Online learning method for adaptive active noise reduction of earphone
By employing online learning methods for recursive denoising and multi-scale residual convolutional networks, combined with privacy protection protocols, the real-time adjustment and personalization issues of active denoising technology in complex environments are addressed, achieving an efficient and secure denoising experience.
Patent Information
- Application Number
- CN202510917364.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-11-11
AI Technical Summary
Existing active noise cancellation technologies struggle to adjust in real time in complex and ever-changing environments, failing to provide a personalized noise cancellation experience, and the issues of computing resources and privacy protection remain unresolved.
We employ a recursive denoising method combined with a multi-scale residual convolutional network and online learning. By optimizing the denoising model through a parameter extraction module and a meta-controller, we achieve personalized denoising and introduce a privacy protection protocol.
It achieves efficient and personalized noise reduction in complex noisy environments, reduces computational complexity, protects user privacy, and provides a more reassuring user experience.
Smart Images
Figure CN120935488A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of adaptive active noise cancellation technology for headphones, and in particular to an online learning method for adaptive active noise cancellation for headphones. Background Technology
[0002] With rapid urbanization and an increasingly fast-paced lifestyle, our daily environment is becoming increasingly noisy. Against this backdrop, headphones with active noise cancellation have gradually become an indispensable device for many people in their work and daily lives. However, despite years of development in active noise cancellation technology, existing technology still has many limitations and struggles to meet the growing needs of users.
[0003] Traditional active noise cancellation technology typically uses filters with fixed parameters to eliminate ambient noise. While this method performs reasonably well in stable noisy environments, its limitations become apparent in the complex and ever-changing real world. For example, when a user moves from a quiet office to a noisy street, fixed-parameter noise cancellation algorithms often struggle to adjust in time, significantly reducing the noise cancellation effect. Furthermore, different users have different ear canal structures and hearing characteristics, making it difficult for fixed-parameter methods to provide a personalized noise cancellation experience for each user.
[0004] In recent years, some researchers have attempted to incorporate machine learning techniques to improve noise cancellation. These methods typically involve training models offline in a laboratory environment and then deploying the trained models into headphones. While this approach improves the adaptability of noise cancellation to some extent, several problems remain. First, offline-trained models struggle to adapt to the various novel noise environments encountered in daily life. Second, these methods often require significant computational resources, making them difficult to run in real-time on headphone devices with strict power and size limitations. Finally, with increasing emphasis on data privacy, how to improve noise cancellation while protecting user privacy has become a pressing issue.
[0005] Furthermore, existing noise reduction technologies often focus on noise in specific frequency bands, such as low-frequency noise or the human voice frequency band. However, real-world noise is often a complex combination of multiple frequency bands and types. Effectively processing various types of noise simultaneously has become a significant challenge.
[0006] Faced with these challenges, the industry urgently needs a new technical solution that can achieve real-time, adaptive, and personalized noise reduction with limited computing resources, while also considering privacy protection. This invention was developed in response to this need. Summary of the Invention
[0007] This invention proposes an online learning method for adaptive active noise cancellation in headphones, aiming to solve many problems existing in the prior art. Specifically, this invention strives to achieve the following objectives: improve the adaptability and personalization of noise cancellation effects, reduce computational complexity to adapt to the limitations of headphone devices, and protect user privacy and security.
[0008] This invention proposes an online learning method for adaptive active noise cancellation in headphones, comprising the following steps:
[0009] S1. Use the parameter extraction module to extract the original signal features of the current headphones and match their dimensions to an acceptable size for the model input;
[0010] S2. The noise reduction model and the parameter extraction module are trained online, and the training parameters of the parameter extraction module and the noise reduction model are optimized by the meta-controller;
[0011] S3. Calculate personalized noise reduction parameters using the trained noise reduction model parameters and send them to the active noise reduction module to complete the noise reduction;
[0012] The parameter extraction module uses a recursive noise reduction method to segment the signal into frames and uses Mel-frequency cepstral coefficients (MFCC) to extract features. The formula for calculating MFCC is as follows:
[0013]
[0014] Where X(n) is the spectrum of the signal, N is the number of FFT points, and k is the index of the MFCC coefficients.
[0015] Preferably, the parameter extraction module uses a recursive method to extract features, including the following steps:
[0016] S11. Signal Acquisition: Acquire the original signal and reference signal during the noise reduction process;
[0017] S12. Signal framing: Divide the original signal and the reference signal into non-overlapping sub-segments;
[0018] S13. Energy Normalization: Calculate the average energy value of the reference signal as a normalization reference, divide the average energy value of the original signal by the average energy value of the reference signal, and input it into each frame of signal;
[0019] S14. Fourier Transform: Perform a Fast Fourier Transform on each frame of signal and take the modulus;
[0020] S15. Logarithmic operation: Take the logarithm of the Fourier transform result to obtain the signal spectrum;
[0021] S16. Calculate Mel frequency: Set the sampling frequency, number of filters, and center frequency of the filters according to the characteristics of the headphones, and calculate the Mel frequency of each frame signal;
[0022] S17. Trigonometric function operation and interpolation: Map the Mel frequency of each frame signal to the spectrum diagram, and then use the trigonometric function approximation method to interpolate the Mel cepstral coefficients of each frame signal;
[0023] S18. Feature Dimension Output: The Mel-frequency cepstral coefficients calculated from each frame of signal are concatenated, and a fixed-length vector is truncated according to the model requirements as the input feature of the model.
[0024] Preferably, the online training includes the following steps:
[0025] S21. Initialization: Load the meta-controller to adjust the controller's learning rate and learning decay rate, and initialize the parameters of the noise reduction model using random numbers;
[0026] S22. Data Acquisition: Play audio files from the training set through a speaker and acquire raw noise signals; while acquiring the raw noise signals, wear headphones on the ear and acquire raw noise signals and raw microphone signals from the ear canal.
[0027] S23. Calculate the loss function value: The input of the noise reduction model is the original microphone signal and the features extracted by the parameter extraction module, and the output is the noise reduction parameter corresponding to the microphone signal. Calculate the loss function value for the noise reduction parameter predicted by the noise reduction model and the noise reduction parameter corresponding to the original noise signal respectively.
[0028] S24. Adjust the meta-controller: Use the Adam optimization algorithm and gradient pruning technique to optimize the parameters of the meta-controller, including the learning rate and the learning decay rate;
[0029] S25. Adjust the parameter extraction module: with the optimization objective being to minimize the loss function value corresponding to the original noise signal, use the Adam optimization algorithm and gradient shearing technique to optimize the parameters of the parameter extraction module;
[0030] S26. Adjust the denoising model: with the optimization objective being to minimize the loss function value predicted by the denoising model, use the Adam optimization algorithm and gradient clipping and gradient shearing techniques to optimize the parameters of the denoising model;
[0031] S27. Update the learning rate and learning decay rate of the parameter extraction module using the meta-controller;
[0032] S28. Determine if all data has been used: If not all data has been used, proceed to step S22; otherwise, proceed to step S29.
[0033] S29. Load the meta controller, update the parameters of the noise reduction model, and obtain the noise reduction model exclusive to each user after training is completed;
[0034] S210. Generate personalized noise reduction parameters: After the original noise signal at the human ear is processed by the parameter extraction module, the features are input into the noise reduction model to obtain the personalized noise reduction parameters for each frequency of the user.
[0035] Preferably, the meta-controller updates the learning rate using the following formula:
[0036]
[0037] Where, η t Let η be the learning rate at time t, η0 be the initial learning rate, decay be the learning rate decay coefficient, and t be the current training step.
[0038] 5. The online learning method according to claim 3, characterized in that the loss function adopts the mean squared error (MSE), and the calculation formula is:
[0039]
[0040] Where n is the number of samples, y i For the true value, These are predicted values.
[0041] Preferably, the update formula of the Adam optimization algorithm is:
[0042]
[0043] Where, θ t Let be the parameters at time t, η be the learning rate, and m be the parameter. t For first-order momentum, v t Let be the second-order momentum, and ∈ be the smoothing term.
[0044] Preferably, the noise reduction model uses a multi-scale residual convolutional network, in which dilated convolution and pooling convolution are used simultaneously in the residual blocks, and the ReLU activation function is employed. The formula for calculating dilated convolution is as follows:
[0045] y[i]=∑ k x[i+r*k]*w[k]
[0046] Where y[i] is the output, x[i] is the input, w[k] is the convolution kernel weight, and r is the dilation rate.
[0047] Preferably, the calculation process of the multi-scale residual convolutional network is as follows:
[0048] S81. The input signal is the feature signal of each time point of 28 channels, which is input into dilated convolution and pooling convolution respectively. The input size is 28*1 and the output size is 14*256.
[0049] S82. The obtained multi-channel signals are merged into a new channel dimension signal with a size of 28*384 by splicing operation;
[0050] S83. Multi-scale dilated convolution: By using convolution kernels of different sizes and dilation ratios, feature signals with different receptive fields can be obtained. The relationship between the receptive field K and the convolution kernel size S is as follows:
[0051] K = S + (S-1)*(R-1)
[0052] Where R is the expansion rate;
[0053] S84. The output of pooling convolution is 28*16, which is concatenated with the features extracted by multi-channel convolution to obtain a 28*512 dimension feature.
[0054] S85. The output layer has 2048 convolutional kernels of size 1*1. After passing through the Flatten layer, it becomes a vector of 2048*1 dimensions. Then, after passing through the output layer, it regresses to obtain a real-time output of 2048 dimensions.
[0055] S86. Outputs noise estimation parameters in 2048 dimensions through the output layer.
[0056] Preferably, the meta-controller simultaneously optimizes the extracted features and the parameters of the denoising model, with optimization objectives being to minimize the loss function value after feature extraction and to minimize the loss function after the corresponding output of the denoising model, respectively. The optimization objective function of the meta-controller is:
[0057] L=λ1*L feature +λ2*L model
[0058] Where L is the total loss, L feature For feature extraction loss, L model λ1 and λ2 are the model prediction loss and weight coefficients, respectively.
[0059] Preferably, the method further includes the following steps:
[0060] S101. Set a privacy protection protocol during model training, wherein the privacy protection protocol uses gradient clipping technology to clip the uploaded gradient data;
[0061] S102. Set the pruning threshold to the loss function value of the parameter extraction module for extracting features from the original signal;
[0062] S103. Perform differential privacy processing on the clipped gradient data, where ε-differential privacy is defined as:
[0063] P(M(D)∈S)≤exp(∈)*P(M(D′)∈S)
[0064] Where M is a random algorithm, D and D′ are any two adjacent datasets, S is any subset of the output space, and ∈ is the privacy budget;
[0065] S104. Use the processed gradient data to update the model, ensuring privacy and security while completing model training.
[0066] The beneficial effects of this invention are mainly reflected in the following aspects:
[0067] The core of this invention lies in its innovative algorithm design, which cleverly solves several seemingly contradictory technical challenges. First, by introducing a parameter extraction module for recursive noise reduction and a multi-scale residual convolutional network, this invention can effectively capture the characteristics of various complex noises, achieving efficient noise reduction across the entire frequency band. This method not only improves the noise reduction effect but also greatly enhances the system's adaptability to novel noise types.
[0068] Secondly, this invention employs an online learning method based on a meta-controller. This method allows the system to be continuously optimized and adjusted during actual use, providing a personalized noise reduction experience for each user. Compared to traditional offline training methods, this online learning method significantly reduces the reliance on pre-collected large amounts of training data while also improving the system's real-time performance.
[0069] More notably, this invention cleverly resolves the conflict between computational efficiency and performance. Through a meticulously designed multi-scale residual convolutional network structure, this invention effectively controls computational complexity while enhancing the model's expressive power. This enables the method to run in real-time on headphone devices with limited computing resources, providing users with a smooth noise cancellation experience.
[0070] Regarding privacy protection, this invention introduces an innovative privacy protection protocol. By combining gradient clipping and differential privacy technology, this invention effectively prevents the leakage of user privacy information while ensuring noise reduction performance. This not only improves product security but also provides users with a more secure user experience.
[0071] In summary, this invention achieves multiple breakthroughs in noise reduction, adaptability, computational efficiency, and privacy protection through a combination of innovative technologies. These innovations are not simply additive but rather synergistically designed. For example, the multi-scale residual convolutional network not only improves noise reduction but also provides efficient feature extraction capabilities for online learning, thereby enhancing the overall system's adaptability. Simultaneously, the introduction of privacy protection protocols does not significantly impact system performance; instead, by increasing user trust, it indirectly promotes the acquisition of more diverse training data, further improving the overall system performance.
[0072] This comprehensive technological breakthrough enables the invention to provide users with an unprecedented noise cancellation experience. Whether in noisy public places or quiet private spaces, users can enjoy clear and natural audio. More importantly, with increased usage time, the system continuously learns and optimizes, providing users with increasingly personalized and precise noise cancellation effects. This not only greatly enhances the user's listening experience but also paves a new path for the future development of smart headphone technology. Attached Figure Description
[0073] Figure 1 This is a block diagram of the overall method logic of the present invention.
[0074] Figure 2 This is a logic block diagram of the parameter extraction module of the present invention.
[0075] Figure 3 This is a block diagram of the online training logic of the present invention.
[0076] Figure 4 This is a block diagram of the multi-scale residual convolutional network of the present invention.
[0077] Figure 5 This is a logic block diagram of the privacy protection protocol of the present invention. Detailed Implementation
[0078] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0079] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0080] Example 1
[0081] See Figure 1-5 This invention provides an online learning method for adaptive active noise cancellation in headphones. This method can adjust noise cancellation parameters in real time based on the user's environment and personal characteristics to provide optimal noise cancellation performance. The online learning method of this invention includes the following steps:
[0082] S1. First, the parameter extraction module extracts the raw signal features of the current headphones and matches their dimensions to an acceptable size for the model input. This step is crucial because it lays the foundation for subsequent noise reduction processing. Specifically, the parameter extraction module uses a recursive noise reduction method to segment the signal into frames and extracts features using Mel-frequency cepstral coefficients (MFCC). The MFCC is calculated using the following formula:
[0083]
[0084] Where X(n) is the signal spectrum, N is the number of FFT points, and k is the index of the MFCC coefficients. This method is effective because it simulates the auditory characteristics of the human ear and can more accurately capture the key features of sound.
[0085] S2. Next, the denoising model and parameter extraction module are trained online. This step uses a meta-controller to optimize the training parameters of the parameter extraction module and the denoising model. The advantage of online training is that it can adapt to changes in the environment in real time, providing more accurate denoising results.
[0086] S3. Finally, personalized noise reduction parameters are calculated using the trained noise reduction model parameters, and these parameters are sent to the active noise reduction module to complete the noise reduction. This personalized approach ensures that each user obtains the most suitable noise reduction effect.
[0087] Now, let's delve into how the parameter extraction module works. The parameter extraction module uses a recursive method to extract features, a process that involves multiple steps:
[0088] S11. The first step is signal acquisition, which involves acquiring the original signal and the reference signal during the noise reduction process.
[0089] S12. Then, signal framing is performed, dividing the original signal and the reference signal into non-overlapping segments. This framing process can capture the time-varying characteristics of the signal and improve the accuracy of feature extraction.
[0090] S13. The next step is energy normalization. In this step, we calculate the average energy value of the reference signal as a normalization reference, then divide the average energy value of the original signal by the average energy value of the reference signal, and apply the result to each frame of the signal. The purpose of this step is to eliminate the influence of different signal intensities on feature extraction.
[0091] S14. Subsequently, perform a fast Fourier transform on each frame of signal and take the modulus;
[0092] S15. Then take the logarithm of the transformation result to obtain the signal spectrum. These steps convert the time-domain signal into a frequency-domain representation, which facilitates subsequent feature extraction.
[0093] S16. When calculating the Mel frequency, we set the sampling frequency, the number of filters, and the filter center frequency according to the characteristics of the headphones. This targeted setting ensures that the extracted features better reflect the performance characteristics of the headphones.
[0094] S17. Map the Mel frequencies of each frame of signal to the spectrum diagram, and then interpolate the Mel cepstral coefficients of each frame of signal using a trigonometric function approximation method. Through trigonometric function operations and interpolation, we map the Mel frequencies of each frame of signal to the spectrum diagram, and then interpolate the Mel cepstral coefficients of each frame of signal using a trigonometric function approximation method. These coefficients constitute the input features of the model.
[0095] S18. Finally, the Mel-Cepstral coefficients calculated from each frame of signal are concatenated, and a vector of fixed length is truncated according to the model requirements as the input feature of the model.
[0096] Now, let's explore the online training process in detail. Online training involves several key steps:
[0097] S21. Initialization: Load the meta-controller to adjust the controller's learning rate and learning decay rate, and initialize the parameters of the denoising model using random numbers; this initialization method can provide a good starting point for the model and accelerate the convergence process.
[0098] S22. Data Acquisition: Play audio files from the training set through a speaker and acquire raw noise signals; while acquiring the raw noise signals, wear headphones on the ear and acquire raw noise signals and raw microphone signals from the ear canal; this real-world data acquisition ensures that the model can learn the noise characteristics of the actual usage environment.
[0099] S23. Calculate the loss function value: The input of the noise reduction model is the original microphone signal and the features extracted by the parameter extraction module, and the output is the noise reduction parameters corresponding to the microphone signal. The loss function value is calculated for the noise reduction parameters predicted by the noise reduction model and the noise reduction parameters corresponding to the original noise signal respectively. The purpose of this step is to evaluate the performance of the model and provide a basis for subsequent parameter adjustment.
[0100] S24. Adjust the meta-controller: Use the Adam optimization algorithm and gradient pruning technique to optimize the parameters of the meta-controller, including the learning rate and the learning decay rate;
[0101] S25. Adjust the parameter extraction module: with the optimization objective being to minimize the loss function value corresponding to the original noise signal, use the Adam optimization algorithm and gradient shearing technique to optimize the parameters of the parameter extraction module;
[0102] S26. Adjusting the denoising model: To optimize the denoising model by minimizing the predicted loss function value, the parameters are optimized using the Adam optimization algorithm and gradient clipping techniques. These adjustment steps utilize the Adam optimization algorithm and gradient clipping techniques to ensure the stability and efficiency of the training process. The optimization objective is to minimize the loss function value, thereby improving the model's performance.
[0103] S27. Update the learning rate and learning decay rate of the parameter extraction module using the meta-controller;
[0104] S28. Determine if all data has been used: If not all data has been used, proceed to step S22; otherwise, proceed to step S29.
[0105] S29. Load the meta controller, update the parameters of the noise reduction model, and obtain the noise reduction model exclusive to each user after training is completed;
[0106] S210. Generate personalized noise reduction parameters: After the original noise signal at the human ear is processed by the parameter extraction module, the features are input into the noise reduction model to obtain the personalized noise reduction parameters for each frequency of the user.
[0107] After completing one round of training, we use the meta-controller to update the learning rate and learning decay rate of the parameter extraction module. This dynamic adjustment mechanism enables the model to better adapt to changes during the training process.
[0108] Finally, we determine whether all data has been used. If not, we return to the data acquisition step to continue training; otherwise, we load the meta-controller, update the parameters of the denoising model, and obtain a denoising model specific to each user.
[0109] A key innovation of this invention is the use of a meta-controller. According to claim 4, the meta-controller updates the learning rate using the following formula:
[0110]
[0111] Where, η t Let ηt be the learning rate at time t, η0 be the initial learning rate, decay be the learning rate decay coefficient, and t be the current training step. This dynamic learning rate adjustment method can maintain a large learning rate in the early stages of training to accelerate convergence, while reducing the learning rate in the later stages of training to improve the model's accuracy.
[0112] In practical applications, we found that setting the initial learning rate η0 to 0.001 and the decay coefficient decay to 0.0001 resulted in good performance for the model in most cases. However, these parameters may need to be fine-tuned depending on the specific application scenario.
[0113] Another important innovation is the choice of loss function. This invention uses mean squared error (MSE) as the loss function, and its calculation formula is as follows:
[0114]
[0115] Where n is the number of samples, y i For the true value, The value is the predicted value. The advantage of MSE is that it can effectively measure the difference between the predicted value and the true value, and it is simple to calculate and the gradient is easy to solve.
[0116] In the optimization process, we use the Adam algorithm. The update formula for the Adam algorithm is:
[0117]
[0118] Where, θ t Let be the parameters at time t, η be the learning rate, and m be the parameter. t For first-order momentum, v t Let be the second-order momentum, and ∈ be the smoothing term. The Adam algorithm combines the advantages of the momentum method and RMSprop, achieving faster convergence and being less sensitive to the choice of hyperparameters.
[0119] Another innovation of this invention is the structure of the denoising model. The denoising model uses a multi-scale residual convolutional network, employing both dilated convolution and pooling convolution within the residual blocks, and adopting the ReLU activation function. The formula for calculating dilated convolution is as follows:
[0120]
[0121] Where y[i] is the output, x[i] is the input, w[k] is the kernel weight, and r is the dilation rate. The advantage of dilated convolution is that it can expand the receptive field without increasing the number of parameters, thereby capturing a wider range of contextual information.
[0122] Specifically, the calculation process for implementing a multi-scale residual convolutional network is as follows:
[0123] First, the input signal consists of feature signals from each of the 28 channels at various time points. These signals are then fed into dilated convolutions and pooling convolutions, with an input size of 28*1 and an output size of 14*256. Next, the resulting multi-channel signals are fused together using a concatenation operation to create a new channel-dimensional signal with a size of 28*384.
[0124] Next, we perform multi-scale dilated convolution. By using convolution kernels of different sizes and dilation ratios, we can obtain feature signals with different receptive fields. The relationship between the receptive field K and the convolution kernel size S is as follows:
[0125] K = S + (S-1)*(R-1)
[0126] Where R is the dilation rate. By adjusting the hole ratio and convolution kernel size of different combinations, we can adjust the receptive field to correspond to different distances between the human ear and the microphone in the training data, thereby extracting feature information from different locations of the signal.
[0127] In practical applications, we found that using three different dilation rates (1, 2, 4) can effectively capture features at different scales. Simultaneously, we used convolutional kernels of three sizes: 1x1, 3x3, and 5x5. This combination allows us to obtain rich feature information while maintaining computational efficiency.
[0128] The output of the pooling convolution is 28*16, which is concatenated with the features extracted by the multi-channel convolution to obtain a 28*512 dimensional feature. Finally, the output layer consists of 2048 1*1 convolutional kernels, which are transformed into 2048*1 dimensional vectors by the Flatten layer, and then passed through the output layer for regression to obtain a 2048-dimensional real-time output. This 2048-dimensional output is the noise estimation parameter.
[0129] Another innovation of this invention is the optimization objective function of the meta-controller. The meta-controller simultaneously optimizes the extracted features and the parameters of the denoising model, with optimization objectives of minimizing the loss function value after feature extraction and minimizing the loss function of the corresponding output of the denoising model, respectively. The optimization objective function of the meta-controller is:
[0130] L=λ1*L feature +λ2*L model
[0131] Where L is the total loss, L feature For feature extraction loss, L model The loss is set to the model prediction loss, and λ1 and λ2 are weighting coefficients. By adjusting the values of λ1 and λ2, we can balance the importance of feature extraction and model prediction. In practice, we found that the model achieves better overall performance when λ1 = 0.3 and λ2 = 0.7.
[0132] Finally, this invention also considers privacy protection. A privacy protection protocol is implemented during model training. This protocol uses gradient pruning to trim the uploaded gradient data, setting the pruning threshold as the loss function value used by the parameter extraction module to extract features from the original signal.
[0133] Furthermore, we apply differential privacy processing to the pruned gradient data. The definition of ε-differential privacy is:
[0134] P(M(D)∈S)≤exp(∈)*P(M(D′)∈S)
[0135] Where M is a randomized algorithm, D and D′ are any two adjacent datasets, S is any subset of the output space, and ∈ represents the privacy budget. In practical applications, we usually set ε between 0.1 and 1, and the specific value needs to be determined based on the balance between the strictness of privacy protection and model performance.
[0136] By combining these innovative features, this invention provides a highly efficient, adaptive, and secure headphone noise cancellation method. This method not only adjusts noise cancellation parameters in real time based on the user's individual characteristics and the usage environment, but also protects the user's privacy. In practical applications, this method can significantly improve the user's listening experience, especially in noisy environments. For example, when using headphones on public transportation or in an open-plan office, users can achieve better noise cancellation while maintaining audio clarity and naturalness.
[0137] In summary, the online learning method for adaptive active noise cancellation in headphones of this invention represents a significant advancement in noise cancellation technology. It not only improves noise cancellation effectiveness but also enhances the system's adaptability and security. This method is expected to be widely applied in high-end headphone products, providing users with a superior listening experience.
[0138] To verify the superiority of the online learning method for adaptive active noise cancellation in headphones proposed in this invention, we conducted a series of experiments. Here, we will focus on a preferred embodiment and a comparative example, demonstrating the innovation of this invention by comparing their performance.
[0139] The preferred embodiment employs the complete set of technologies proposed in this invention, including a parameter extraction module for recursive denoising, a denoising model based on a multi-scale residual convolutional network, an online learning method based on a meta-controller, and a privacy protection protocol. The comparative embodiment, however, uses a traditional fixed-parameter denoising method with a pre-trained model, without online learning or adaptive adjustment.
[0140] We selected several key metrics to evaluate the performance of the two methods: noise reduction effect, adaptability, computational efficiency, and privacy protection. These metrics are all closely related to the core innovations of this invention and can comprehensively reflect its superiority.
[0141] First, let's look at the noise reduction effect. We use Signal-to-Noise Ratio Improvement (SNRI) as the evaluation metric. The higher the SNRI, the better the noise reduction effect. We tested the performance of the two methods in different noisy environments (such as cafes, subways, and offices).
[0142] Secondly, there's adaptability. We designed a dynamic noise environment where the noise type and intensity change periodically. We used adaptation time to measure the speed at which the two methods respond to environmental changes; a shorter adaptation time is better.
[0143] In terms of computational efficiency, we measured the amount of audio data processed per second, in MB / s. This metric reflects the algorithm's real-time performance in practical applications.
[0144] Finally, to assess the level of privacy protection, we used the ε value of differential privacy as an indicator. The smaller the ε value, the higher the level of privacy protection.
[0145] Below is a detailed comparison table of the test results:
[0146] index Method of the present invention Traditional methods SNRI (cafe ambiance) 25.3dB 18.7dB SNRI (Metro Environment) 28.1dB 20.5dB SNRI (Office Environment) 22.6dB 17.2dB Adaptation time 0.5 seconds 5.2 seconds computational efficiency 12.8MB / s 15.3MB / s Privacy protection (ε value) 0.3 not applicable
[0147] From these test results, we can draw the following important conclusions:
[0148] First, in terms of noise reduction performance, the method of this invention significantly outperforms traditional methods in all test environments. Taking the subway environment as an example, the SNRI of this invention reached 28.1 dB, which is nearly 8 dB higher than that of traditional methods. This means that in noisy subway environments, users can obtain a clearer and more comfortable listening experience using our method. This advantage mainly stems from the multi-scale residual convolutional network and online learning method we employ, which can more accurately capture and eliminate various complex noise patterns.
[0149] Secondly, the method of this invention demonstrates excellent adaptability. Faced with dynamically changing noise environments, our method can adapt to new noise characteristics in just 0.5 seconds, while traditional methods require 5.2 seconds. This rapid adaptability is attributed to our online learning mechanism and the dynamic adjustment of the meta-controller, enabling the system to respond to environmental changes in real time and provide users with continuously optimized noise reduction effects.
[0150] In terms of computational efficiency, our method is slightly slower than traditional methods, processing 12.8 MB of audio data per second, while traditional methods can process 15.3 MB. This is mainly because our method involves more complex computational processes, including real-time feature extraction and model updates. However, considering the processing power of modern mobile devices, this level of computation is still perfectly feasible in real time and will not affect the user experience. Moreover, the significant performance improvement brought by our method is well worth the slightly increased computational cost.
[0151] Finally, regarding privacy protection, our method introduces a differential privacy mechanism with an ε value of 0.3, providing a high level of privacy protection. In contrast, traditional methods completely disregard privacy protection. In today's context of increasing emphasis on data privacy, this feature will undoubtedly enhance product competitiveness and user trust.
[0152] In summary, these test results fully demonstrate the superiority of our invention. Our method significantly improves the core noise reduction effect and adaptability, while also taking into account the important issue of privacy protection. Although there is a slight sacrifice in computational efficiency, this sacrifice is completely acceptable considering the overall performance improvement.
[0153] These results not only validate the effectiveness of our proposed technical solution but also demonstrate the immense potential of this method in practical applications. It can provide users with a superior, smarter, and safer noise cancellation experience, making it particularly suitable for use in the complex and ever-changing noise environments of modern life. We believe this technology will open up new possibilities for next-generation smart headphone products, bringing users an unprecedented listening experience.
[0154] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An online learning method for adaptive active noise cancellation in headphones, characterized in that... Includes the following steps: S1. Use the parameter extraction module to extract the original signal features of the current headphones and match their dimensions to an acceptable size for the model input; S2. The noise reduction model and the parameter extraction module are trained online, and the training parameters of the parameter extraction module and the noise reduction model are optimized by the meta-controller; S3. Calculate personalized noise reduction parameters using the trained noise reduction model parameters and send them to the active noise reduction module to complete the noise reduction; The parameter extraction module uses a recursive noise reduction method to segment the signal into frames and uses Mel-frequency cepstral coefficients (MFCC) to extract features. The formula for calculating MFCC is as follows: Where X(n) is the spectrum of the signal, N is the number of FFT points, and k is the index of the MFCC coefficients.
2. The online learning method according to claim 1, characterized in that... The parameter extraction module uses a recursive method to extract features, including the following steps: S11. Signal Acquisition: Acquire the original signal and reference signal during the noise reduction process; S12. Signal framing: Divide the original signal and the reference signal into non-overlapping sub-segments; S13. Energy Normalization: Calculate the average energy value of the reference signal as a normalization reference, divide the average energy value of the original signal by the average energy value of the reference signal, and input it into each frame of signal; S14. Fourier Transform: Perform a Fast Fourier Transform on each frame of signal and take the modulus; S15. Logarithmic operation: Take the logarithm of the Fourier transform result to obtain the signal spectrum; S16. Calculate Mel frequency: Set the sampling frequency, number of filters, and center frequency of the filters according to the characteristics of the headphones, and calculate the Mel frequency of each frame signal; S17. Trigonometric function operation and interpolation: Map the Mel frequency of each frame signal to the spectrum diagram, and then use the trigonometric function approximation method to interpolate the Mel cepstral coefficients of each frame signal; S18. Feature Dimension Output: The Mel-frequency cepstral coefficients calculated from each frame of signal are concatenated, and a fixed-length vector is truncated according to the model requirements as the input feature of the model.
3. The online learning method according to claim 1 or 2, characterized in that... The online training includes the following steps: S21. Initialization: Load the meta-controller to adjust the controller's learning rate and learning decay rate, and initialize the parameters of the noise reduction model using random numbers; S22. Data Acquisition: Play audio files from the training set through a speaker and acquire raw noise signals; while acquiring the raw noise signals, wear headphones on the ear and acquire raw noise signals and raw microphone signals from the ear canal. S23. Calculate the loss function value: The input of the noise reduction model is the original microphone signal and the features extracted by the parameter extraction module, and the output is the noise reduction parameter corresponding to the microphone signal. Calculate the loss function value for the noise reduction parameter predicted by the noise reduction model and the noise reduction parameter corresponding to the original noise signal respectively. S24. Adjust the meta-controller: Use the Adam optimization algorithm and gradient pruning technique to optimize the parameters of the meta-controller, including the learning rate and the learning decay rate; S25. Adjust the parameter extraction module: with the optimization objective being to minimize the loss function value corresponding to the original noise signal, use the Adam optimization algorithm and gradient shearing technique to optimize the parameters of the parameter extraction module; S26. Adjust the denoising model: with the optimization objective being to minimize the loss function value predicted by the denoising model, use the Adam optimization algorithm and gradient clipping and gradient shearing techniques to optimize the parameters of the denoising model; S27. Update the learning rate and learning decay rate of the parameter extraction module using the meta-controller; S28. Determine if all data has been used: If not all data has been used, proceed to step S22; otherwise, proceed to step S29. S29. Load the meta controller, update the parameters of the noise reduction model, and obtain the noise reduction model exclusive to each user after training is completed; S210. Generate personalized noise reduction parameters: After the original noise signal at the human ear is processed by the parameter extraction module, the features are input into the noise reduction model to obtain the personalized noise reduction parameters for each frequency of the user.
4. The online learning method according to claim 3, characterized in that... The meta-controller updates the learning rate using the following formula: Where, η t Let η be the learning rate at time t, η0 be the initial learning rate, decay be the learning rate decay coefficient, and t be the current training step.
5. The online learning method according to claim 3, characterized in that... The loss function uses the mean squared error (MSE), and the calculation formula is as follows: Where n is the number of samples, y i For the true value, These are predicted values.
6. The online learning method according to claim 3, characterized in that... The update formula for the Adam optimization algorithm is: Where, θ t Let be the parameters at time t, η be the learning rate, and m be the parameter. t For first-order momentum, v t Let be the second-order momentum, and ∈ be the smoothing term.
7. The online learning method according to claim 1, characterized in that... The denoising model uses a multi-scale residual convolutional network, employing both dilated convolution and pooling convolution within the residual blocks, and adopting the ReLU activation function. The formula for calculating dilated convolution is as follows: y[i]=∑ k x[i+r*k]*w[k] Where y[i] is the output, x[i] is the input, w[k] is the convolution kernel weight, and r is the dilation rate.
8. The online learning method according to claim 7, characterized in that... The computation process of the multi-scale residual convolutional network is as follows: S81. The input signal is the feature signal of each time point of 28 channels, which is input into dilated convolution and pooling convolution respectively. The input size is 28*1 and the output size is 14*256. S82. The obtained multi-channel signals are merged into a new channel dimension signal with a size of 28*384 by splicing operation; S83. Multi-scale dilated convolution: By using convolution kernels of different sizes and dilation ratios, feature signals with different receptive fields can be obtained. The relationship between the receptive field K and the convolution kernel size S is as follows: K = S + (S-1)*(R-1) Where R is the expansion rate; S84. The output of pooling convolution is 28*16, which is concatenated with the features extracted by multi-channel convolution to obtain a 28*512 dimension feature. S85. The output layer has 2048 convolutional kernels of size 1*1. After passing through the Flatten layer, it becomes a vector of 2048*1 dimensions. Then, after passing through the output layer, it regresses to obtain a real-time output of 2048 dimensions. S86. Outputs noise estimation parameters in 2048 dimensions through the output layer.
9. The online learning method according to claim 1, characterized in that... The meta-controller simultaneously optimizes the extracted features and the parameters of the denoising model, with optimization objectives of minimizing the loss function value after feature extraction and minimizing the loss function after the corresponding output of the denoising model. The optimization objective function of the meta-controller is: L=λ1*L feature +λ2*L model Where L is the total loss, L feature For feature extraction loss, L model λ1 and λ2 are the model prediction loss and weight coefficients, respectively.
10. The online learning method according to claim 1, characterized in that... The method further includes the following steps: S101. Set a privacy protection protocol during model training, wherein the privacy protection protocol uses gradient clipping technology to clip the uploaded gradient data; S102. Set the pruning threshold to the loss function value of the parameter extraction module for extracting features from the original signal; S103. Perform differential privacy processing on the clipped gradient data, where ε-differential privacy is defined as: P(M(D)∈S)≤exp(∈)*P(M(D′)∈S) Where M is a random algorithm, D and D′ are any two adjacent datasets, S is any subset of the output space, and ∈ is the privacy budget; S104. Use the processed gradient data to update the model, ensuring privacy and security while completing model training.