Wireless earphone audio transmission method based on Bluetooth

Abstract

The invention discloses a Bluetooth-based wireless earphone audio transmission method. The method comprises the following steps: acquiring environment audio matrix data through a sensor group in a wireless earphone; extracting noise spectrum features in the environment audio matrix data by using an improved CNN convolutional neural network, and eliminating motion friction noise in the matrix data through an adaptive filtering algorithm to obtain dual-channel noise reduction data; inputting the dual-channel noise reduction data into the fusion model to recognize a current scene of a user, predicting an available bandwidth within 5ms in the future in combination with a Bluetooth RSS I signal and historical channel data, and switching a codec according to a scene recognition result and the available bandwidth; fragmentation is carried out based on a codec, and an audio data stream is segmented into a high-priority data packet and a low-priority data packet, and the high-priority data packet and the low-priority data packet are transmitted through a Bluetooth main channel and an auxiliary channel respectively. The transmission success rate of the audio data is improved, the phenomena of audio lagging and distortion are effectively reduced, and the auditory experience and the use satisfaction of a user are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless transmission, and in particular to a Bluetooth-based wireless earphone audio transmission method. Background Art

[0002] During audio transmission, environmental noise and friction noise generated by the movement of headphones seriously affect the sound quality. Traditional noise reduction technology often uses filters with fixed parameters, which cannot adjust the noise reduction strategy in real time according to the complex and changing environment and the user's movement status, resulting in poor noise reduction effect. Users find it difficult to get a clear audio experience in noisy environments or sports scenes. At the same time, the bandwidth of the Bluetooth channel is limited and dynamically changing. Traditional methods find it difficult to accurately predict the available bandwidth. When the channel quality deteriorates, the codec cannot be adjusted in time, which can easily cause audio freezes, distortion and other problems. In addition, existing audio data transmission methods usually do not distinguish between data priorities. When bandwidth is limited, the reliable transmission of key audio information cannot be guaranteed, which seriously affects the user's auditory experience. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and design a Bluetooth-based wireless headset audio transmission method.

[0004] To achieve the above-mentioned purpose, the technical solution of the present invention is as follows: further, in the above-mentioned Bluetooth-based wireless headset audio transmission method, the wireless headset audio transmission method includes the following steps:

[0005] The sensor group in the wireless headset obtains the user's motion status, head posture and environmental noise to obtain sensor data. The audio data and sensor data are aligned by time stamp to obtain environmental audio matrix data.

[0006] The improved CNN convolutional neural network is used to extract the noise spectrum characteristics in the environmental audio matrix data, and the motion friction noise in the matrix data is eliminated by an adaptive filtering algorithm to obtain dual-channel noise reduction data;

[0007] The dual-channel noise reduction data is input into the BERT-ResNet fusion model to identify the user's current scenario, combine the Bluetooth RSS I signal and historical channel data, predict the available bandwidth within the next 5ms, and switch the codec based on the scenario recognition result and the available bandwidth;

[0008] The audio data stream is segmented based on the codec and divided into high-priority and low-priority data packets, which are transmitted through the Bluetooth primary channel and auxiliary channel respectively.

[0009] Furthermore, in the above-mentioned Bluetooth-based wireless headset audio transmission method, aligning the audio data and the sensor data by timestamp to obtain the ambient audio matrix data includes:

[0010] Clock synchronization is performed through the Bluetooth protocol. A bidirectional synchronization algorithm based on a reference timestamp reduces the clock deviation of sensor data and audio data to obtain synchronized data.

[0011] Establishing a timestamp mapping table, and performing interpolation processing on the sensor data in the synchronization data based on the timestamp of the audio data to obtain interpolated data;

[0012] The linear interpolation method is used to estimate the sensor state at the intermediate moment; the sensor data whose timestamp is later than the audio frame is time-aligned to obtain aligned data;

[0013] Based on the alignment data, environmental audio matrix data including audio data, motion state data, head posture data and environmental noise data is established, where the matrix rows represent time series and the matrix columns correspond to data dimensions.

[0014] Furthermore, in the above-mentioned Bluetooth-based wireless headset audio transmission method, the step of extracting noise spectrum features from the ambient audio matrix data using an improved CNN convolutional neural network includes:

[0015] The improved CNN convolutional neural network is based on a multi-layer structure that alternates convolutional and pooling layers, including three convolutional layers, two maximum pooling layers, and two fully connected layers. The convolutional layers use 3×3 convolution kernels to capture local spectral features, and the pooling layers use 2×2 windows for downsampling.

[0016] Add skip connections between the input and output of adjacent convolutional layers to add the input to the convolutional output;

[0017] Add a batch normalization layer after each convolutional layer and fully connected layer to normalize the data, and use the LeakyReLU activation function to replace the ReLU activation function of the CNN convolutional neural network;

[0018] An improved CNN convolutional neural network is used to extract noise spectrum features in the environmental audio matrix data to obtain noise spectrum features.

[0019] Furthermore, in the above-mentioned Bluetooth-based wireless headset audio transmission method, eliminating motion friction noise in the matrix data by an adaptive filtering algorithm to obtain dual-channel noise reduction data includes:

[0020] Obtain the ambient audio matrix data, determine whether there is motion friction noise and generate the corresponding noise reference signal;

[0021] The noise spectrum characteristics are obtained, the coefficients of the LMS filter are adjusted in real time, the adjusted filter is applied to the ambient audio matrix data, and the motion friction noise is removed through the adaptive filtering algorithm to obtain dual-channel noise reduction data.

[0022] Furthermore, in the above-mentioned Bluetooth-based wireless headset audio transmission method, inputting the dual-channel noise reduction data into the BERT-ResNet fusion model to identify the user's current scene includes:

[0023] Use the residual block in the deep residual network ResNet-18 to extract the spatial features of audio data and output a high-dimensional spatial feature vector;

[0024] The pre-trained BERT model is used to process audio sequence data, capture the contextual dependencies of audio signals in time series, and extract sequence features.

[0025] The spatial feature vector output by ResNet and the sequence feature vector output by BERT are concatenated to obtain a fused feature vector containing spatial and temporal information.

[0026] The fused feature vector is input into the fully connected layer for scene classification to identify the scene the user is currently in.

[0027] Furthermore, in the above-mentioned Bluetooth-based wireless headset audio transmission method, the method combines the Bluetooth RSS I signal and historical channel data to predict the available bandwidth within the next 5 ms, and switches the codec based on the scene recognition result and the available bandwidth, including:

[0028] Establish a correspondence between scenarios and available bandwidth. Choose a high-quality codec for quiet indoor scenarios, a codec that balances sound quality and anti-interference capabilities for sports scenarios, and a codec that prioritizes low bitrate and reliability for complex outdoor scenarios.

[0029] When the scene recognition result changes and the predicted available bandwidth exceeds the applicable range of the current codec, the codec switching process is triggered.

[0030] Furthermore, in the above-mentioned Bluetooth-based wireless headset audio transmission method, the codec-based segmentation is performed to divide the audio data stream into high-priority and low-priority data packets, which are transmitted through the Bluetooth primary channel and auxiliary channel respectively, including:

[0031] FHSS frequency hopping spread spectrum technology is used to select frequencies with good signal quality for transmission. High-priority data packets are protected by Reed-Solomon coding, with a code length of n=255 and an RS code with k=223 information bits.

[0032] Low-priority data packets use the ARQ automatic repeat request mechanism. After sending the data packet, the sender starts a timer to wait for the confirmation signal from the receiver.

[0033] Furthermore, in a system for implementing a Bluetooth-based wireless headset audio transmission method, the system includes the following modules:

[0034] The matrix data acquisition module is used to obtain the user's motion status, head posture and environmental noise through the sensor group in the wireless headset to obtain sensor data, and align the audio data and sensor data through timestamps to obtain environmental audio matrix data;

[0035] A matrix data denoising module is used to extract noise spectrum features from the ambient audio matrix data using an improved CNN convolutional neural network, eliminate motion friction noise from the matrix data using an adaptive filtering algorithm, and obtain dual-channel denoised data;

[0036] A decoder switching module is used to input the dual-channel noise reduction data into the BERT-ResNet fusion model to identify the user's current scenario, combine the Bluetooth RSSI signal and historical channel data, predict the available bandwidth within the next 5ms, and switch the codec based on the scenario recognition result and the available bandwidth;

[0037] The audio channel transmission module is used to perform segmentation based on the codec, divide the audio data stream into high-priority and low-priority data packets, and transmit them through the Bluetooth main channel and auxiliary channel respectively.

[0038] Furthermore, in a system for implementing a Bluetooth-based wireless headset audio transmission method, the system includes the following submodules:

[0039] The synchronization submodule is used to synchronize the clocks through the Bluetooth protocol. The bidirectional synchronization algorithm based on the reference timestamp reduces the clock deviation of the sensor data and audio data to obtain synchronized data.

[0040] An interpolation submodule, configured to establish a timestamp mapping table, and perform interpolation processing on the sensor data in the synchronization data based on the timestamp of the audio data to obtain interpolated data;

[0041] The alignment submodule is used to estimate the sensor state at the intermediate moment using the linear interpolation method; the sensor data whose timestamp is later than the audio frame is time-aligned to obtain aligned data;

[0042] A submodule is established to establish environmental audio matrix data including audio data, motion state data, head posture data and environmental noise data based on the alignment data, where the matrix rows represent time series and the matrix columns correspond to data dimensions.

[0043] Furthermore, in a system for implementing a Bluetooth-based wireless headset audio transmission method, the system includes the following submodules:

[0044] Set up submodules to determine the improved CNN convolutional neural network structure based on alternating layers of convolution and pooling layers, including three convolutional layers, two max pooling layers, and two fully connected layers. The convolutional layers use 3×3 convolution kernels to capture local spectral features, and the pooling layers use 2×2 windows for downsampling.

[0045] Add a submodule to add skip connections between the input and output of adjacent convolutional layers, adding the input to the convolutional output;

[0046] The replacement submodule is used to add a batch normalization layer after each convolutional layer and fully connected layer to normalize the data and use the LeakyReLU activation function to replace the ReLU activation function of the CNN convolutional neural network;

[0047] The extraction submodule is used to extract the noise spectrum features in the environmental audio matrix data using an improved CNN convolutional neural network to obtain the noise spectrum features.

[0048] The beneficial effect is that the user's motion state, head posture and environmental noise are obtained through the sensor group in the wireless headset to obtain sensor data, and the audio data and sensor data are aligned by timestamp to obtain environmental audio matrix data; the noise spectrum characteristics in the environmental audio matrix data are extracted using an improved CNN convolutional neural network, and the motion friction noise in the matrix data is eliminated by an adaptive filtering algorithm to obtain dual-channel noise reduction data; the dual-channel noise reduction data is input into the BERT-ResNet fusion model to identify the user's current scene, and the available bandwidth within the next 5ms is predicted by combining the Bluetooth RSS I signal and historical channel data. The codec is switched according to the scene recognition result and the available bandwidth; based on the codec, the audio data stream is segmented into high-priority and low-priority data packets, which are transmitted through the Bluetooth main channel and auxiliary channel respectively. 1. It can accurately extract noise spectrum characteristics and eliminate motion friction noise. Compared with traditional noise reduction methods, the noise reduction effect is improved, the purity of the audio is effectively improved, and clearer sound quality is provided to users. 2. Dynamically switches codecs based on the scenario and available bandwidth, improving audio transmission fluency in complex environments and matching audio quality with the current usage scenario and channel conditions, meeting the user's personalized needs in different scenarios. 3. Improves the success rate of audio data transmission, effectively reducing audio freezes and distortion, and enhancing the user's listening experience and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0050] Figure 1 Schematic diagram of a first embodiment of a Bluetooth-based wireless headset audio transmission method in an embodiment of the present invention;

[0051] Figure 2 2 is a schematic diagram of a second embodiment of a Bluetooth-based wireless headset audio transmission method according to an embodiment of the present invention;

[0052] Figure 3 FIG3 is a schematic diagram of a third embodiment of a Bluetooth-based wireless headset audio transmission method in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0054] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a", "an", "" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0055] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 As shown, a Bluetooth-based wireless headset audio transmission method includes the following steps:

[0056] Step 101: Obtain user motion status, head posture, and ambient noise through a sensor group in the wireless headset to obtain sensor data, and align the audio data and sensor data through timestamps to obtain ambient audio matrix data;

[0057] Specifically, in this embodiment, clock synchronization is performed through the Bluetooth protocol, and a bidirectional synchronization algorithm based on a reference timestamp is used to reduce clock deviations between sensor data and audio data to obtain synchronized data;

[0058] Establish a timestamp mapping table, and interpolate the sensor data in the synchronization data based on the timestamp of the audio data to obtain interpolated data;

[0059] The linear interpolation method is used to estimate the sensor state at the intermediate moment; the sensor data whose timestamp is later than the audio frame is time-aligned to obtain aligned data;

[0060] Based on the aligned data, environmental audio matrix data including audio data, motion state data, head posture data and environmental noise data is established. The matrix rows represent time series and the matrix columns correspond to data dimensions.

[0061] Step 102: Using an improved CNN convolutional neural network to extract noise spectrum features in the ambient audio matrix data, and using an adaptive filtering algorithm to eliminate motion friction noise in the matrix data to obtain dual-channel noise reduction data;

[0062] Specifically, the improved CNN convolutional neural network in this embodiment is based on a multi-layer structure that alternates convolutional layers and pooling layers, including three convolutional layers, two maximum pooling layers, and two fully connected layers. The convolutional layer uses a 3×3 convolution kernel to capture local spectral features, and the pooling layer uses a 2×2 window for downsampling.

[0063] Add skip connections between the input and output of adjacent convolutional layers to add the input to the convolutional output;

[0064] Add a batch normalization layer after each convolutional layer and fully connected layer to normalize the data, and use the LeakyReLU activation function to replace the ReLU activation function of the CNN convolutional neural network;

[0065] The improved CNN convolutional neural network is used to extract the noise spectrum features in the environmental audio matrix data to obtain the noise spectrum features.

[0066] Obtain the ambient audio matrix data, determine whether there is motion friction noise and generate the corresponding noise reference signal;

[0067] The noise spectrum characteristics are obtained, the coefficients of the LMS filter are adjusted in real time, the adjusted filter is applied to the ambient audio matrix data, and the motion friction noise is removed through the adaptive filtering algorithm to obtain dual-channel noise reduction data.

[0068] Step 103: Input the dual-channel noise reduction data into the BERT-ResNet fusion model to identify the user's current scenario. Combine the Bluetooth RSSI signal and historical channel data to predict the available bandwidth within the next 5ms. Switch the codec based on the scenario identification result and the available bandwidth.

[0069] Specifically, in this embodiment, the residual block in the deep residual network ResNet-18 is used to extract the spatial features of the audio data and output a high-dimensional spatial feature vector;

[0070] The pre-trained BERT model is used to process audio sequence data, capture the contextual dependencies of audio signals in time series, and extract sequence features.

[0071] The spatial feature vector output by ResNet and the sequence feature vector output by BERT are concatenated to obtain a fused feature vector containing spatial and temporal information.

[0072] The fused feature vector is input into the fully connected layer for scene classification to identify the scene the user is currently in.

[0073] Establish a correspondence between scenarios and available bandwidth. Choose a high-quality codec for quiet indoor scenarios, a codec that balances sound quality and anti-interference capabilities for sports scenarios, and a codec that prioritizes low bitrate and reliability for complex outdoor scenarios.

[0074] When the scene recognition result changes and the predicted available bandwidth exceeds the applicable range of the current codec, the codec switching process is triggered.

[0075] Step 104: Segment the audio data stream into high-priority and low-priority data packets based on the codec, and transmit them through the Bluetooth primary channel and auxiliary channel respectively.

[0076] Specifically, in this embodiment, FHSS frequency hopping spread spectrum technology is used to select a frequency with good signal quality for transmission, and high-priority data packets are protected by Reed-Solution coding for error correction. The code length is n=255, and the information bits are RS codes with k=223.

[0077] Low-priority data packets use the ARQ automatic repeat request mechanism. After sending the data packet, the sender starts a timer to wait for the confirmation signal from the receiver.

[0078] Its beneficial effects are as follows: the user's motion status, head posture, and ambient noise are acquired through the sensor group in the wireless headset to obtain sensor data, and the audio data and sensor data are aligned by timestamp to obtain ambient audio matrix data; the noise spectrum characteristics in the ambient audio matrix data are extracted using an improved CNN convolutional neural network, and the motion friction noise in the matrix data is eliminated through an adaptive filtering algorithm to obtain dual-channel noise reduction data; the dual-channel noise reduction data is input into the BERT-ResNet fusion model to identify the user's current scene, and the available bandwidth within the next 5ms is predicted by combining the Bluetooth RSS I signal and historical channel data. The codec is switched based on the scene recognition result and the available bandwidth; the audio data stream is segmented based on the codec, and high-priority and low-priority data packets are transmitted through the Bluetooth main channel and auxiliary channel respectively. 1. It can accurately extract noise spectrum characteristics and eliminate motion friction noise. Compared with traditional noise reduction methods, the noise reduction effect is improved, effectively improving the purity of the audio and providing users with clearer sound quality. 2. Dynamically switches codecs based on the scenario and available bandwidth, improving audio transmission fluency in complex environments and matching audio quality with the current usage scenario and channel conditions, meeting the user's personalized needs in different scenarios. 3. Improves the success rate of audio data transmission, effectively reducing audio freezes and distortion, and enhancing the user's listening experience and user satisfaction.

[0079] See also Figure 2 In a Bluetooth-based wireless headset audio transmission method, aligning audio data and sensor data by timestamps to obtain ambient audio matrix data includes the following steps:

[0080] Step 201: Perform clock synchronization via the Bluetooth protocol, and use a bidirectional synchronization algorithm based on a reference timestamp to reduce clock deviations between sensor data and audio data to obtain synchronized data.

[0081] Step 202: Establish a timestamp mapping table, and perform interpolation processing on the sensor data in the synchronization data based on the timestamp of the audio data to obtain interpolated data;

[0082] Step 203: Estimate the sensor state at the intermediate moment using a linear interpolation method; perform time point alignment on the sensor data whose timestamp is later than the audio frame to obtain aligned data;

[0083] Step 204: Create ambient audio matrix data including audio data, motion state data, head posture data, and ambient noise data based on the aligned data. The matrix rows represent time series, and the matrix columns correspond to data dimensions.

[0084] See also Figure 3In a Bluetooth-based wireless headset audio transmission method, inputting dual-channel noise reduction data into a BERT-ResNet fusion model to identify the user's current scene includes the following steps:

[0085] Step 301: Use the residual block in the deep residual network ResNet-18 to extract the spatial features of the audio data and output a high-dimensional spatial feature vector;

[0086] Step 302: Process the audio sequence data using the pre-trained BERT model to capture the contextual dependencies of the audio signal in the time series and extract sequence features.

[0087] Step 303: Concatenate the spatial feature vector output by ResNet and the sequence feature vector output by BERT to obtain a fused feature vector containing spatial and temporal information.

[0088] Step 304: Input the fused feature vector into the fully connected layer for scene classification to identify the scene the user is currently in.

[0089] Specifically, the present invention can also be implemented through the following steps:

[0090] 1. Data Collection and Alignment

[0091] (1) Sensor group composition and data acquisition

[0092] The built-in sensor group of wireless headphones consists of multiple types of sensors working together to fully obtain user status and environmental information. Specifically, they include:

[0093] Motion and posture sensors: These include a three-axis accelerometer, a three-axis gyroscope, and a three-axis geomagnetic sensor. The accelerometer measures acceleration in three orthogonal directions during user movement, enabling identification of different motion states, such as stillness, walking, running, and jumping. The gyroscope senses the head's angular velocity along its three rotational axes in real time. Combined with the orientation information provided by the geomagnetic sensor, it accurately determines the head's pitch, yaw, and roll, including movements such as up and down, left and right, and side to side.

[0094] Environmental noise collection sensor: It uses a multi-microphone array, usually arranged on the surface of the earphone shell, to collect environmental noise at different angles and positions to ensure that it can capture surrounding sound signals such as traffic noise, human voices, wind noise, etc. in all directions.

[0095] Each sensor collects data at a unified sampling frequency, which is set to 44.1kHz, consistent with the common audio sampling frequency, facilitating subsequent data processing and synchronization.

[0096] (2) Timestamp alignment mechanism

[0097] To achieve accurate synchronization between audio data and sensor data, the following timestamp alignment method is used:

[0098] Timestamp generation: When collecting audio data and sensor data, the high-precision clock module built into the headset generates a timestamp for each frame of data. The timestamp is accurate to the microsecond level to ensure the accuracy of the time identification.

[0099] Clock synchronization: During the connection initialization phase, the headset and paired Bluetooth device (mobile phone, player, etc.) synchronize their clocks through the Bluetooth protocol. A two-way synchronization algorithm based on a reference timestamp is used to reduce the clock deviation between the two and ensure the consistency of the entire system time.

[0100] Data alignment: A timestamp mapping table is established, and sensor data is interpolated or resampled based on the audio data's timestamp. For sensor data with timestamps earlier than the current audio frame's timestamp, linear interpolation is used to estimate the sensor state at the intermediate moments. For sensor data with timestamps later than the audio frame's timestamp, alignment is performed until an appropriate time point. This ultimately forms an ambient audio matrix containing audio data, motion data, head posture data, and ambient noise data. The rows of the matrix represent time series, and the columns correspond to different types of data dimensions.

[0101] 2. Noise Cancellation Processing

[0102] (1) Improved CNN convolutional neural network

[0103] To extract the noise spectrum features of ambient audio matrix data, an improved CNN convolutional neural network is designed. Its network structure and improvements are as follows:

[0104] Network Architecture: This model employs a multi-layer structure alternating between convolutional and pooling layers, consisting of three convolutional layers, two max-pooling layers, and two fully connected layers. The convolutional layers use 3×3 convolution kernels to capture local spectral features, while the pooling layers use 2×2 windows for downsampling to reduce data dimensionality.

[0105] Improvement measures:

[0106] Residual connection: Residual connections are introduced between adjacent convolutional layers, that is, skip connections are added between the input and output, and the input is directly added to the output after convolution processing. This effectively solves the gradient vanishing problem in deep network training and enables the network to learn more complex noise spectrum characteristics.

[0107] Batch Normalization: Add a batch normalization layer after each convolutional layer and fully connected layer to normalize the data, accelerate network convergence, and improve the training stability and generalization ability of the model.

[0108] Activation function: LeakyReLU activation function is used instead of the traditional ReLU function to avoid the "death" of neurons in the negative interval and enhance the nonlinear expression ability of the network.

[0109] (2) Adaptive filtering algorithm

[0110] Based on the extracted noise spectrum characteristics, an adaptive least mean square (LMS) filtering algorithm is used to eliminate motion friction noise. The specific steps are as follows:

[0111] Noise reference signal generation: The ambient noise data collected by the microphone array is combined with the motion status and head posture information obtained by the sensor group to determine whether there is motion friction noise (noise generated by the friction between the headphones and the head) and generate a corresponding noise reference signal.

[0112] Filter Coefficient Adjustment: Based on the noise spectrum characteristics extracted by the improved CNN, the LMS filter coefficients are adjusted in real time. The filter adopts a transverse FIR structure. Noise Removal: The adjusted filter is applied to the ambient audio matrix data to remove motion friction noise through filtering, resulting in dual-channel (left and right channel) noise reduction data.

[0113] 3. Scene Recognition and Codec Switching

[0114] (1) BERT-ResNet fusion model

[0115] To accurately identify the user's current scenario, a BERT-ResNet fusion model was constructed. The fusion process is as follows:

[0116] Data preprocessing: The dual-channel denoised data is converted into a format suitable for both network inputs. For the ResNet component, the audio data for each channel is treated as a two-dimensional matrix (in the time-frequency domain) as the input to the ResNet. For the BERT component, the audio data is converted into a sequence format, and the audio features at each time point are used as a token and input into the BERT.

[0117] Feature extraction:

[0118] ResNet part: The deep residual network ResNet-18 is used to extract the spatial features of audio data (spatial distribution of the spectrum, energy concentration area, etc.) through multiple residual blocks, and output a high-dimensional spatial feature vector.

[0119] BERT part: Use the pre-trained BERT model to process audio sequence data, capture the contextual dependencies of audio signals in time series, and extract sequence features.

[0120] Feature fusion: The spatial feature vector output by ResNet and the sequence feature vector output by BERT are concatenated to form a fused feature vector containing spatial and temporal information. This vector is then input into the fully connected layer for scene classification to identify the user's current scene, such as indoor, outdoor, active, or quiet.

[0121] (2) Available bandwidth prediction

[0122] The Bluetooth RSSI signal and historical channel data are combined to predict the available bandwidth within the next 5ms. The specific method is as follows:

[0123] Bluetooth RSSI signal acquisition: The Bluetooth signal's received signal strength indicator (RSSI) is collected in real time at a frequency of 100 Hz, reflecting the signal strength and channel quality of the current Bluetooth link.

[0124] Historical channel data management: Establish a historical channel database to store channel parameters such as RSSI values, transmission rate, bit error rate, etc. within the past period of time (1 minute), and index them in chronological order.

[0125] Prediction Model: A long short-term memory (LSTM) network is used as the prediction model. It inputs the current RSSI value and historical channel data and outputs the available bandwidth within the next 5 ms. The LSTM model can capture the time series correlation of channel parameters and adapt to dynamic changes in the channel.

[0126] (3) Codec switching strategy

[0127] Different codec schemes are pre-set based on the scene recognition results and available bandwidth. The specific switching rules are as follows:

[0128] Scenario-Bandwidth Mapping Table: Establishes the correspondence between scenarios and available bandwidth, for example:

[0129] Quiet indoor scene: The available bandwidth is high (>2Mbps), and a high-quality codec (LDAC) is selected;

[0130] Sports: For moderate bandwidth (1-2 Mbps), choose a codec that balances sound quality and interference immunity (aptX Adaptive).

[0131] Complex outdoor scenarios: Available bandwidth is low (<1Mbps). Select a low-bitrate, high-reliability codec (SBC).

[0132] Real-time switching mechanism: When the scene recognition results change or the predicted available bandwidth exceeds the applicable range of the current codec, the codec switching process is triggered. During the switching process, the current audio transmission is first stopped, and the codec is reconfigured and negotiated to ensure the stability and compatibility of the switching.

[0133] 4. Data Sharding and Transmission

[0134] (1) Packet Prioritization

[0135] Based on the importance of the audio data and the real-time requirements, the audio data stream is divided into high-priority and low-priority data packets:

[0136] High-priority packets contain key audio information, such as the fundamental pitch frequency and formant parameters of speech. This data is crucial to audio intelligibility and quality. For music signals, high-priority packets include the main melody and harmony information in the low-frequency band.

[0137] Low-priority data packets: Contains secondary audio information, such as detailed sound effects in the high-frequency band and environmental reverberation. These data improve audio quality to a certain extent, but can be sacrificed appropriately when bandwidth is limited.

[0138] (2) Channel allocation and transmission strategy

[0139] Channel division: The Bluetooth link is divided into a primary channel and a secondary channel. The primary channel is used to transmit high-priority data packets. It adopts frequency hopping spread spectrum (FHSS) technology and selects frequencies with better signal quality for transmission to ensure transmission reliability. The secondary channel is used to transmit low-priority data packets. It adopts fixed frequency transmission to improve spectrum utilization.

[0140] Encoding and retransmission mechanism:

[0141] High-priority data packets use Reed-Solomon (RS) coding for error correction. The RS code, with a code length of n = 255 and information bits of k = 223, can correct up to 16 symbol errors. During transmission, RS-encoded redundant information is added to each high-priority data packet. The receiving end uses RS decoding to recover any erroneous data.

[0142] Low-priority packets: The Automatic Repeat Request (ARQ) mechanism is used. After sending a packet, the sender starts a timer to wait for an acknowledgment signal from the receiver. If no acknowledgment is received within the specified time, the packet is retransmitted up to three times to avoid delays caused by frequent retransmissions.

[0143] (3) Sharding strategy

[0144] The appropriate fragment size is determined based on the codec's output data rate and available bandwidth. The transmission interval is set to 1ms to ensure real-time fragmentation. The audio data stream is split into multiple packets based on the fragment size, and each packet is marked with a priority and sequence number to facilitate reassembly and processing at the receiving end.

[0145] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A Bluetooth-based wireless headset audio transmission method, characterized in that: The wireless headset audio transmission method comprises the following steps: The sensor group in the wireless headset obtains the user's motion status, head posture and environmental noise to obtain sensor data. The audio data and sensor data are aligned by time stamp to obtain environmental audio matrix data. The improved CNN convolutional neural network is used to extract the noise spectrum characteristics in the environmental audio matrix data, and the motion friction noise in the matrix data is eliminated by an adaptive filtering algorithm to obtain dual-channel noise reduction data; The dual-channel noise reduction data is input into the BERT-ResNet fusion model to identify the user's current scenario, combine the Bluetooth RSS I signal and historical channel data, predict the available bandwidth within the next 5ms, and switch the codec based on the scenario recognition result and the available bandwidth; The audio data stream is segmented based on the codec and divided into high-priority and low-priority data packets, which are transmitted through the Bluetooth primary channel and auxiliary channel respectively.

2. The Bluetooth-based wireless headset audio transmission method according to claim 1, wherein: The method of aligning the audio data and the sensor data by timestamps to obtain the ambient audio matrix data includes: Clock synchronization is performed through the Bluetooth protocol. A bidirectional synchronization algorithm based on a reference timestamp reduces the clock deviation of sensor data and audio data to obtain synchronized data. Establishing a timestamp mapping table, and performing interpolation processing on the sensor data in the synchronization data based on the timestamp of the audio data to obtain interpolated data; The linear interpolation method is used to estimate the sensor state at the intermediate moment; the sensor data whose timestamp is later than the audio frame is time-aligned to obtain aligned data; Based on the alignment data, environmental audio matrix data including audio data, motion state data, head posture data and environmental noise data is established, where the matrix rows represent time series and the matrix columns correspond to data dimensions.

3. The Bluetooth-based wireless headset audio transmission method according to claim 1, wherein: The method of extracting noise spectrum features from the ambient audio matrix data using an improved CNN convolutional neural network includes: The improved CNN convolutional neural network is based on a multi-layer convolutional layer and pooling layer alternating structure, including three convolutional layers, two maximum pooling layers and two fully connected layers. The convolutional layer uses a 3×3 convolution kernel to capture local spectral features, and the pooling layer uses a 2×2 window for downsampling. Add skip connections between the input and output of adjacent convolutional layers to add the input to the convolutional output; Add a batch normalization layer after each convolutional layer and fully connected layer to normalize the data, and use the LeakyReLU activation function to replace the ReLU activation function of the CNN convolutional neural network; An improved CNN convolutional neural network is used to extract noise spectrum features in the environmental audio matrix data to obtain noise spectrum features.

4. The Bluetooth-based wireless headset audio transmission method according to claim 2, wherein: The method of eliminating motion friction noise in the matrix data by an adaptive filtering algorithm to obtain dual-channel noise reduction data includes: Obtain the ambient audio matrix data, determine whether there is motion friction noise and generate the corresponding noise reference signal; The noise spectrum characteristics are obtained, the coefficients of the LMS filter are adjusted in real time, the adjusted filter is applied to the ambient audio matrix data, and the motion friction noise is removed through the adaptive filtering algorithm to obtain dual-channel noise reduction data.

5. The Bluetooth-based wireless headset audio transmission method according to claim 1, wherein: Inputting the dual-channel noise reduction data into the BERT-ResNet fusion model to identify the user's current scene includes: Use the residual block in the deep residual network ResNet-18 to extract the spatial features of audio data and output a high-dimensional spatial feature vector; The pre-trained BERT model is used to process audio sequence data, capture the contextual dependencies of audio signals in time series, and extract sequence features. The spatial feature vector output by ResNet and the sequence feature vector output by BERT are concatenated to obtain a fused feature vector containing spatial and temporal information; The fused feature vector is input into the fully connected layer for scene classification to identify the scene the user is currently in.

6. The Bluetooth-based wireless headset audio transmission method according to claim 1, wherein: The method combines the Bluetooth RSS I signal and historical channel data to predict the available bandwidth within the next 5 ms, and switches the codec based on the scene recognition result and the available bandwidth, including: Establish a correspondence between scenarios and available bandwidth. Choose a high-quality codec for quiet indoor scenarios, a codec that balances sound quality and anti-interference capabilities for sports scenarios, and a codec that prioritizes low bitrate and reliability for complex outdoor scenarios. When the scene recognition result changes and the predicted available bandwidth exceeds the applicable range of the current codec, the codec switching process is triggered.

7. The Bluetooth-based wireless headset audio transmission method according to claim 1, wherein: The codec-based segmentation method includes: dividing the audio data stream into high-priority and low-priority data packets, and transmitting the data packets through the Bluetooth primary channel and the auxiliary channel respectively. FHSS frequency hopping spread spectrum technology is used to select frequencies with good signal quality for transmission. High-priority data packets are protected by Reed-Solomon coding, with a code length of n=255 and an RS code with k=223 information bits. Low-priority data packets use the ARQ automatic repeat request mechanism. After sending the data packet, the sender starts a timer to wait for the confirmation signal from the receiver.

8. A system for implementing the Bluetooth-based wireless headset audio transmission method as claimed in claim 1, characterized in that: The system includes the following modules: The matrix data acquisition module is used to obtain the user's motion status, head posture and environmental noise through the sensor group in the wireless headset to obtain sensor data, and align the audio data and sensor data through timestamps to obtain environmental audio matrix data; A matrix data denoising module is used to extract noise spectrum features from the ambient audio matrix data using an improved CNN convolutional neural network, eliminate motion friction noise from the matrix data using an adaptive filtering algorithm, and obtain dual-channel denoised data; A decoder switching module is used to input the dual-channel noise reduction data into the BERT-ResNet fusion model to identify the user's current scenario, combine the Bluetooth RSSI signal and historical channel data, predict the available bandwidth within the next 5ms, and switch the codec based on the scenario recognition result and the available bandwidth; The audio channel transmission module is used to perform segmentation based on the codec, divide the audio data stream into high-priority and low-priority data packets, and transmit them through the Bluetooth main channel and auxiliary channel respectively.

9. A system for implementing the Bluetooth-based wireless headset audio transmission method as claimed in claim 1, characterized in that: The matrix data acquisition module includes the following submodules: The synchronization submodule is used to synchronize the clocks through the Bluetooth protocol. The bidirectional synchronization algorithm based on the reference timestamp reduces the clock deviation of the sensor data and audio data to obtain synchronized data. An interpolation submodule, configured to establish a timestamp mapping table, and perform interpolation processing on the sensor data in the synchronization data based on the timestamp of the audio data to obtain interpolated data; The alignment submodule is used to estimate the sensor state at the intermediate moments using a linear interpolation method; Perform time point alignment on sensor data whose timestamp is later than the audio frame to obtain aligned data; A submodule is established to establish environmental audio matrix data including audio data, motion state data, head posture data and environmental noise data based on the alignment data, where the matrix rows represent time series and the matrix columns correspond to data dimensions.

10. A system for implementing the Bluetooth-based wireless headset audio transmission method as claimed in claim 1, characterized in that: The system includes the following modules: Set up submodules to determine the improved CNN convolutional neural network structure based on alternating multiple convolutional layers and pooling layers, including 3 convolutional layers, 2 maximum pooling layers, and 2 fully connected layers. The convolutional layer uses a 3×3 convolution kernel to capture local spectral features, and the pooling layer uses a 2×2 window for downsampling. Add a submodule to add skip connections between the input and output of adjacent convolutional layers, adding the input to the convolutional output; The replacement submodule is used to add a batch normalization layer after each convolutional layer and fully connected layer to normalize the data and use the LeakyReLU activation function to replace the ReLU activation function of the CNN convolutional neural network; The extraction submodule is used to extract the noise spectrum features in the environmental audio matrix data using an improved CNN convolutional neural network to obtain the noise spectrum features.

Images