing module is used to perform label printing operations based on the label keywords.

15. The label generating device according to claim 14, characterized in that, The recognition mode selection module is specifically used for: In response to the user's manual selection of a voice recognition mode, identify whether the current voice recognition mode is a network recognition mode or a local recognition mode; Alternatively, the system can identify whether the current voice recognition mode is network recognition mode or local recognition mode based on the current network connection status. If the current network connection is disconnected or the network signal strength is lower than a preset threshold, the local recognition mode is triggered; if the current network connection is normal, the network recognition mode is triggered.

16. The label generating device according to claim 14, characterized in that, The device also includes: The local recognition processing module is used to decode and segment the noise-reduced user speech when the local recognition mode is triggered, and to call the user's personalized word frequency table corresponding to the current target usage scenario for matching and recognition, and extract tag keywords.

17. The label generating device according to claim 16, characterized in that, The device also includes: The user identity verification module is used to verify the identity of the current user; Correspondingly, the local identification processing module is also used for: When the local recognition mode is triggered, the noise-reduced user speech is decoded and segmented, and the user's personalized word frequency table under the current target usage scenario is called for matching and recognition to extract tag keywords.

18. A label generation system, characterized in that, Includes the label generating device and cloud platform as described in any one of claims 14-17, wherein the label generating device is communicatively connected to the cloud platform; The cloud platform is used for: The system receives the noise-reduced user voice sent by the tag generation device, performs speech recognition on the user voice, extracts tag keywords, and stores the tag keyword information in the user's personalized word frequency table on the cloud platform. Update the usage frequency and recent usage time of each keyword in the user's personalized keyword frequency table based on the information of the tag keywords, and adjust the matching priority of each keyword in the user's personalized keyword frequency table; The words in the user-personalized word frequency table are replaced according to the priority of each keyword, and the user-personalized word frequency table is updated after the word replacement. The tag keyword information and the user-personalized word frequency table are sent to the tag generation device.

Images