A bidirectional real-time translation system, device and method based on exclusive voiceprint and hardware identity

Abstract

The application relates to a bidirectional real-time translation system, equipment and method based on exclusive voiceprints and hardware identity, which comprises target individual special real-time translation equipment, human special real-time translation equipment, a mobile terminal APP and a cloud service system. The target individual special real-time translation equipment is internally provided with an unalterable identity recognition ID hardware, and locally stores a voiceprint feature white list library only for a single target individual (including pets, babies, speechless and disabled double obstacle crowds and the like), which is used for collecting the voice of the individual, translating the voice into a human understandable language in a local offline mode and outputting; the human special equipment locally stores a voiceprint white list of a specified user, which is used for collecting the voice of the user, converting the voice into a signal perceivable by the target individual in a local offline mode and outputting; the APP communicates with the equipment and the cloud, and is used for configuring parameters, uploading samples and receiving model updates. The application is widely applied to the communication scene of pets, babies, speechless and disabled double obstacle crowds and other people.

Description

Technical Field

[0001] This invention relates to the fields of voiceprint recognition, offline speech processing, and animal interaction technology. Specifically, it relates to a two-way real-time translation system, device, and method based on a unique voiceprint and hardware identity, which is suitable for two-way real-time communication between pets and humans, infants and guardians, and people with dual disabilities (aphasia and disability) and caregivers. Background Technology

[0002] With the increasing pet ownership rate and the growing demand for care for special groups in an aging society, translation devices that enable effective communication between humans and non-verbal individuals (pets, infants, and people with both aphasia and disability) have become a hot research topic. Currently, related technologies can be mainly categorized as follows:

[0003] 1. Pet bark recognition device based on a general acoustic model

[0004] Existing pet translation devices typically employ cloud-based or local universal acoustic models to uniformly identify the cries of pets of different breeds and individuals. For example, Chinese invention patent CN121122241A discloses a baby cry detection and translation system based on sound feature recognition. This system collects multimodal data such as the baby's voiceprint, body temperature, pulse, and movements, and performs comprehensive analysis in conjunction with a cloud database to determine the cause of the baby's crying and translate it into text. However, this solution only performs one-way recognition of the baby's cry and does not establish a dedicated voiceprint whitelist for each individual. Any baby's cry can trigger recognition, posing a high risk of false triggers. Furthermore, this solution relies on a cloud database for real-time analysis; system functionality is limited when the network is interrupted, and uploading user data to the cloud poses a privacy risk.

[0005] 2. Professional domain machine translation system based on dynamic management of terminology database

[0006] Chinese invention patent CN121543608A discloses an intelligent translation method and system. Through a dynamic terminology management module, a RAG-enhanced semantic calibration system, and a translation quality evaluation module, it achieves high-fidelity translation of texts in specialized fields. This system employs a hierarchical terminology database and a locally pre-trained model, enabling partial translation offline. However, its translation targets are limited to conversions between human languages ​​(e.g., English to Chinese), and it does not address the recognition and translation of pet vocalizations, nor does it convert human language into signals perceptible to pets. Furthermore, the system still requires access to a cloud-based translation interface when processing complex sentences, failing to guarantee real-time response across all scenarios.

[0007] 3. Adaptive machine translation system based on human correction

[0008] Chinese invention patent CN1573741A discloses an adaptive machine translation method that continuously improves translation accuracy by sending the automatic translation results to a reliable source of correction (such as a human translator) for correction, and then using the corrected bilingual corpus to train a machine translation model. This system introduces the cloud-edge collaborative concept of "cloud training and local inference," but its training process relies on human correction, and the translation objects are still limited to human languages. More importantly, the system's real-time translation process still requires interaction with the cloud, such as uploading low-confidence segments and receiving updated parameters, making completely offline real-time translation impossible.

[0009] In summary, existing technologies share the following common drawbacks:

[0010] First, there is a lack of a dedicated voiceprint isolation mechanism for each individual: existing devices do not build a dedicated voiceprint whitelist for each pet or designated user. Any ambient sound (including the barking of other pets of the same breed) can trigger the device, resulting in a high false wake-up rate and low recognition accuracy. Although CN121122241A collects the voiceprint of an infant, it does not establish a dedicated whitelist for each individual, and cannot physically prevent interference from non-target sounds.

[0011] Secondly, true two-way translation is not achieved: existing solutions are mostly one-way recognition: from pet / baby vocalizations to human language, or limited to machine translation between human languages, lacking a complete reverse channel to convert human language into signals that the pet can perceive. Users cannot actively give commands to their pets or express emotions, and human-pet communication remains in a one-way receiving state.

[0012] Third, network dependence and privacy risks coexist: CN121122241A relies on a cloud database for real-time analysis, and its functionality is limited when the network is interrupted; CN121543608A needs to call the cloud translation interface when processing complex sentences; CN1573741A's real-time translation still requires interaction with the cloud. None of the above solutions can complete real-time translation independently in a completely offline state, and uploading the original audio data to the cloud poses a risk of privacy leakage.

[0013] Fourth, lack of scalable hardware identity identifiers: Existing devices are all isolated terminals without built-in, tamper-proof hardware-level identity recognition IDs, which makes it impossible to achieve identity authentication and data interaction between devices, limiting extended application scenarios such as multi-device collaboration and multi-pet management.

[0014] Fifth, the model update mechanism is not suitable for real-time translation scenarios: Although CN1573741A and CN121543608A involve model updates, neither of them has designed a low-power, incremental update mechanism for "real-time translation devices". The update process may interrupt the use of the device and does not support version rollback. Summary of the Invention

[0015] This invention designs a two-way real-time translation system, device, and method based on a unique voiceprint and hardware identity. The technical problems it solves are: the lack of individual unique voiceprint isolation in existing technologies leads to a high false trigger rate; reliance on real-time network computing poses privacy risks and offline unavailability; only supports one-way recognition and cannot achieve two-way communication; lack of hardware identity identification makes it difficult to expand; and the model update mechanism is not suitable for real-time devices. It cannot meet the needs of accurate, real-time, secure, and two-way communication in scenarios such as pets, infants, and people with dual disabilities such as aphasia and disability.

[0016] To solve the aforementioned technical problems, the present invention adopts the following solution:

[0017] A two-way real-time translation system based on a unique voiceprint and hardware identity includes: a real-time translation device dedicated to the target individual, an optional real-time translation device dedicated to humans, a mobile terminal APP, and a cloud service system. The real-time translation device dedicated to humans is an optional component: in pet applications, this device is used to perform reverse translation, converting human commands into sound wave signals that the pet can perceive, forming a complete two-way translation loop; in applications involving infants or individuals with both aphasia and disability, only the real-time translation device dedicated to the target individual needs to perform forward translation, and guardians, caregivers, or third parties can take corresponding actions (such as feeding, turning over, or comforting) based on the translation results, without the need for reverse translation.

[0018] The dedicated real-time translation device for the target individual has an built-in tamper-proof identity recognition ID hardware and locally stores a voiceprint feature whitelist library that is only for a single target individual; the dedicated real-time translation device for the target individual is used to collect the voice of the target individual, translate it offline locally into human-understandable language and output it.

[0019] The dedicated real-time translation device for humans has a local storage of a whitelist of voiceprint features of a specified user, which is used to collect the voice of the specified user, convert it into a signal that can be perceived by the target individual locally offline, and then output it.

[0020] The mobile terminal APP is connected to the target individual's dedicated real-time translation device, the human dedicated real-time translation device, and the cloud service system, respectively, and is used to configure device parameters, receive and upload sound samples and annotation data, and receive and send model update packages.

[0021] The cloud service system is used to receive and encrypt the uploaded data, train a personalized translation model based on a single individual's exclusive dataset, and push incremental update packages to the corresponding local device through the mobile terminal APP.

[0022] Preferably, both the target individual-specific real-time translation device and the human-specific real-time translation device include a low-power sleep / wake-up module with a standby power consumption of ≤1mA. Normally, only the sound pickup module and the voiceprint fast comparison submodule are running. Silence signals are eliminated through short-time energy analysis, and the entire module is only woken up when a valid sound source is detected and the voiceprint is successfully matched.

[0023] A two-way real-time translation device based on a unique voiceprint and hardware identity, wherein the device is a real-time translation device dedicated to a target individual or a real-time translation device dedicated to humans, and the same device supports only a single purpose;

[0024] When it is a real-time translation device specifically designed for a target individual, it includes:

[0025] Unalterable identity verification hardware;

[0026] A locally stored whitelist of voiceprint features specific to each individual;

[0027] The sound acquisition module is used to collect ambient sound.

[0028] The local processing chip has a built-in voiceprint comparison submodule, feature extraction submodule and local acoustic model, which is used to preprocess the collected target individual voice, extract the three-dimensional features of spectrum, duration and intensity, and map them into emotion-need combination tags through the local acoustic model;

[0029] The voice output module is used to convert the emotion-needs tags into human-understandable language and broadcast them.

[0030] The low-power sleep / wake-up module is used to control the device to remain in sleep mode when there is no valid trigger, and to wake up the entire module after successful voiceprint matching. After the translation is completed, a soft reset is performed to restore sleep mode.

[0031] When it is a real-time translation device for human use only, it includes:

[0032] A locally stored whitelist of user voiceprint features;

[0033] The sound acquisition module is used to collect ambient sound.

[0034] The local processing chip has a built-in voiceprint comparison submodule, semantic extraction submodule and signal conversion module, which are used to preprocess the collected speech, extract phonemes and keywords, summarize human intentions, and map them into a combination of frequency, duration and pitch signals that can be perceived by the target individual based on preset fixed instruction rules.

[0035] The signal output module is used to output signals that can be perceived by the target individual;

[0036] The low-power sleep / wake-up module has the same function as the target individual-specific real-time translation device.

[0037] Preferably, the voiceprint matching submodule uses a combination of primary and secondary methods and dual thresholds for voiceprint matching:

[0038] Using cosine similarity as the primary matching method, a core matching threshold of ≥85% is set.

[0039] Using Euclidean distance and Manhattan distance as auxiliary verification methods, respectively set acceptable threshold ranges;

[0040] Voiceprint matching is considered successful only when the cosine similarity is ≥85% and both the Euclidean distance and the Manhattan distance fall within the acceptable threshold range; otherwise, it is considered a mismatch and the voice data is discarded.

[0041] A bidirectional real-time translation method based on a unique voiceprint and hardware identity, applied to the aforementioned system or device, includes the following steps:

[0042] Step S1: Device initialization and voiceprint recording:

[0043] The device is set to be dedicated to a specific target individual or user, and then enters a low-power sleep / standby state.

[0044] Collect voice clips of a specified target individual or user, extract voiceprint feature vectors and store them locally, build a unique voiceprint whitelist, and set a similarity threshold of ≥85%;

[0045] Step S2: Real-time sound pickup and voiceprint matching:

[0046] The device normally operates in low-power sleep mode, continuously picking up audio and performing short-term energy analysis.

[0047] When a valid sound source is detected, the voiceprint features are extracted and matched with the whitelist;

[0048] A dual threshold verification method is adopted, with cosine similarity as the primary criterion and Euclidean distance and Manhattan distance as secondary criteria: if the cosine similarity is ≥85% and both the Euclidean distance and Manhattan distance fall within the qualified threshold range, the match is successful and the entire module is triggered to wake up; otherwise, the audio data is discarded and the system returns to low-power sleep mode.

[0049] Step S3: Branch processing based on device type:

[0050] If it is a real-time translation device specifically for a target individual, forward translation is performed, including:

[0051] Collect all original audio data of the target individual;

[0052] Preprocessing includes pre-emphasis, frame-segmentation windowing, and spectral subtraction for noise reduction.

[0053] Extract quantized feature vectors of three-dimensional features: spectrum, duration, and intensity;

[0054] Input the local acoustic model and output emotion-demand combination labels according to the principle of maximum probability.

[0055] The tags are converted into human-understandable language and broadcast via TTS voice within 5 seconds after the speech ends, with text display also supported.

[0056] The forward translation is applicable to all types of target individuals: pets, infants, and people with dual disabilities who are both aphasic and disabled. It is used to translate the voice of the target individual into human-understandable language so that a third party can understand it and take appropriate action.

[0057] If the device is for human use only and is applied to a pet scenario, a reverse conversion is performed, including:

[0058] Collect all raw voice data of users;

[0059] Perform the same preprocessing;

[0060] Extract phonemes and keywords, remove redundant words, and complete effective instructions;

[0061] Human language can be categorized as human intention, including emotions, needs, or instructions.

[0062] Based on preset fixed instruction rules, the target individual is mapped to a corresponding perceptible signal according to its type, and the signal is output within 5 seconds after the sound is emitted.

[0063] The reverse conversion is only performed in pet scenarios, converting human commands into sound signals that the pet can perceive. For infants and individuals with both aphasia and disability, no dedicated human-specific equipment or reverse conversion steps are required. It should be noted that the reverse conversion step is only performed for pet applications. For infants and individuals with both aphasia and disability, since their communication needs can be met directly through actions taken by a third party, there is no need to convert human language into signals perceptible to the target individual. Therefore, in these application scenarios, no dedicated human-specific equipment or reverse conversion steps are required.

[0064] Step S4: Device soft reset

[0065] After the translation or conversion output is completed, the non-core modules are shut down, leaving only the sound pickup module and the voiceprint fast comparison submodule. The system then automatically returns to a low-power sleep state and repeats step S2.

[0066] Preferably, in the forward translation, the emotion type output by the local acoustic model is preset according to the target individual type, and the demand type is also preset; if the maximum probability output by the model is lower than the preset threshold, it is determined as unrecognized and a prompt message is output.

[0067] Preferably, in the reverse conversion, the example of mapping with preset fixed instruction rules includes: for pets, mapping to an ultrasonic frequency band signal that is compatible with the pet's hearing sensitive range.

[0068] Preferably, it also includes an iterative step for the cloud-edge collaborative model:

[0069] The local device associates the sound sample, translation / conversion result, and identity ID, and then synchronizes them to the mobile terminal APP via Bluetooth.

[0070] Users manually annotate samples in the app, including emotions, needs, or semantics;

[0071] The app will encrypt and upload valid samples, labeled data, and identity IDs to the cloud service system;

[0072] Personalized translation models are trained in the cloud based on a single individual's exclusive dataset, and a new version is released when the recognition accuracy is improved by ≥5%.

[0073] Incremental update packages (≤10MB) are generated in the cloud and pushed to the corresponding apps based on the user's identity ID;

[0074] The app sends the update package to the local device via Bluetooth. The local device completes the incremental update within 10 seconds and supports version rollback.

[0075] Preferably, in the low-power sleep standby state, the standby power consumption is ≤1mA; the voiceprint matching time is ≤500ms; and the total time for a single round of forward translation or reverse conversion is ≤5 seconds.

[0076] Preferably, after a soft reset, the device returns to a low-power sleep state and continues real-time audio monitoring, forming a continuous loop.

[0077] A two-way real-time translation device based on a unique voiceprint and hardware identity, comprising:

[0078] Unalterable identity verification hardware;

[0079] The locally stored voiceprint feature whitelist database includes a whitelist of target individual voiceprints and a whitelist of human user voiceprints.

[0080] The sound acquisition module is used to collect ambient sound.

[0081] The sound source category recognition module is used to classify the collected valid sound sources into biological categories and output category labels and confidence scores.

[0082] The voiceprint comparison module is used to select the corresponding voiceprint whitelist for matching based on the category label;

[0083] The local processing chip has a built-in forward translation unit and a reverse conversion unit;

[0084] The output module is used to output translation results or conversion signals;

[0085] The low-power sleep / wake-up module is used to control the device to remain in sleep mode when there is no effective trigger, and to wake up the entire module after the sound source category is matched and the voiceprint is successfully matched;

[0086] The device is configured as follows:

[0087] When the category label output by the sound source category identification module is the target individual category and the confidence level is not lower than the preset threshold, the voiceprint comparison module selects the target individual voiceprint whitelist for matching. After successful matching, the forward translation unit translates the collected sound into human-understandable language and outputs it.

[0088] When the category label output by the sound source category recognition module is a human user category and the confidence level is not lower than a preset threshold, the voiceprint comparison module selects the human user voiceprint whitelist for matching. After successful matching, the reverse conversion unit converts the collected speech into a signal that can be perceived by the target individual and outputs it.

[0089] When the category label does not match the target individual category preset by the device, the audio data is discarded and translation is not triggered.

[0090] The bidirectional real-time translation system, device, and method based on unique voiceprints and hardware identity have the following beneficial effects:

[0091] (1) This invention constructs a unique voiceprint whitelist for each individual and employs a voiceprint matching method combining primary and secondary methods with dual thresholds (cosine similarity as the primary matching, and Euclidean distance and Manhattan distance as secondary verification). It only responds to the voices of pre-recorded single pets or designated users. Unbound voices (including other pets of the same breed, environmental noise, and non-designated users) are directly discarded, without waking the device, storing, or responding. The measured false wake-up rate is ≤0.01%, which eliminates the problem of false triggering from a physical perspective and is significantly better than the ≥15% false triggering rate of general voiceprint recognition or non-isolation schemes in the prior art.

[0092] (2) In this invention, all real-time translation and conversion operations are completed on the local chip without relying on any network communication. After the device is woken up, the entire process, including voiceprint comparison, feature extraction, model inference, and signal output, is executed locally, and the results are output quickly after the sound is emitted. Compared with the prior art, this invention does not require real-time cloud computing, and the device function is completely unaffected when the network is interrupted. It can be applied to any scenario, such as home, outdoors, or travel. At the same time, the original sound data is not uploaded to the cloud, thus avoiding the risk of privacy leakage from the source.

[0093] (3) This invention provides differentiated translation modes according to application scenarios: For pets, it achieves complete two-way translation—forward translation maps pet barks to emotional and demand combination tags and converts them into human language broadcasts, while reverse conversion maps human commands to ultrasonic frequency band sound wave signals adapted to the pet's auditory sensitive range, realizing two-way communication between humans and pets; for infants and people with dual disabilities such as aphasia and disability, it adopts a one-way translation mode—only the target individual's voice needs to be translated into human-understandable language, and the guardian, caregiver, or third party can directly take corresponding actions based on the translation results without the need for reverse conversion. This differentiated design not only meets the needs of two-way communication in pet scenarios, but also avoids adding unnecessary equipment complexity and signal interference in infant and patient scenarios, reflecting the flexibility and pertinence of the technical solution.

[0094] (4) The present invention incorporates an immutable identity recognition ID hardware into a dedicated real-time translation device for the target individual. This ID is bound one-to-one with a single individual's voiceprint whitelist, and its consistency is automatically verified when the device is powered on. Compared with existing technologies that rely solely on software or cloud-based identification, the hardware ID of the present invention is unique and immutable, providing a physical-level security foundation for identity authentication, data interaction, and collaborative work among multiple pet devices in the future, and significantly improving the scalability of the system.

[0095] (5) This invention adopts a cloud-edge collaborative architecture that combines real-time edge processing with iterative optimization in the cloud. The local device runs real-time translation completely offline, while the cloud is only used for non-real-time data storage and model training. Samples are collected through the APP and manually labeled. The cloud triggers personalized model training based on a single individual's exclusive dataset (≥500 samples). When the recognition accuracy improves by ≥5%, a new version is released, generating an incremental update package (≤10MB) and accurately pushing it to the corresponding device based on the identity ID. The local device completes the incremental update within ≤10 seconds and supports version rollback. This mechanism allows the translation accuracy to continuously improve over time (initially ≥85%, and can reach ≥95% after multiple iterations), solving the problem of model solidification and inability to improve accuracy in existing technologies.

[0096] (6) This invention has wide applicability. By changing the voiceprint whitelist and signal mapping rules, it can be seamlessly applied to pets, infants and people who are unable to express their needs in a conventional way due to aphasia and disability. It provides an effective two-way communication solution for the above groups and has significant social benefits. Attached Figure Description

[0097] Figure 1 This invention is based on the overall architecture diagram of a two-way real-time translation system using a unique voiceprint and hardware identity.

[0098] Figure 2 : A schematic diagram of the module connections of the real-time translation device specifically for the target individual in this invention;

[0099] Figure 3 : A schematic diagram of the module connections of the human-dedicated real-time translation device in this invention;

[0100] Figure 4 : A schematic diagram of the module connection of the mobile terminal APP in this invention;

[0101] Figure 5 : A schematic diagram of the module connections of the cloud service system in this invention;

[0102] Figure 6 This invention is based on a flowchart of a two-way translation process using a unique voiceprint and hardware identity.

[0103] Figure 7 This invention is based on an iterative flowchart of a cloud-edge collaborative model using exclusive voiceprints and hardware identities. Detailed Implementation

[0104] The following is combined with Figures 1 to 7 The present invention will be further described as follows:

[0105] like Figure 1 As shown, the bidirectional real-time translation system based on a unique voiceprint and hardware identity provided by this invention includes: a real-time translation device, a mobile terminal APP, and a cloud service system. The real-time translation device includes a real-time translation device specifically for the target individual and a real-time translation device specifically for humans. The four devices work together to achieve bidirectional real-time translation between the target individual and humans.

[0106] Both the target-specific real-time translation device and the human-specific real-time translation device establish a two-way Bluetooth connection with the mobile terminal APP through their internal communication modules:

[0107] The device sends the following data to the app: sound samples, translation / conversion results, identity ID, and device operation logs; the app sends the following data to the device: model incremental update packages and configuration parameters, such as voiceprint matching threshold and output volume.

[0108] Mobile apps establish two-way communication with cloud service systems via the internet:

[0109] The app sends the following data to the cloud: encrypted valid sound samples, manually labeled data, and identity ID association information; the cloud sends the following data to the app: incremental model update packages, version notifications, and training status feedback.

[0110] like Figure 2As shown, the target individual dedicated real-time translation device 1 internally includes: identity recognition ID hardware 11, voiceprint feature whitelist library 12, sound acquisition module 13, local processing chip 14, voice output module 15, communication module 16, and low-power sleep / wake-up module 17. Among them, the local processing chip 14 further includes a voiceprint comparison module, a feature extraction module, and a local acoustic model.

[0111] The data transfer path between modules is as follows:

[0112] The identity recognition ID hardware 11 uses an internal bus, such as I 2 C or SPI, bound to the voiceprint feature whitelist database 12. When the device is powered on, the main control chip reads the ID and verifies its consistency with the ID in the whitelist database.

[0113] The sound acquisition module 13 acquires ambient sound in real time at a sampling rate of 16kHz and transmits the raw audio data to the low-power sleep / wake-up module 17.

[0114] The low-power sleep / wake-up module 17 first performs a short-term energy analysis on the audio:

[0115] If the energy is below the preset threshold, it is determined to be silent, the data is discarded, and other modules are not woken up; if a valid sound source is detected, the audio data is transmitted to the voiceprint comparison module.

[0116] The voiceprint comparison module extracts the voiceprint features of the audio and reads the pre-stored single-individual voiceprint feature vector from the voiceprint feature whitelist library 12, performing main and auxiliary combination and dual threshold matching:

[0117] Main match: cosine similarity (threshold ≥ 85%);

[0118] Auxiliary verification: Euclidean distance (threshold ≤ 0.25) and Manhattan distance (threshold ≤ 0.35).

[0119] The matching result (success / failure) is sent back to the low-power sleep / wake-up module 17.

[0120] If the matching fails, the low-power sleep / wake-up module 17 discards the audio, and the device remains in a low-power sleep state without further processing. If the matching succeeds, the low-power sleep / wake-up module 17 sends a wake-up signal to all other modules in the device (local processing chip, voice output module, communication module, etc.) to enable all modules to work.

[0121] Upon wake-up, the raw audio data is transmitted from the sound acquisition module 13 or the low-power sleep / wake-up module 17 to the feature extraction module. The feature extraction module then performs the following steps:

[0122] Pre-emphasis (coefficient 0.97); framing (20-30ms / frame); windowing (rectangular window or Hamming window); spectral subtraction denoising; extraction of quantized feature vectors of three-dimensional features of spectrum, duration, and intensity. The feature vectors are then transmitted to the local acoustic model 143.

[0123] The local acoustic model 143 is a pre-built lightweight neural network model (such as CNN and LSTM) that outputs a unique emotion-need combination label according to the principle of maximum probability, such as "anxiety and eating". This label is then sent to the speech output module 15.

[0124] The voice output module 15 converts the tags into standardized human-understandable sentences and broadcasts them through the local TTS voice library, such as "The cat is very anxious now and wants to eat," while also generating text information.

[0125] The translation results (text), along with the original audio sample, identity ID, and other information, are transmitted to the communication module 16 via the internal bus. The communication module packages the data and sends it to the mobile terminal APP 3 via Bluetooth.

[0126] After each translation is completed, the low-power sleep wake-up module 17 performs a soft reset: shuts down all modules except itself and the voiceprint comparison module, and the device returns to a low-power sleep state (standby power consumption ≤1mA).

[0127] like Figure 3 As shown, the human-dedicated real-time translation device 2 internally includes: a voiceprint feature whitelist library 21, a sound acquisition module 22, a local processing chip 23, a signal output module 24, a communication module 25, and a low-power sleep / wake-up module 26. The local processing chip 23 further includes a voiceprint comparison module, a semantic extraction submodule, and a signal conversion module.

[0128] The data transfer path between modules is as follows:

[0129] The sound acquisition module 22 acquires the user's voice (sampling rate 16kHz) and transmits the audio data to the low-power sleep-wake module 26.

[0130] The low-power sleep / wake-up module 26 performs short-term energy analysis and voiceprint matching: audio data is sent to the voiceprint comparison module, which reads the pre-stored user voiceprint features from the voiceprint feature whitelist library 21 and performs matching (the matching method is the same as that of the target individual's dedicated real-time translation device, using primary and secondary dual thresholds). The matching result is sent back to the sleep / wake-up module.

[0131] If a match fails, the audio is discarded and the device remains in sleep mode; if a match succeeds, all modules are awakened.

[0132] Upon wake-up, the audio data is transmitted to the semantic extraction module. This submodule performs the following:

[0133] Preprocessing (pre-emphasis, frame-by-frame windowing, noise reduction); phoneme and keyword extraction; redundant word removal (such as "um" and "that"); valid instruction completion. If no valid keywords are found, an "unrecognized" message is output, and the conversion ends.

[0134] The extracted semantic features (keywords and summarized intent) are transmitted to the signal conversion module. This module has a built-in preset fixed instruction rule library that maps human intent into signal parameters (frequency, duration, pitch) that are perceptible to the target individual. For example:

[0135] For pets:

[0136] "Eating" - a 10kHz sine wave lasting 0.5 seconds;

[0137] "Come here" - 8kHz pulse wave, repeated 3 times;

[0138] "Praise" - Frequency sweep from low to high (2kHz→12kHz, lasting 1 second).

[0139] The converted digital signal parameters are transmitted to the signal output module 24. For pets and infants, the signal output module generates corresponding sound wave signals through a high-fidelity speaker; for people with both aphasia and disability, the signal output module outputs tactile vibration signals through bone conduction headphones or a vibration motor.

[0140] The conversion result (text description), along with the original voice sample, identity ID, and other information, is transmitted to the communication module 25 via the internal bus and synchronized to the mobile terminal APP3 via Bluetooth.

[0141] Upon completion, the low-power sleep / wake-up module 26 performs a soft reset, shuts down non-core modules, and resumes sleep mode.

[0142] The human-specific real-time translation device 2 and its reverse conversion function of the present invention are mainly designed for pet application scenarios. In pet scenarios, the target individual-specific real-time translation device 1 translates the pet's barks into human language, and then the human-specific real-time translation device 2 converts human commands into ultrasonic frequency sound wave signals that the pet can perceive, forming a complete two-way translation closed loop and realizing two-way communication between humans and pets.

[0143] For applications involving infants and individuals with both aphasia and disability, the communication needs are unidirectional: caregivers or guardians only need to understand the needs expressed by the infant's cries or the patient's vocalizations (such as hunger, pain, or needing water) to take direct action (such as feeding, turning over, or giving water), without needing to convert human language into signals perceptible to the infant or patient. Therefore, in these application scenarios, only a dedicated real-time translation device 1 for the target individual is required for forward translation; a dedicated real-time translation device 2 for humans is not needed, nor is a reverse conversion step required.

[0144] like Figure 4 As shown, the mobile terminal APP 3 includes: a hardware configuration module 31, a data acquisition and upload module 32, a manual annotation module 33, a model update module 34, a data visualization module 35, and a system settings module 36. The data flow between the modules is as follows:

[0145] Hardware configuration module 31 communicates with the target individual-specific real-time translation device 1 and the human-specific real-time translation device 2 via Bluetooth:

[0146] Read and display the device ID; send configuration parameters such as voiceprint matching threshold, sleep power consumption, output volume, and signal parameters; receive voiceprint samples uploaded by the device and complete voiceprint recording and binding.

[0147] The data acquisition and upload module 32 receives samples, translation / conversion results, and ID information from the target individual-dedicated real-time translation device 1 and the human-dedicated real-time translation device 2, and stores them locally. After user confirmation, valid samples are filtered, encrypted, and uploaded to the cloud service system 4. Resume interrupted uploads are supported.

[0148] The manual annotation module 33 reads unannotated samples from the local database and provides a visual interface for users to annotate: Pet samples: annotate the emotion type (excitement, comfort, anxiety, fear, alertness) and the need type (eating, drinking, resting, going out, playing, comforting); Human samples: annotate the core semantics and keywords.

[0149] The labeled data is transmitted to the data acquisition and upload module 32 and uploaded to the cloud along with the original samples.

[0150] The model update module 34 receives incremental update packages pushed from the cloud and distributes them via Bluetooth to the corresponding target individual-specific real-time translation device 1 or human-specific real-time translation device 2. It also maintains a local model version record, supporting manual triggering of updates or automatic update checks.

[0151] The data visualization module 35 receives translation / conversion results in real time from the target individual's dedicated real-time translation device 1 and the human-dedicated real-time translation device 2, displays them in text and voice formats, and stores historical logs. Users can query and export data by ID and time.

[0152] The system settings module 36 manages account binding between the APP and the cloud, data privacy permissions, ID security management, etc., and the configuration parameters can be synchronized to the hardware configuration module.

[0153] like Figure 5 As shown, the cloud service system 4 includes: a data encryption storage module 41, a single individual personalized model training module 42, a model generalization optimization module 43, a model version management and distribution module 44, and a cloud monitoring module 45.

[0154] The data flow is as follows:

[0155] The data encryption and storage module 41 receives uploaded data (samples, labels, IDs) from the mobile terminal APP3, performs classified encryption and storage, and establishes a mapping between the identity ID and the exclusive dataset of a single individual.

[0156] The individual personalized model training module 42 periodically scans the storage module. When the number of labeled samples corresponding to a certain ID reaches a preset threshold (e.g., 500 samples), it automatically reads the dataset for that ID, performs mini-batch gradient descent training, and fine-tunes the local acoustic model parameters. The new model parameters after training are sent to the model version management and distribution module 44.

[0157] The model generalization optimization module 43 uses anonymized multi-pet data (excluding personal privacy information) to train the base model and optimize the human speech conversion rules. The optimized base model parameters are also sent to the model version management and distribution module 44.

[0158] The model version management and distribution module 44 performs accuracy verification on new models (using a test set, releasing the model when the recognition accuracy is ≥5% higher than the previous version), assigns a version number, and pushes incremental update packages (≤10MB) to the corresponding mobile terminal APP3 based on the identity ID.

[0159] The cloud monitoring module 45 collects the operating status, training progress, and device connection status of each module in real time, and automatically issues warnings for abnormal tasks.

[0160] Figure 2 The target individual dedicated real-time translation device 1 as a whole Figure 1 One of the independent nodes, whose internal communication module 16 is responsible for communicating with... Figure 4 Data exchange is performed on the mobile terminal APP3.

[0161] Similarly, Figure 3 Human-dedicated real-time translation device 2 as Figure 1 Another node in the process interacts with the mobile terminal APP3 through the communication module 25.

[0162] Figure 4 The mobile terminal APP 3 serves as the hub connecting the target individual-specific real-time translation device 1 and the human-specific real-time translation device 2 with the cloud service system 4. Its hardware configuration module 31, data acquisition module 32, and model update module 34 communicate via Bluetooth with the communication modules 16 and 25 of the target individual-specific real-time translation device 1 and the human-specific real-time translation device 2, respectively; its data acquisition and upload module 32 and model update module 34 also communicate with… Figure 5 The data encryption storage module 41 and the model version management and distribution module 44 communicate with each other via the network.

[0163] Figure 5 The cloud service system 4 receives uploaded data from the mobile terminal APP 3 through the data encryption storage module 41, and pushes update packages to the mobile terminal APP 3 through the model version management and distribution module 44.

[0164] Figure 6 In this invention, the bidirectional real-time translation process starts from the top and proceeds downwards. The specific steps are as follows:

[0165] S101: Low-power sleep state;

[0166] Only the sound acquisition module 13 and its voiceprint fast comparison module in the target individual dedicated real-time translation device 1 and the sound acquisition module 22 and its voiceprint fast comparison module in the human dedicated real-time translation device 2 operate, with standby power consumption ≤1mA.

[0167] S102: Sound pickup and short-time energy analysis;

[0168] Sound acquisition modules 13 and 22 acquire ambient sound and calculate short-term energy; if the energy is below the threshold, it is discarded; otherwise, proceed to S103.

[0169] S103: Voiceprint matching;

[0170] Combining primary and secondary methods with dual thresholds—extracting voiceprint features and comparing them with a whitelist database:

[0171] Primary matching: Cosine similarity ≥ 85%? Secondary validation: Euclidean distance ≤ 0.25 and Manhattan distance ≤ 0.35? Matching is successful if all three conditions are met; otherwise, it fails.

[0172] S104: Discard unbound audio;

[0173] If the match fails, discard the sound and return to S101.

[0174] S105: Full module wake-up;

[0175] If a match is successful, the local processing chip, output module, communication module, etc., will be activated.

[0176] Branch decision: Proceed to S106 or S107 depending on the device type (target individual dedicated real-time translation device 1 and human dedicated real-time translation device 2).

[0177] S106: Forward Translation Process (Target Individual / Pet → Human)

[0178] Collect all raw vocalizations of pets;

[0179] Preprocessing (pre-emphasis, frame-by-frame windowing, spectral subtraction denoising);

[0180] Extract the three-dimensional feature vectors of spectrum, duration, and intensity;

[0181] The local acoustic model infers and outputs a combined emotion-needs label.

[0182] It is translated into human-understandable language and broadcast through a local TTS speech library, while also supporting text display.

[0183] S107: Reverse Conversion Process (Human → Target Individual / Pet)

[0184] Collect all raw human speech data;

[0185] Preprocessing (same as above);

[0186] Extract phoneme and keyword features, remove redundant words, and complete effective instructions;

[0187] Human language can be categorized into the target individual's emotions, needs, and instructions.

[0188] Based on preset fixed instruction rules, the system maps the target individual's type (pet / infant / aphasic / disabled person) to a corresponding perceptible signal (ultrasound frequency sound waves for pets, soothing sound waves for infants, and vibration signals for aphasic / disabled people), and outputs the signal within 5 seconds after the vocalization ends. It should be noted that the reverse conversion is only performed for pet applications. For infants and individuals with both aphasia and disability, the reverse conversion step is unnecessary.

[0189] S108: Soft reset, return to low-power sleep mode: After the output is completed, non-core modules (model inference, output, communication, etc.) are shut down, only the sound pickup and voiceprint comparison submodule is retained, and the system returns to S101.

[0190] Figure 6 In the process, S101→S102→S103→(if it fails)S104→S101; if it succeeds, then S105→branch→S106 or S107→S108→S101. The entire single-round processing takes ≤5 seconds, of which voiceprint matching takes ≤500ms.

[0191] like Figure 7The process demonstrates a complete closed loop, starting with real-time translation / conversion on a local device, through sample collection, manual annotation, cloud training, model push, and finally achieving incremental updates of the local device model. Figure 7 It consists of two main parts: edge (local devices and apps) and cloud.

[0192] The specific steps are as follows:

[0193] T101: Real-time translation / conversion on local devices;

[0194] Real-time translation devices specifically designed for individuals or for humans are used in the execution of... Figure 2 During the process shown, sound samples and translation / conversion results are automatically collected, associated with the identity ID, and stored in the local cache.

[0195] T102: Samples and results are synchronized to the APP;

[0196] When the device connects to the mobile app via Bluetooth, the cached samples, results, and ID information are automatically synchronized to the app. Synchronization uses a lightweight compression format.

[0197] T103: Manual screening and labeling;

[0198] Users can filter and label samples through the annotation module of the app:

[0199] Pet vocalization samples: labeled with emotion type (excitement, comfort, anxiety, fear, alertness) and need type (eating, drinking, resting, going out, playing, comforting);

[0200] Human speech samples: labeled with core semantics and keywords.

[0201] T104: Sample and labeled data are encrypted and uploaded to the cloud;

[0202] After user confirmation, the APP will encrypt and upload the valid original sample, manually labeled data, and identity ID association information to the cloud service system (supporting breakpoint resume).

[0203] T105: Cloud-based encrypted data storage and ID association;

[0204] Uploaded data is categorized and encrypted for storage in the cloud, establishing a mapping between identity IDs and individual-specific datasets.

[0205] T106: Personalized model training for a single individual;

[0206] When the accumulated labeled dataset for a specific target individual (such as a pet) reaches a preset number (e.g., 500 records), the cloud automatically triggers the individual-specific model training module, which uses mini-batch gradient descent to fine-tune the model parameters. Simultaneously, the model generalization optimization module utilizes anonymized multi-pet data to optimize the base model.

[0207] T107: Model accuracy verification and version release;

[0208] After training, the new model's accuracy is verified using a test set in the cloud. If the recognition accuracy is improved by ≥5% compared to the previous version, it is released as a new version and assigned a version number.

[0209] T108: Push incremental update packages based on ID;

[0210] The cloud-based model version management and distribution module pushes incremental update packages (≤10MB in size) to the user's app for the corresponding target individual (such as a pet) based on the identity ID.

[0211] T109: The APP receives the update package and sends it to the local device;

[0212] After receiving the update package, the app prompts the user to update. Once the user confirms, the app transmits the incremental update package to the local device via Bluetooth.

[0213] T110: Local device incremental update, completing the iteration loop;

[0214] After receiving the update package, the local device enters a brief "update mode" (≤10 seconds) to merge incremental parameters into the local acoustic model. Upon completion, a soft reset is automatically performed, restoring the device to a low-power sleep state. The previous version model is retained during the update process, and one-click rollback is supported.

[0215] like Figure 7 In the middle, T101→T102→T103→T104→T105→T106→T107→T108→

[0216] T109→T110→(Return to T101, next iteration).

[0217] The following are examples of application extensions:

[0218] Example 1: Applied to translating baby cries;

[0219] The individual device is worn on the infant's wrist or placed next to the crib to record the infant's voiceprint. The device only responds to the infant's cries. When a cry is detected, the spectral density, duration, and intensity features are extracted, and the local model identifies it as a "hunger" demand, broadcasting "Baby is hungry, please feed him / her" within 5 seconds. The entire process is completed offline, unaffected by other infant cries or environmental noise. After hearing the broadcast, the caregiver can directly take action to feed the infant without needing to send a signal to the infant through a reverse conversion device.

[0220] Example 2: Application in assisting communication for patients with dual disabilities of aphasia and disability;

[0221] Individualized devices, such as lanyard-style devices, are worn by patients with both aphasia and disability. These devices record the patient's unique vocalizations, including indistinct syllables. When the patient makes a specific sound (such as "ah"), the device recognizes and translates it as "I want water," broadcasting it through a speaker or sending it to a caregiver's mobile phone. Upon hearing the broadcast or receiving the notification, caregivers can directly provide water to the patient without needing to send a signal back to the patient via a reverse conversion device.

[0222] Example 3: Automatic switching between one-way and two-way translation

[0223] This invention does not mechanically apply reverse conversion to all scenarios, but rather makes differentiated configurations based on the different nature of communication: for pets that cannot understand human language, reverse conversion is required to make sound wave signals that they can perceive; for infants / disabled people who cannot express complex semantics, but have the ability to understand human language or simple instructions, guardians can take action directly, and reverse conversion is neither necessary nor may cause signal interference.

[0224] Based on the existence of two modes for different objects, the present invention can also add a sound source category recognition module, which can automatically identify the sound source type (whether it is a pet / infant / disabled person or normal human user) and automatically switch the translation direction.

[0225] The sound source category recognition module is executed before voiceprint matching, or in parallel with voiceprint matching and based on the classification results. It performs biological category classification on valid sound sources, determines whether the speaker belongs to the "target individual category" (pet / infant / disabled person) or the "human user category", and thus decides whether to perform forward translation or reverse conversion.

[0226] The sound source category identification module may include an acoustic feature extraction submodule, a lightweight classifier, and a category whitelist. The acoustic feature extraction submodule extracts category-identifying features from the acquired sound signals, including but not limited to: fundamental frequency (F0) range, formant distribution, harmonic noise ratio (HNR), Mel-frequency cepstral coefficients (MFCC), duration distribution, and energy envelope. The lightweight classifier uses a pre-trained lightweight neural network (such as MobileNetV2 or TinyML architecture) or a traditional machine learning model (such as SVM or random forest) for classification, outputting category labels and confidence scores. The category whitelist is a pre-stored identifier of the "target individual category" bound to the device (such as "pet," "infant," or "disabled person"), used for matching and verification with the classification results.

[0227] Three scenarios are handled: If the output of the lightweight classifier is "target individual category" (such as pet barking) and the confidence level is higher than the threshold (such as ≥80%), then the forward translation process is initiated.

[0228] If the lightweight classifier outputs "human user" and the confidence level is higher than the threshold, then proceed to the reverse conversion process.

[0229] If the classification result does not match the device's preset target individual category (e.g., the device is bound to a pet, but a baby's cry is detected), the sound is discarded and translation is not triggered.

[0230] Specifically, a bidirectional real-time translation device based on a unique voiceprint and hardware identity includes: an immutable identity recognition ID hardware; a locally stored voiceprint feature whitelist library, including a whitelist of target individual voiceprints and a whitelist of human user voiceprints; a sound acquisition module for acquiring ambient sound; a sound source category identification module for classifying the acquired valid sound sources into biological categories and outputting category labels and confidence levels; a voiceprint comparison module for selecting the corresponding voiceprint whitelist for matching based on the category labels; a local processing chip with a built-in forward translation unit and a reverse conversion unit; an output module for outputting translation results or conversion signals; and a low-power sleep / wake-up module for controlling the device to remain in sleep mode when there is no valid trigger, and to wake up when the sound source category matches and the voiceprint matches. The device is configured as follows: when the category label output by the sound source category recognition module matches the target individual category and the confidence level is not lower than a preset threshold, the voiceprint comparison module selects the target individual's voiceprint whitelist for matching. After successful matching, the forward translation unit translates the collected sound into human-understandable language and outputs it. When the category label output by the sound source category recognition module matches the human user category and the confidence level is not lower than a preset threshold, the voiceprint comparison module selects the human user's voiceprint whitelist for matching. After successful matching, the reverse conversion unit converts the collected speech into a signal perceptible to the target individual and outputs it. When the category label does not match the target individual category preset by the device, the sound data is discarded and translation is not triggered.

[0231] The working principle of this two-way real-time translation device based on a unique voiceprint and hardware identity is as follows:

[0232] S101: Low-power sleep mode (only the pickup module operates);

[0233] S102: Short-time energy analysis to detect valid sound sources;

[0234] S103-A: Sound Source Category Identification:

[0235] Classify valid sound sources and output category labels and confidence levels; such as the specific range of the confidence threshold (e.g., ≥80%).

[0236] S103-B: Category Matching Validation:

[0237] Compare the classification results with the "target individual category" associated with the device:

[0238] If categorized as "target individual category" (e.g., pet) → perform voiceprint matching (positive whitelist);

[0239] If categorized as "human user" → perform voiceprint matching (reverse whitelist, i.e., specify the user's voiceprint).

[0240] If classification fails or the category does not match, discard the sound and return to sleep mode.

[0241] S104: Voiceprint matching (using the original dual threshold scheme, but the whitelist for matching is selected according to the category);

[0242] S105: If a match is successful, all modules will be activated;

[0243] S106 (forward translation) or S107 (reverse conversion) is automatically selected to be executed based on the classification results of step S103-A;

[0244] S108: Soft reset, return to hibernation.

[0245] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the implementation of the present invention is not limited to the above-described manner. Any improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A two-way real-time translation system based on a unique voiceprint and hardware identity, characterized in that, include: The system includes a target individual-specific real-time translation device (1), an optional human-specific real-time translation device (2), a mobile terminal APP (3), and a cloud service system (4); the human-specific real-time translation device (2) is an optional component, used to achieve reverse conversion when applied to pet scenarios, and may not be configured for reverse conversion when applied to infants or people with dual disabilities who are aphasic and disabled. The target individual dedicated real-time translation device (1) has an untamperable identity recognition ID hardware (11) built in and a voiceprint feature whitelist library (12) that is only for a single target individual stored locally; the target individual dedicated real-time translation device (1) is used to collect the voice of the target individual, translate it into human-understandable language offline locally and output it. The human-dedicated real-time translation device (2) locally stores a whitelist of voiceprint features of a specified user (21), which is used to collect the voice of the specified user, convert it into a signal that can be perceived by the target individual locally offline, and output it. The mobile terminal APP (3) is connected to the target individual dedicated real-time translation device (1), the human dedicated real-time translation device (2) and the cloud service system (4) respectively, and is used to configure device parameters, receive and upload sound samples and annotation data, and receive and send model update packages; The cloud service system (4) is used to receive and encrypt the uploaded data, train a personalized translation model based on a single individual's exclusive dataset, and push incremental update packages to the corresponding local device through the mobile terminal APP (3).

2. The bidirectional real-time translation system based on exclusive voiceprint and hardware identity as described in claim 1, characterized in that: Both the target individual-specific real-time translation device (1) and the human-specific real-time translation device (2) include a low-power sleep-wake module with a standby power consumption of ≤1mA. Normally, only the sound pickup module and the voiceprint fast comparison sub-module are running. The silent signal is eliminated through short-time energy analysis. The entire module is only woken up when a valid sound source is detected and the voiceprint is successfully matched.

3. A two-way real-time translation device based on a unique voiceprint and hardware identity, characterized in that: The device is a real-time translation device (1) for a target individual or a real-time translation device (2) for a human, and the same device supports only a single purpose; When it is a real-time translation device (1) specifically for a target individual, it includes: Unalterable identity recognition ID hardware (11); Locally stored whitelist of individual voiceprint features (12); The sound acquisition module (13) is used to acquire ambient sound. The local processing chip (14) has a built-in voiceprint matching submodule (141), feature extraction submodule (142) and local acoustic model (143) for preprocessing the collected target individual voice, extracting the three-dimensional features of spectrum, duration and intensity, and mapping them into emotion-need combination tags through the local acoustic model; The voice output module (15) is used to convert the emotion-needs tags into human-understandable language and broadcast them. The low-power sleep-wake module (17) is used to control the device to remain in sleep when there is no effective trigger, and to wake up the entire module after successful voiceprint matching, and to perform a soft reset to restore sleep after translation is completed; When it is a real-time translation device for human use (2), it includes: Locally stored user voiceprint feature whitelist (21); The sound acquisition module (22) is used to acquire ambient sound; The local processing chip (23) has a built-in voiceprint comparison submodule (231), semantic extraction submodule (232) and signal conversion module (233) for preprocessing the collected speech, extracting phonemes and keywords, summarizing them into human intentions, and mapping them into a combination of frequency, duration and pitch signals that can be perceived by the target individual based on preset fixed instruction rules. The signal output module (24) is used to output a signal that can be perceived by the target individual; The low-power sleep-wake module (26) has the same function as the target individual dedicated real-time translation device.

4. The bidirectional real-time translation device based on a unique voiceprint and hardware identity as described in claim 3, characterized in that: The voiceprint matching submodule uses a combination of primary and secondary methods and dual thresholds for voiceprint matching. Using cosine similarity as the primary matching method, a core matching threshold of ≥85% is set. Using Euclidean distance and Manhattan distance as auxiliary verification methods, respectively set acceptable threshold ranges; Voiceprint matching is considered successful only when the cosine similarity is ≥85% and both the Euclidean distance and the Manhattan distance fall within the acceptable threshold range; otherwise, it is considered a mismatch and the voice data is discarded.

5. A bidirectional real-time translation method based on a unique voiceprint and hardware identity, applied to the system or device described in any one of claims 1 to 4, characterized in that, Includes the following steps: Step S1: Device initialization and voiceprint recording: The device is set to be dedicated to a specific target individual or user, and then enters a low-power sleep / standby state. Collect voice clips of a specified target individual or user, extract voiceprint feature vectors and store them locally, build a unique voiceprint whitelist, and set a similarity threshold of ≥85%; Step S2: Real-time sound pickup and voiceprint matching: The device normally operates in low-power sleep mode, continuously picking up audio and performing short-term energy analysis. When a valid sound source is detected, the voiceprint features are extracted and matched with the whitelist; A dual threshold verification method is adopted, with cosine similarity as the primary criterion and Euclidean distance and Manhattan distance as secondary criteria: if the cosine similarity is ≥85% and both the Euclidean distance and Manhattan distance fall within the qualified threshold range, the match is successful and the entire module is triggered to wake up; otherwise, the audio data is discarded and the system returns to low-power sleep mode. Step S3: Branch processing based on device type: If it is a real-time translation device specifically for a target individual, forward translation is performed, including: Collect all original audio data of the target individual; Preprocessing includes pre-emphasis, frame-segmentation windowing, and spectral subtraction for noise reduction. Extract quantized feature vectors from the three-dimensional features of spectrum, duration, and intensity; Input the local acoustic model and output the emotion-need combination label according to the principle of maximum probability. The tags are converted into human-understandable language and broadcast via TTS voice within 5 seconds after the speech ends, with text display also supported. The forward translation is applicable to all types of target individuals: pets, infants, and people with dual disabilities who are both aphasic and disabled. It is used to translate the voice of the target individual into human-understandable language so that a third party can understand it and take appropriate action. If the device is for human use only and is applied to a pet scenario, a reverse conversion is performed, including: Collect all raw voice data of users; Perform the same preprocessing; Extract phonemes and keywords, remove redundant words, and complete effective instructions; Human language can be categorized as human intention, including emotions, needs, or instructions. Based on preset fixed instruction rules, the target individual is mapped to a corresponding perceptible signal according to its type, and the signal is output within 5 seconds after the sound is emitted. The reverse conversion is only performed in pet scenarios and is used to convert human commands into sound wave signals that pets can perceive. For infants and people with dual disabilities such as aphasia and disability, there is no need to configure human-specific equipment or perform the reverse conversion step. Step S4: Device soft reset After the translation or conversion output is completed, the non-core modules are shut down, leaving only the sound pickup module and the voiceprint fast comparison submodule. The system then automatically returns to a low-power sleep state and repeats step S2.

6. The bidirectional real-time translation method based on exclusive voiceprint and hardware identity according to claim 5, characterized in that: In the forward translation, the emotion type output by the local acoustic model is preset according to the target individual type, and the demand type is also preset; if the maximum probability output by the model is lower than the preset threshold, it is determined as unrecognized and a prompt message is output.

7. The bidirectional real-time translation method based on exclusive voiceprint and hardware identity according to claim 5, characterized in that: In the reverse conversion, the mapping example of the preset fixed instruction rule includes: for pets, it is mapped to an ultrasonic frequency band signal that is adapted to the pet's hearing sensitive range.

8. The bidirectional real-time translation method based on exclusive voiceprint and hardware identity according to claim 5, characterized in that: It also includes the iterative steps of the cloud-edge collaborative model: The local device associates the sound sample, translation / conversion result, and identity ID, and then synchronizes them to the mobile terminal APP via Bluetooth. Users manually annotate samples in the app, including emotions, needs, or semantics; The app will encrypt and upload valid samples, labeled data, and identity IDs to the cloud service system; A personalized translation model is trained in the cloud based on a single individual's exclusive dataset, and a new version is released when the recognition accuracy is improved by ≥5%. Incremental update packages (≤10MB) are generated in the cloud and pushed to the corresponding apps based on the user's identity ID; The app sends the update package to the local device via Bluetooth. The local device completes the incremental update within 10 seconds and supports version rollback.

9. The bidirectional real-time translation method based on exclusive voiceprint and hardware identity according to claim 5, characterized in that: In the low-power sleep standby state, the standby power consumption is ≤1mA; the voiceprint matching time is ≤500ms; and the total time for a single round of forward translation or reverse conversion is ≤5 seconds.

10. A two-way real-time translation device based on a unique voiceprint and hardware identity, characterized in that, include: Unalterable identity verification hardware; The locally stored voiceprint feature whitelist database includes a whitelist of target individual voiceprints and a whitelist of human user voiceprints. The sound acquisition module is used to collect ambient sound. The sound source category recognition module is used to classify the collected valid sound sources into biological categories and output category labels and confidence scores. The voiceprint comparison module is used to select the corresponding voiceprint whitelist for matching based on the category label; The local processing chip has a built-in forward translation unit and a reverse conversion unit; The output module is used to output translation results or conversion signals; The low-power sleep / wake-up module is used to control the device to remain in sleep mode when there is no effective trigger, and to wake up the entire module after the sound source category is matched and the voiceprint is successfully matched; The device is configured as follows: When the category label output by the sound source category identification module is the target individual category and the confidence level is not lower than the preset threshold, the voiceprint comparison module selects the target individual voiceprint whitelist for matching. After successful matching, the forward translation unit translates the collected sound into human-understandable language and outputs it. When the category label output by the sound source category recognition module is a human user category and the confidence level is not lower than a preset threshold, the voiceprint comparison module selects the human user voiceprint whitelist for matching. After successful matching, the reverse conversion unit converts the collected speech into a signal that can be perceived by the target individual and outputs it. When the category label does not match the target individual category preset by the device, the audio data is discarded and translation is not triggered.

Images