Bluetooth headset multi-device switching and context-aware control system based on speech recognition

Abstract

The application belongs to the technical field of intelligent audio device control and speech recognition, and specifically discloses a Bluetooth earphone multi-device switching and scene sensing control system based on speech recognition. The system comprises a speech collection unit, an edge semantic understanding unit, a device state sensing unit, a heterogeneous computing scheduling unit and a dynamic power management unit. Through an end-side lightweight large language model, a speech instruction is analyzed and fused with a device state and a context to generate a structured intent, and the heterogeneous computing scheduling unit allocates computing power resources at a microsecond level to realize energy efficiency optimization in cooperation with dynamic power management. The application can accurately recognize complex speech intent without relying on the cloud, quickly switch audio source devices, reduce power consumption caused by intelligent functions, and improve the endurance and interactive experience.

Description

Technical Field

[0001] This invention belongs to the field of intelligent audio device control and voice recognition technology, specifically relating to a voice recognition-based Bluetooth headset multi-device switching and scene perception control system. Background Technology

[0002] With the rapid evolution of smart wearable devices and IoT technology, Bluetooth headsets, as a key entry point for human-computer interaction, are increasingly widely used in multi-device collaborative work and intelligent voice control. Modern wireless audio transmission systems have evolved from simple audio playback to intelligent terminals integrating complex sensing capabilities, aiming to achieve seamless switching and automatic control between multiple terminals through accurate voice recognition and scene understanding. This technological transformation from passive response to active perception places high demands on the localized processing capabilities and cross-device collaborative efficiency of the headset.

[0003] Context-aware control technology based on intent recognition has become a core path to improve the user experience of Bluetooth headsets. This technology requires the system to be able to automatically infer and execute cross-device audio stream switching or specific function adjustments by parsing the user's complex voice commands and combining them with the real-time status of the currently connected devices. To achieve deeper semantic understanding and personalized interactive services, integrating lightweight large-scale language models and dedicated neural network processing units on the headset side, along with refined resource scheduling strategies, has become a key technological direction for exploration in the current wireless audio field.

[0004] However, traditional Bluetooth headsets have extremely limited computing resources and energy reserves, creating a problem between the high computing power requirements of high-performance semantic processing models and the need for long battery life. Existing technologies often lead to a surge in system power consumption when processing complex intent recognition and lack dynamic scheduling mechanisms for heterogeneous computing cores, making it impossible to achieve microsecond-level precise allocation of computing resources based on task load. Furthermore, traditional edge-side processing solutions often suffer from recognition feedback delays or misjudgments in multi-device switching logic when parsing nonlinear contextual semantics, as the algorithm complexity and hardware energy efficiency are difficult to balance, affecting the continuity of interaction in complex dynamic scenarios.

[0005] Therefore, a voice recognition-based Bluetooth headset multi-device switching and scene perception control system is needed. Summary of the Invention

[0006] The purpose of this invention is to provide a voice recognition-based Bluetooth headset multi-device switching and scene perception control system, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The Bluetooth headset multi-device switching and scene awareness control system based on voice recognition includes a voice acquisition unit, an edge semantic understanding unit, a device status awareness unit, a heterogeneous computing scheduling unit, and a dynamic power management unit, wherein: The voice acquisition unit is configured to acquire the user's voice commands in real time, and to perform front-end noise reduction and feature extraction on the acquired audio signals to generate voice feature data that is suitable for subsequent semantic parsing. The edge semantic understanding unit is integrated inside the headphone main control chip. It has a built-in lightweight large language model with quantized compression and a dedicated neural network processing unit. It is used to receive the voice feature data, perform deep analysis of the user's intent in combination with the current context information, and output a structured intent command containing the target device identifier, switching action type and scene semantic label. The device status awareness unit is used to continuously monitor the operating status of multiple external devices that have established a Bluetooth connection with the headphones, including but not limited to the audio playback status, communication link quality, system activity and user interaction behavior of each device, and to synchronize the status information to the edge semantic understanding unit and heterogeneous computing scheduling unit in real time. The heterogeneous computing scheduling unit is configured to dynamically allocate computing tasks to one or more of the main control system-on-a-chip, artificial intelligence coprocessor, or neural network processing unit within microseconds based on the complexity, real-time requirements, and current system load of the structured intent instruction. This ensures that high-priority semantic parsing tasks receive dedicated computing resources while avoiding redundant computing overhead during low-load periods. The dynamic power management unit works in conjunction with the heterogeneous computing scheduling unit to dynamically adjust the power supply voltage, clock frequency, and module sleep state based on the task scheduling strategy, the power consumption characteristics of each computing core, and the remaining battery power, thereby optimizing the overall energy efficiency of the system.

[0008] Preferably, the lightweight large language model running by the edge semantic understanding unit undergoes knowledge distillation and parameter pruning, compressing its model size to allow it to run efficiently in resource-constrained embedded environments, while retaining its ability to understand complex speech commands with strong context dependence and high semantic ambiguity.

[0009] Furthermore, the device status awareness unit periodically polls the status information of each connected device through a low-power Bluetooth broadcast channel, and uses an event-driven mechanism to immediately trigger an interrupt when a critical state change is detected, so as to reduce the additional power consumption caused by continuous monitoring.

[0010] Furthermore, the heterogeneous computing scheduling unit has a built-in task priority evaluation engine that can automatically determine whether to enable the neural network processing unit for accelerated inference based on the time sensitivity, semantic complexity and historical interaction patterns implied in the voice commands. For simple commands, the main control system-on-a-chip will process them in the lowest power mode.

[0011] Preferably, the dynamic power management unit is equipped with a multi-level power consumption threshold strategy. When the battery level is lower than the first preset threshold, the full-speed operation mode of the neural network processing unit is automatically disabled; when the battery level is lower than the second preset threshold, only the basic audio transmission function is allowed to be executed, and all semantic understanding and scene perception tasks are suspended.

[0012] Furthermore, when parsing user voice commands, the edge semantic understanding unit can integrate multi-source contextual information from the device state perception unit, including the device currently playing audio, the type of terminal the user has recently interacted with, and the level of ambient noise, to accurately distinguish between semantically similar but different commands such as "switch to mobile phone" and "switch to laptop".

[0013] Furthermore, the system supports a cross-device intent continuation mechanism. When a user initiates an incomplete voice interaction on one device and switches to another, the edge semantic understanding unit can automatically restore and continue the original intent execution process based on the historical dialogue context and device role mapping relationship, thereby achieving a seamless cross-terminal voice control experience.

[0014] Preferably, the voice acquisition unit employs dual-microphone beamforming technology and combines it with an adaptive noise suppression algorithm, which can still extract clear voice features in noisy environments, ensuring that the input quality of subsequent semantic understanding is not affected by environmental interference.

[0015] Furthermore, a high-speed feedback channel is established between the heterogeneous computing scheduling unit and the dynamic power management unit, which can monitor the deviation between the actual power consumption and the expected model in real time during task execution, and dynamically correct subsequent scheduling decisions accordingly, forming a closed-loop energy efficiency optimization mechanism.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The Bluetooth headset multi-device switching and scene perception control system based on voice recognition provided by this invention achieves localized deep understanding of complex contextual semantics by integrating a lightweight large language model and a dedicated neural network processing unit on the terminal side. It can accurately interpret the user's true intention in a multi-device environment without relying on cloud services. Through close collaboration between the heterogeneous computing scheduling unit and the dynamic power management unit, the system can activate different computing resources on demand at the microsecond time scale, reducing the additional power consumption brought by high intelligence functions and alleviating the problem between artificial intelligence capabilities and device battery life; the introduction of the device status awareness unit enables the system to have real-time cognitive ability of multi-terminal collaborative environment, and combined with the context fusion semantic parsing mechanism, improves the accuracy and response speed of multi-device audio stream switching. The overall architecture deeply integrates edge artificial intelligence and computer architecture technologies, achieving a high-performance, low-latency, and long-lasting intelligent voice interaction experience with limited hardware resources. It solves key defects of traditional solutions such as high recognition latency, switching logic errors, and rapid battery consumption in complex dynamic scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of edge semantic understanding based on a lightweight large language model in this invention; Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow of multi-source context information fusion and structured intent instruction parsing in this invention; Figure 4 This is a logical flow diagram of the closed-loop coordination between heterogeneous computing task allocation and dynamic power management in this invention. Figure 5 This is a logical flow diagram of cross-device intent continuation execution based on device state awareness in this invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] A voice recognition-based Bluetooth headset multi-device switching and scene perception control system includes a voice acquisition unit, an edge semantic understanding unit, a device status perception unit, a heterogeneous computing scheduling unit, and a dynamic power management unit. The voice acquisition unit is configured to acquire user voice commands in real time and perform front-end noise reduction and feature extraction on the acquired audio signals to generate voice feature data suitable for subsequent semantic parsing. The voice acquisition unit is physically integrated inside the Bluetooth headset housing, and its core component includes a microelectromechanical system (MEMS) microphone array with high signal-to-noise ratio. The microphone array is arranged at preset spatial geometric intervals to form a physical basis for directional sound pickup. At the hardware logic level, the voice acquisition unit includes an adaptive beamforming logic module. This module analyzes the phase difference of signals received by different microphones and dynamically adjusts the direction of the pickup beam to accurately lock onto the user's mouth direction, achieving preliminary suppression of ambient background noise at the physical level. The voice acquisition unit also includes a high-performance digital signal processing kernel configured to run an adaptive noise suppression algorithm. This algorithm subtracts non-steady-state noise components from the mixed audio stream in real time by establishing a statistical model of ambient noise.

[0020] Furthermore, the speech acquisition unit is equipped with an acoustic feature extraction engine, which can convert the processed time-domain speech signal into a Mel-cepstral coefficient feature vector in the frequency domain, or directly extract it into high-dimensional tensor data that can be read by a deep learning model. This process is implemented through a hardware accelerator, ensuring that the conversion from raw analog audio to structured speech feature data is completed with low latency, providing high-quality, low-redundancy data input for subsequent deep semantic parsing.

[0021] The edge semantic understanding unit, integrated within the headphone's main control chip, incorporates a quantized and compressed lightweight large language model and a dedicated neural network processing unit. It receives the speech feature data, performs deep analysis of the user's intent based on current context information, and outputs a structured intent command containing the target device identifier, switching action type, and scene semantic tags. This edge semantic understanding unit is the cognitive core of the entire system, supported by hardware computing power from the dedicated neural network processing unit. This dedicated neural network processing unit features a highly parallel computing architecture, specifically designed for handling matrix operations and nonlinear activation functions. The lightweight large language model is optimized for storage space and computational overhead. It employs advanced knowledge distillation technology, using a teacher model with massive parameters as a benchmark and guiding a very small student model to learn its logical mapping relationships through a loss function. This allows the lightweight large language model to understand complex logical connections and long-distance contextual dependencies while retaining only key parameters. Furthermore, the model utilizes parameter pruning and quantization techniques, converting weight parameters from high-precision floating-point format to low-bit fixed-point format, reducing memory bandwidth requirements and memory access power consumption. In terms of operational logic, the edge semantic understanding unit is configured to maintain a continuously updated context buffer, which stores the interaction history, device status change records, and environmental parameters within the most recent preset time window. When new voice feature data is input, the edge semantic understanding unit not only analyzes the current semantic content but also uses an attention mechanism to weightedly fuse the current input with the historical information in the buffer. For example, when a user issues the vague command "switch to the past," the edge semantic understanding unit can identify, by retrieving context information, that "past" refers to the specific external device that recently generated the notification or audio activity, and generate a precise, structured intent command containing the specific device's Media Access Control address.

[0022] The device status awareness unit continuously monitors the operational status of multiple external devices connected to the headset via Bluetooth, including but not limited to the audio playback status, communication link quality, system activity, and user interaction behavior of each device, and synchronizes this status information in real time to the edge semantic understanding unit and the heterogeneous computing scheduling unit. The device status awareness unit maintains a logical connection with multiple paired devices through the underlying Bluetooth protocol stack and includes a multi-channel concurrent status monitoring module. The status monitoring module utilizes a low-power Bluetooth broadcast channel to acquire system broadcast data from each connected device in a non-connected or semi-connected manner. This data includes the device's current audio focus status, i.e., whether the device is playing media, in a call, or has a high-priority system alarm.

[0023] The device status awareness unit also integrates a link quality assessment module. This module monitors the received signal strength indicator, bit error rate, and data packet retransmission frequency in real time, calculating the stability score of each connection link. Furthermore, the device status awareness unit senses user activity on each device through interaction protocols with the operating system's underlying layer, such as touchscreen timestamps and keyboard input frequency. This multi-dimensional status information is uniformly encapsulated into a device status vector, which is transmitted in real time via a high-speed on-chip bus to the edge semantic understanding unit as a priori conditions for semantic inference, and simultaneously transmitted to the heterogeneous computing scheduling unit as a reference for adjusting computing strategies. The device status awareness unit also has anomaly detection capabilities. Once the connection quality of a device falls below a preset stability threshold, the device status awareness unit immediately sends an interrupt signal to the scheduling system, triggering a contingency plan to prevent audio interruptions or command failures during handover.

[0024] The heterogeneous computing scheduling unit is configured to dynamically allocate computing tasks to one or more of the main control system-on-a-chip, artificial intelligence coprocessor, or neural network processing unit within microseconds based on the complexity, real-time requirements, and current system load of the structured intent instructions. This ensures that high-priority semantic parsing tasks receive dedicated computing resources while avoiding redundant computing overhead during low-load periods. The heterogeneous computing scheduling unit includes a task priority evaluation engine and a hardware resource mapping module. The task priority evaluation engine automatically determines the urgency of tasks by parsing the semantic tags in the structured intent instructions. For example, tasks involving "answering a phone call" or "emergency voice interruption" are assigned the highest priority, while tasks involving "checking battery level" or "adjusting volume" are assigned a normal priority.

[0025] The hardware resource mapping module maintains a real-time load status table for the main control system-on-a-chip (SoC), AI coprocessor, and neural network processing unit, recording the current frequency setting, utilization rate, and estimated temperature rise of each core. At the moment a task is issued, the heterogeneous computing scheduling unit calculates the optimal task mapping relationship using a heuristic scheduling algorithm. For highly complex intent reasoning, the scheduling unit quickly activates the neural network processing unit and increases its clock frequency to achieve parallel acceleration; for simple control commands, they are passed to the low-power AI coprocessor or even the microcontroller unit of the SoC via direct instruction set transmission, while keeping the neural network processing unit in sleep or ultra-low power standby mode. This microsecond-level dynamic switching mechanism shortens the physical feedback loop from voice input to command execution and also finely manages the system's computing power.

[0026] The dynamic power management unit (DPMU) works in conjunction with the heterogeneous computing scheduling unit to dynamically adjust the power supply voltage, clock frequency, and module sleep state based on task scheduling strategies, the power consumption characteristics of each computing core, and the remaining battery power, thereby optimizing the overall system energy efficiency. The DPMU has a built-in set of sophisticated programmable power management integrated circuit interfaces, enabling independent level adjustment for different voltage domains. The unit internally stores multiple preset power profiles, which define the optimal voltage and frequency combination curves for each computing core under different load levels. When the heterogeneous computing scheduling unit decides to enable the neural network processing unit for high-intensity computation, the DPMU synchronously receives a voltage boost request and uses a fast regulator to raise the target core's power supply voltage to a stable operating point within nanoseconds. Conversely, when the task is completed, the unit immediately executes fast discharge logic, reducing the voltage to a hold state or directly shutting down the power rail. Furthermore, the DPMU has adaptive power sensing logic, which dynamically corrects the upper limit of power output based on the battery's coulomb counter data and internal resistance variation. If the battery level is below a first preset threshold, the unit will forcibly limit the maximum operating frequency of each core and enter a "long battery life mode". In this mode, the system will prioritize the power consumption requirements of basic audio transmission and save energy by reducing the inference depth or sampling frequency of the large language model.

[0027] The lightweight large language model running on the edge semantic understanding unit undergoes knowledge distillation and parameter pruning, compressing its model size to allow for efficient operation in resource-constrained embedded environments while retaining its ability to understand complex speech commands with strong context dependencies and high semantic ambiguity. The knowledge distillation process minimizes the cross-entropy loss between the teacher and student models in their output probability distributions, enabling the student model to capture complex semantic features such as irony, metaphor, or ellipsis based on specific scenarios. The parameter pruning process, after model training, uses an importance scoring mechanism to remove connection branches with weights close to zero, increasing the model's sparsity to a preset ratio. When this sparsified model is executed on a dedicated neural network processing unit, it can utilize a hardware-level zero-skip mechanism to further reduce the number of computational loops.

[0028] Furthermore, the device status sensing unit periodically polls the status information of each connected device through a low-power Bluetooth broadcast channel, and employs an event-driven mechanism to immediately trigger an interrupt when a critical state change is detected, thereby reducing the additional power consumption caused by continuous monitoring. The polling period is dynamically adjustable; when the system detects that the user is stationary and there is no voice interaction, the polling period automatically extends; while when the voice acquisition unit detects a user wake-up word, the polling period immediately shortens to the millisecond level. The event-driven mechanism is based on a hardware comparator logic. When a specific flag bit in the received device status message flips, an interrupt pulse is directly generated by the hardware circuit to wake up the relevant processing logic without intervention from the main control chip.

[0029] Furthermore, the heterogeneous computing scheduling unit incorporates a task priority evaluation engine, which can automatically determine whether to activate the neural network processing unit for accelerated inference based on the time sensitivity, semantic complexity, and historical interaction patterns implied in the voice commands. For simple commands, the main control system-on-a-chip (SoC) is used to process them in the lowest power mode. The historical interaction patterns refer to the system's statistics on the frequency of user usage of different functions. For frequently used simple control commands, the system pre-loads the corresponding lightweight recognition template into the SoC's local registers to achieve instantaneous response.

[0030] Preferably, the dynamic power management unit is configured with a multi-level power consumption threshold strategy. When the battery level is lower than a first preset threshold, the full-speed operation mode of the neural network processing unit is automatically disabled; when the battery level is lower than a second preset threshold, only basic audio transmission functions are allowed, and all semantic understanding and scene awareness tasks are suspended. This hierarchical defense mechanism ensures that the Bluetooth headset's most basic function as an audio transmission device remains reliable even under extremely low battery conditions, avoiding unexpected system shutdowns caused by running energy-intensive deep learning algorithms.

[0031] Furthermore, when parsing user voice commands, the edge semantic understanding unit can fuse multi-source contextual information from the device state awareness unit, including the device currently playing audio, the type of terminal the user has recently interacted with, and the ambient noise level, to accurately distinguish between semantically similar but target-different commands such as "switch to mobile phone" and "switch to laptop." The multi-source contextual information fusion process employs a weighted feature fusion method, where the system dynamically adjusts the weights of various features based on the signal-to-noise ratio of the current environment. For example, in a high-noise environment, the confidence level of voice features decreases, and the system automatically increases the reference weight for device state information, utilizing the user's current physical interaction behavior to assist in determining their intent.

[0032] Furthermore, the system supports a cross-device intent continuation mechanism. When a user initiates an incomplete voice interaction on one device and switches to another, the edge semantic understanding unit can automatically restore and continue the original intent execution flow based on the historical dialogue context and device role mapping relationship, achieving a seamless cross-terminal voice control experience. This intent continuation mechanism relies on a distributed state synchronization protocol. When a switch occurs, a snapshot of the interaction state of the original connected device is encapsulated as an intent continuation packet and transmitted to the processing logic of the new target device via the internal bus or Bluetooth data channel, allowing the user to continue the previous operation flow without repeating commands.

[0033] Preferably, the voice acquisition unit employs dual-microphone beamforming technology combined with an adaptive noise suppression algorithm, enabling the extraction of clear voice features even in noisy environments, ensuring that the input quality for subsequent semantic understanding is unaffected by environmental interference. The beamforming technology calculates the incident angle of the sound source by estimating the time delay between the two microphones, and weights and attenuates signals from non-target angles in the digital domain, forming a narrow "sound pickup window" in the spatial domain. This improves the purity of sound pickup in complex environments such as wind noise and human voice interference.

[0034] Furthermore, a high-speed feedback channel is established between the heterogeneous computing scheduling unit and the dynamic power management unit. This channel enables real-time monitoring of the deviation between actual power consumption and the expected model during task execution, and dynamically corrects subsequent scheduling decisions accordingly, forming a closed-loop energy efficiency optimization mechanism. The high-speed feedback channel transmits real-time current sensing data for each core. If the actual power consumption of a task on the neural network processing unit exceeds a preset percentage of the predicted value, the scheduling unit will determine that the current model is running inefficient and attempt to balance energy efficiency by adjusting memory read strategies or reducing computational precision.

[0035] Example 2: As an alternative implementation of the system architecture of the present invention, this example describes a Bluetooth headset control system based on distributed computing offloading logic. In this implementation, although the core logic remains at the headset edge, the system adds the flexibility to coordinate with smart terminals.

[0036] A voice recognition-based Bluetooth headset multi-device switching and scene perception control system includes a front-end acquisition module, an edge inference module, a collaborative communication module, a global resource management module, and a hierarchical power supply control module. The front-end acquisition module functions identically to the voice acquisition unit in Embodiment 1, but with the addition of a bone conduction sensor in its hardware implementation. The bone conduction sensor is positioned close to the user's ear canal wall to acquire vibration signals from human tissue caused by the user's vocalization. The front-end acquisition module is configured to fuse the bone conduction signal with the air conduction microphone signal. This dual-modal acquisition method effectively distinguishes between the user's own voice and ambient noise, especially in high-noise industrial environments or densely populated areas. Bone conduction signals, as a reliable basis for voice activity detection, can significantly reduce the false wake-up rate of the edge inference module.

[0037] The edge inference module integrates a miniaturized hybrid of convolutional neural networks and recurrent neural networks. Unlike Embodiment 1, this edge inference module is primarily responsible for the initial classification and categorization of voice commands. When a command is identified as belonging to a preset simple command set, logical processing is performed directly locally. However, when a received voice command is determined to be a "supercomputing power demand task," the edge inference module automatically extracts key tensors of the voice features.

[0038] The collaborative communication module establishes a high-speed data transmission channel between the headset and the connected device. When the edge inference module determines that the current computing load exceeds the local processing limit, the collaborative communication module sends the extracted key speech feature tensors to the currently active smartphone or tablet via Bluetooth high-speed mode or a proprietary communication protocol. Leveraging the powerful CPU and GPU capabilities of the connected device, larger-scale semantic understanding models can be executed in the cloud or locally on the device. This distributed processing approach enables complex natural language interactions without significantly increasing headset power consumption.

[0039] The global resource management module, acting as the command center of the entire heterogeneous system, not only manages the computing resources within the headset but also monitors the available computing power of connected terminals. This module includes a latency prediction engine that calculates the difference between local processing latency and data transmission latency plus remote processing latency in real time. If it determines that latency caused by data transmission would impair the user experience, the global resource management module will force a degraded, lightweight model to be used locally for a faster response, achieving a dynamic balance between response speed and recognition depth.

[0040] The tiered power supply control module is optimized for distributed computing scenarios. When computing tasks are offloaded to external devices, the module sets the neural network processing unit inside the headset to deep sleep mode and shuts down the relevant memory bus. Simultaneously, it monitors the transmission power of the Bluetooth wireless RF module and dynamically adjusts the output power level based on the distance to the connected device to ensure energy efficiency during data transmission. If the battery voltage is detected to be in a low-voltage warning range, the tiered power supply control module sends a command to the global resource management module to completely stop external offloading operations and revert to the most basic low-power local control mode.

[0041] In this embodiment, the edge semantic understanding process is no longer limited to the computing power of a single headset, but rather constructs a fluid computing environment with the headset as the sensing antenna and connected devices as the computing power backing. The device status perception logic is also extended to perceive the battery level of external devices. If the connected smartphone is low on battery, the collaborative communication module will proactively reduce the frequency of sending uninstallation requests to that phone, demonstrating system-level context awareness and collaborative intelligence.

[0042] Example 3: This example focuses on describing a system implementation optimized for highly dynamic mobile scenarios, emphasizing the deep integration of motion state and semantic understanding.

[0043] The Bluetooth headset multi-device switching and scene perception control system based on voice recognition includes a motion perception voice enhancement unit, an environmental semantic fusion understanding unit, a multi-link adaptive scheduling unit, a behavior prediction intent parsing unit, and an intelligent power management unit.

[0044] The motion-aware speech enhancement unit, in addition to basic audio acquisition capabilities, integrates a six-axis inertial measurement unit (IMU). This unit monitors the user's motion status in real time, including cadence, head posture angle, and speed. Since microphones are subject to strong wind noise and increased breathing noise when running or cycling, the unit is configured to dynamically adjust the filter coefficients of the adaptive noise reduction algorithm based on the motion parameters output by the IMU. For example, when high-speed movement is detected, the system automatically activates a wind noise suppression algorithm and uses gravitational acceleration information to correct beamforming pointing offset, ensuring that the acquired speech features maintain high confidence even in dynamic environments.

[0045] The environmental semantic fusion understanding unit is configured to perceive the characteristics of the surrounding physical environment. It collects the spectral characteristics of ambient noise through a microphone to identify the user's current scene, such as an office, subway, coffee shop, or outdoor street. These scene semantic labels are input into the edge semantic understanding unit. When the user issues a "speak louder" command in a subway environment, the unit not only adjusts the volume but also automatically enhances the gain of the human voice frequency range and activates passive noise isolation compensation logic. This understanding based on physical environment semantics makes the headphone's control logic more aligned with the user's actual perceptual needs.

[0046] The multi-link adaptive scheduling unit is optimized for scenarios involving frequent switching between multiple devices. Internally, it includes a link stability prediction model that uses time-series analysis of historical signal strength to predict the probability of each Bluetooth link dropping within the next few seconds. When a user moves between rooms, causing the signal of the currently connected device to weaken, the multi-link adaptive scheduling unit pre-establishes a connection with a backup device with a stronger signal in the background and synchronizes some semantic understanding results to the backup link, achieving a smooth, "zero-aware" task switching.

[0047] The behavior prediction intent parsing unit utilizes a long short-term memory network to analyze the user's historical behavior sequences. For example, if a user habitually listens to music before answering a phone call during their morning commute, when the system detects an incoming call in the morning while the user is in motion, its intent parsing logic will pre-increase the weight of the intent to "switch to call." This behavior prediction-based feedforward control mechanism enables the system to achieve high response accuracy with low computational cost.

[0048] The intelligent power management unit incorporates energy recovery monitoring logic. While Bluetooth headsets typically lack built-in energy recovery hardware, this unit generates a detailed energy flow diagram by monitoring impedance changes in the audio output stage and the real-time power consumption of each system module. Based on this diagram, the unit dynamically adjusts the depth of semantic understanding. During intense exercise, when other sensors such as heart rate sensors may be activated simultaneously, the unit dynamically reduces the power consumption of the voice processing module to ensure that the system does not trigger battery protection due to excessive instantaneous current during multi-tasking.

[0049] Example 4: This example describes a system implementation method based on enhanced security encryption and privacy protection.

[0050] The Bluetooth headset multi-device switching and scene perception control system based on voice recognition includes a security domain isolation acquisition unit, a privacy compliance semantic understanding unit, an encrypted command scheduling unit, and a secure power monitoring unit.

[0051] The secure domain isolation acquisition unit employs an isolation design based on a trusted execution environment in its hardware architecture. After the voice signal is converted from analog to digital, it directly enters a hardware-level secure area for preprocessing. All data involving user biometric features (such as voiceprint information) is processed within the isolated area and is never exposed to the ordinary system bus, thus preventing the possibility of voice commands being eavesdropped on or tampered with by malicious third-party programs from the source.

[0052] The privacy-compliant semantic understanding unit operates a lightweight, large language model that has undergone de-identification training. When processing voice commands containing personal privacy information (such as schedules, payment instructions, or private contact names), this lightweight, large language model automatically triggers localized encryption processing logic. This privacy-compliant semantic understanding unit does not upload such sensitive semantics to any external devices; all intent parsing is completed in a closed loop within the trusted computing domain of the headset. Only verified structured control commands are allowed to be sent to the Bluetooth protocol stack for device switching operations.

[0053] The encrypted instruction scheduling unit is responsible for performing end-to-end encryption when transmitting switching instructions between different devices. It uses a shared key established with each connected device to sign and encrypt each structured intent instruction. Upon receiving the instruction, the target device must authenticate before executing the corresponding switching or control action. This prevents attackers from illegally controlling multiple user devices by forging Bluetooth control protocols.

[0054] The aforementioned secure power monitoring unit functions to prevent attacks exploiting power consumption side channels. This unit incorporates randomized noise jitter during dynamic power management, preventing external devices from inferring the weights of the running semantic understanding model or specific user input by monitoring real-time current fluctuations in the headphones. This power-level protection mechanism provides physical-layer security for the operation of edge AI.

[0055] Through the detailed description of the above embodiments, it can be seen that the present invention solves the problem of computing power anxiety and battery life contradiction in the process of intelligent transformation of traditional Bluetooth headsets by deeply integrating edge large language models, NPU acceleration, heterogeneous computing scheduling, and dynamic power management. Each level of the system reflects the efficient use of limited resources and the precise control of complex interaction scenarios.

[0056] In the above embodiments, the communication logic between the system units adopts a lossless data message format, and all processes involving numerical comparison and logical judgment are described in a purely textual manner. For example, when describing heterogeneous computing scheduling, the "load status score" mentioned refers to a scalar value obtained by weighted summation of processor utilization, memory usage percentage, and the number of currently running tasks. When this scalar value exceeds a preset first evaluation threshold, the system determines it to be in a high-load state and calls upon computing cores with higher computing power. This descriptive method fully complies with the specifications for patent drafting, transforming complex mathematical relationships into clear logical steps.

[0057] Furthermore, the lightweight large language model involved in this invention uses fixed-point conversion in its quantization process, which refers to projecting the weight values ​​represented by 32-bit floating-point numbers into an 8-bit or 4-bit integer space through a linear mapping algorithm. During inference, the system uses bitwise operations instead of traditional floating-point multiplication, achieving high energy efficiency at the hardware circuit level. This textual engineering description clearly reveals the physical implementation path of the technical solution, providing a complete and feasible reference guide for those skilled in the art.

[0058] This invention provides a voice recognition-based Bluetooth headset multi-device switching and scene perception control system. Through the collaborative work of multiple units, it achieves a complete intelligent closed loop from perception, understanding, scheduling to execution. The localized processing capability of the edge semantic understanding unit ensures user privacy and real-time interaction; the closed-loop collaboration between the heterogeneous computing scheduling unit and the dynamic power management unit overcomes the bottleneck of limited battery power in wearable devices; and the global perspective of the device status perception unit endows the headset with multi-terminal command capabilities, acting as a smart hub. The overall solution achieves an intelligent interactive experience comparable to cloud processing while maintaining a lightweight hardware structure, demonstrating advanced technology and practical value.

[0059] It should be understood that the above embodiments are merely illustrative examples of the system architecture of the present invention, and not limitations on the scope of protection of the present invention. Within the technical framework of the present invention, any variations based on the same edge intent computation and context-aware scheduling concepts, including equivalent replacements of module names and rearrangements of hardware components, should be included within the scope of protection of the present invention. The various algorithmic strategies employed in the system can be specifically optimized according to the instruction set of the specific chip platform in actual engineering implementation. Such adaptive adjustments based on specific hardware environments also fall within the scope of disclosure of the present invention.

[0060] Finally, the various units and modules described in this invention can be implemented using discrete hardware components or integrated into a single system-on-a-chip (SoC) through software-defined hardware. In modern semiconductor processes, integrating the voice acquisition, semantic understanding, state awareness, task scheduling, and power management logic onto the same silicon substrate and utilizing a high-speed on-chip network for data interaction is the optimal path to realizing the system described in this invention. This highly integrated system-on-a-chip (SoC) approach can further reduce interconnect power consumption and improve the overall system response bandwidth, representing an important direction for the industrial application of this invention.

[0061] In summary, this invention, through the deep integration and innovation of edge artificial intelligence and computer architecture, provides a high-performance, low-power, and highly perceptive control solution for next-generation smart wearable devices. By deeply mining contextual information and micro-scheduling hardware resources, it improves user interaction efficiency in multi-device environments, solves the core pain points of wireless audio systems in complex scenarios such as low recognition rate, high switching latency, and short battery life, and lays a solid technical foundation for achieving truly seamless intelligent interaction.

[0062] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A Bluetooth headset multi-device switching and scene perception control system based on voice recognition, characterized in that, include: The voice acquisition unit is configured to acquire users' voice commands in real time and perform front-end noise reduction and feature extraction on the acquired audio signals to generate voice feature data that is suitable for subsequent semantic parsing. The edge semantic understanding unit is integrated into the headphone main control chip. It has a built-in lightweight large language model with quantized compression and a dedicated neural network processing unit. It is used to receive the voice feature data, perform deep analysis of the user's intent in combination with the current context information, and output a structured intent command containing the target device identifier, switching action type and scene semantic label. The device status awareness unit is used to continuously monitor the operating status of multiple external devices that have established a Bluetooth connection with the earphone, and synchronize the status information to the edge semantic understanding unit and the heterogeneous computing scheduling unit in real time. The heterogeneous computing scheduling unit is configured to dynamically allocate computing tasks to one or more of the main control system-on-a-chip, artificial intelligence coprocessor, or neural network processing unit within microseconds based on the complexity, real-time requirements, and current system load of the structured intent instructions, ensuring that high-priority semantic parsing tasks obtain dedicated computing resources. The dynamic power management unit works in conjunction with the heterogeneous computing scheduling unit to dynamically adjust the power supply voltage, clock frequency, and module sleep state based on the task scheduling strategy, the power consumption characteristics of each computing core, and the remaining battery power.

2. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The voice acquisition unit includes: a microelectromechanical system microphone array, which consists of multiple microphones with high signal-to-noise ratio arranged at a preset spatial geometric spacing, forming the physical basis for directional sound pickup; The adaptive beamforming logic module is configured to analyze the phase difference of signals received by different microphones, dynamically adjust the direction of the pickup beam to lock the direction of the user's mouth, and suppress ambient background noise at the physical level. The digital signal processing kernel is configured to run an adaptive noise suppression algorithm that subtracts non-steady-state noise components from the mixed audio stream in real time by establishing a statistical model of the ambient noise. The acoustic feature extraction engine is configured to convert the processed time-domain speech signal into a Mel-Cepstral Coefficient feature vector in the frequency domain, or directly extract it as high-dimensional tensor data for deep learning models to read.

3. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The edge semantic understanding unit is configured to execute the following logic: the lightweight large language model is constructed through knowledge distillation technology, using a preset teacher model as a benchmark, and a loss function is used to guide a small student model to learn its logical mapping relationship, so that it can understand long-distance contextual dependencies while retaining key parameters. The lightweight large language model applies parameter pruning and quantization techniques to convert the weight parameters from high-precision floating-point format to low-bit fixed-point format, thereby reducing memory bandwidth requirements and memory access power consumption. The edge semantic understanding unit maintains a rolling updated context buffer to store interaction history, device status change records, and environmental parameters within a preset time window; The current speech feature data is weighted and fused with historical information in the context buffer through an attention mechanism. The external device pointed to by the fuzzy instruction is identified by retrieving context information, and a structured intent instruction containing the specific device's media access control address is generated.

4. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The device status sensing unit includes: a multi-channel concurrent status monitoring module, which uses a low-power Bluetooth broadcast channel to acquire system broadcast data of each connected device in a non-connected or semi-connected manner. The system broadcast data includes the audio focus status, media playback status, call status and system alarm priority of each device. The link quality assessment module is used to monitor the received signal strength indication, bit error rate and data packet retransmission frequency in real time, and calculate the stability score of each connection link. The user activity perception module obtains the user's touch timestamps and input frequency on each device through the interaction protocol with the operating systems of each external device, and encapsulates the multi-dimensional state information into a device state vector, which is then transmitted to the edge semantic understanding unit and the heterogeneous computing scheduling unit through the on-chip bus.

5. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The heterogeneous computing scheduling unit includes: a task priority evaluation engine, used to parse the semantic tags in the structured intent instruction, and automatically determine the urgency of the task and assign corresponding priority weights based on the task's time sensitivity, semantic complexity and historical interaction patterns. The hardware resource mapping module is used to maintain the load status tables of the main control system-on-a-chip, artificial intelligence coprocessor and neural network processing unit in real time, and record the current frequency setting, utilization and estimated temperature rise of each core. The heuristic scheduling algorithm module is configured to calculate the optimal task mapping relationship at the moment the task is issued. For high-complexity intention reasoning tasks, the neural network processing unit is activated and its clock frequency is increased. For simple control instructions, the processing is handled by the artificial intelligence coprocessor or the main control system chip, and the neural network processing unit is kept in a sleep or low-power standby state.

6. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The dynamic power management unit includes: a programmable power management integrated circuit interface, used to independently adjust the level of different voltage domains in the system; The power profile memory stores profiles that define the optimal voltage and frequency combination curves for each computing core under different load levels. The voltage regulation logic module is configured to, upon receiving a voltage boost request from the heterogeneous computing scheduling unit, boost the power supply voltage of the target core to a stable operating point within nanoseconds, and execute fast discharge logic after the task is completed. The adaptive power sensing logic module dynamically adjusts the upper limit of power output based on the battery's coulomb data and internal resistance variation. When the battery power is lower than the first preset threshold, it forcibly limits the maximum operating frequency of each core.

7. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The system also has a cross-device intent continuation mechanism: when parsing voice commands, the edge semantic understanding unit adopts a weighted feature fusion method, which combines the device type of the currently playing audio, the type of the most recently interacted terminal, and the level of environmental noise, and dynamically adjusts the weight ratio of voice features and device status information to distinguish target devices with similar semantics. When a user initiates an incomplete voice interaction on the first device and switches to the second device, the system encapsulates a snapshot of the interaction state of the original connected device into an intent continuation package, which is then transmitted to the processing logic of the new target device via the internal bus or Bluetooth data channel. Based on the historical dialogue context and the mapping relationship between the device roles, the original intent execution process is automatically restored and continued.

8. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The system also includes a distributed computing offloading module: the voice acquisition unit is equipped with a bone conduction sensor to collect the vibration signal of human tissue caused by the user's voice, and fuse the bone conduction signal with the air conduction microphone signal to distinguish the user's own voice from environmental interference sound; The collaborative communication module is used to establish a high-speed data pass-through channel between the headset and the connected device. When the edge inference load exceeds the local processing limit, the key tensors of the voice features are sent to the external smart terminal to execute a large-scale semantic understanding model. The global resource management module has a built-in latency prediction engine, which is used to calculate the difference between local processing latency and data transmission plus remote processing latency in real time, and to dynamically balance response speed and recognition depth.

9. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The system also includes a motion-aware speech enhancement unit and a multi-link adaptive scheduling unit: the motion-aware speech enhancement unit integrates a six-axis inertial measurement unit to monitor the user's step frequency, head posture angle and movement speed, and dynamically adjusts the filter coefficients of the adaptive noise reduction algorithm according to the motion parameters, and uses gravity acceleration information to correct the beamforming pointing offset. The environmental semantic fusion understanding unit identifies the physical scene label where the user is located by collecting the spectral characteristics of the environmental background noise, and inputs it into the edge semantic understanding unit to adjust the volume gain or enable the passive noise reduction compensation logic. The multi-link adaptive scheduling unit has a built-in link stability prediction model. It predicts the probability of each Bluetooth link dropping by analyzing the time sequence of historical signal strength. When the signal of the currently connected device weakens, it establishes a pre-connection state with a backup device with a higher signal strength in advance.

10. The Bluetooth headset multi-device switching and scene perception control system based on voice recognition according to claim 1, characterized in that, The system also includes a security and privacy protection module: a security domain isolation acquisition unit, which adopts a hardware isolation design based on a trusted execution environment, directly sending the digitized data after the voice signal is converted into a trusted area for preprocessing, ensuring that biometric feature data is not exposed to the ordinary system bus; The privacy-compliant semantic understanding unit is configured to automatically trigger localized encryption logic when processing voice commands containing personal privacy information, and to prohibit sensitive semantic data from being uploaded to external devices; The safety power monitoring unit is configured to introduce randomized noise jitter during dynamic power management. By masking the real-time current fluctuation characteristics of the headphones, it prevents external devices from inferring the operating status of the semantic understanding model or the content of user commands through power consumption side-channel analysis.

Images