A denoising system and method based on echo state network adaptive retraining

Abstract

The application discloses a noise reduction system and method based on echo state network adaptive retraining; the system comprises a feedforward microphone, a feedback microphone, a loudspeaker and a processor; the processor filters environmental noise by using a neural network filter containing an echo state network, and outputs an anti-phase noise signal, which is superimposed with residual noise in the ear collected by the feedback microphone, and generates a noise cancellation signal after adaptive gain adjustment. When the detection of the earphone wearing mode or the acoustic leakage amount changes, online retraining is triggered: training samples are constructed based on the buffered feedforward environmental noise and feedback residual noise, the output weight of the ESN is updated by using the recursive least square method, and the noise reduction filter is kept uninterrupted during the training; the retraining driven by the mixed error of the feedforward and feedback signals, the fast-converging recursive least square method and the adaptive gain control significantly improve the noise reduction accuracy, speed and system stability under the change of the acoustic environment.

Description

Technical Field

[0001] This invention relates to the field of active noise reduction technology, specifically to a noise reduction system and method based on adaptive retraining of an echo state network. Background Technology

[0002] Active noise cancellation (ANC) technology cancels noise by generating sound waves with the opposite phase to the noise through a speaker, and is widely used in devices such as headphones. Feedforward, feedback, and hybrid are common ANC architectures. As application scenarios become more complex, traditional noise reduction schemes based on linear filters struggle to accurately model nonlinear acoustic paths, limiting the noise reduction depth and robustness.

[0003] Therefore, noise reduction using neural network filters, especially echo state networks (ESNs), has become a research hotspot. ESNs have the characteristics of large-scale sparse recursive reservoir structure and only need to train output weights. They have both strong nonlinear modeling capabilities and low training complexity, making them suitable for embedded ANC implementation.

[0004] However, in practical use, changes in headphone wearing methods (such as loosening or displacement) or alterations in acoustic leakage can lead to significant deviations in the secondary path transfer function from the speaker to the ear and the noise transmission path, resulting in a sharp decline in noise reduction performance based on the factory-pre-trained model. To address this issue, existing technologies have proposed several adaptive adjustment schemes.

[0005] An existing solution (such as Chinese patent CN116095553B) discloses using an ESN to filter noise collected by a feedforward microphone and triggering ESN retraining when a change in wearing conditions is detected. However, when performing online retraining, this solution only uses the feedforward ambient noise signal as the training input and does not fully integrate the residual noise signal in the ear collected by the feedback microphone. This results in the optimization target of retraining not being able to directly and accurately reflect the actual noise reduction error at the ear, and the compensation accuracy for changes in leakage state is limited, and the convergence process is slow. At the same time, the lack of adaptive gain adjustment involving feedback signals may lead to system instability or overshooting during the retraining process.

[0006] Another existing approach uses a neural network-based feedforward filter, but it only uses a fixed-training model and lacks the ability to adapt to changes in wearing status online. Furthermore, existing retraining methods mostly employ traditional gradient descent algorithms, which struggle to meet real-time noise reduction requirements in the face of rapidly changing acoustic environments due to their limited convergence speed and tracking capabilities. Summary of the Invention

[0007] This invention aims to at least partially solve one of the technical problems in related technologies. To this end, one object of this invention is to propose a noise reduction system based on adaptive retraining of an echo-state network, comprising: A feedforward microphone is used to collect ambient noise signals; Feedback microphone, used to collect residual noise signals in the ear; A loudspeaker, used to play noise cancellation signals; The processor is configured to: acquire the ambient noise signal collected by the feedforward microphone and the residual noise signal in the ear collected by the feedback microphone; The environmental noise signal is filtered using a trained neural network filter containing an echo state network to output a filtered inverse noise signal. The echo state network includes an input layer, a reservoir, and an output layer. The reservoir is composed of a large number of sparsely connected internal neurons. The input weights and internal connection weights of the reservoir remain unchanged after initialization, and only the output weights of the output layer are determined through training. The inverted noise signal is superimposed with the residual noise signal in the ear collected by the feedback microphone, and an adaptive gain adjustment is applied to generate the noise cancellation signal, so that the noise cancellation signal is played through the speaker to cancel the noise in the ear of the earphone; Under the condition of meeting the preset retraining triggering conditions, the echo state network is retrained online, wherein the preset retraining triggering conditions include at least detecting a change in the way the headphones are worn or a change in the amount of acoustic leakage. The online retraining includes: constructing a training sample set from the environmental noise signal and the residual noise signal in the ear within a preset time window before the current moment; and updating the output weights of the echo state network using the recursive least squares method based on the training sample set to redetermine the output weights of the echo state network.

[0008] Preferably, the online retraining further includes: using the ambient noise signal as the training input of the echo state network, using the error signal between the ambient noise signal and the residual noise signal in the ear as the training target, and using recursive least squares to minimize the following cost function: in, The output weight matrix of the echo state network is... The desired signal is determined based on the ambient noise signal and the residual noise signal in the ear. Let be the state vector of the storage pool at time step t. Forgetting factor, This is the length of the training window.

[0009] Preferably, the preset retraining trigger condition further includes at least one of the following: the noise reduction performance index is lower than a preset threshold, a user manual trigger command is received, a significant change in the environmental noise characteristics is detected, and the earphone is taken out of the charging case or powered on.

[0010] Preferably, the detection of changes in headphone wearing method or acoustic leakage is achieved through at least one of the following methods: The change in the signal transfer function between the feedforward microphone and the feedback microphone is monitored. When the change exceeds a first preset threshold, it is determined that the wearing method has changed. Alternatively, the low-frequency energy value of the signal collected by the feedback microphone can be monitored. When the change in the low-frequency energy value exceeds a second preset threshold, it is determined that the acoustic leakage has changed.

[0011] Preferably, the processor is further configured to: during the online retraining of the echo state network, maintain the filtering processing of the ambient noise signal by the neural network filter until the retraining is completed and the current output weights are replaced by the updated echo state network output weights, thereby achieving seamless switching.

[0012] Preferably, the adaptive gain adjustment is dynamically adjusted based on the energy level of the residual noise signal in the ear collected by the feedback microphone. When the energy level of the residual noise signal in the ear is lower than a preset lower threshold, the gain is reduced, and when it is higher than a preset upper threshold, the gain is increased.

[0013] Preferably, the processor further includes a secondary path compensation module for processing the secondary path transfer function of the noise cancellation signal from the speaker to the feedback microphone. Estimation is performed, and during the online retraining process, the following is utilized: The environmental noise signal is filtered and compensated before being used as the training input for the echo state network.

[0014] A noise reduction method based on adaptive retraining of an echo-state network is applied to the aforementioned noise reduction system based on adaptive retraining of an echo-state network. The method includes: The ambient noise signal is acquired through the feedforward microphone, and the residual noise signal in the ear is acquired through the feedback microphone. The environmental noise signal is filtered using a trained neural network filter containing an echo state network to output a filtered inverse noise signal. The echo state network includes an input layer, a reservoir, and an output layer. The input weights and internal connection weights of the reservoir remain unchanged after initialization, and only the output weights are determined through training. The inverted noise signal is superimposed with the residual noise signal in the ear and adaptive gain adjustment is applied to generate a noise cancellation signal, and the speaker is driven to play the noise cancellation signal to cancel the noise in the ear. Under the condition of meeting the preset retraining triggering conditions, the echo state network is retrained online. The preset retraining triggering conditions include at least detecting a change in the way the headphones are worn or a change in the amount of acoustic leakage. The online retraining includes: using the environmental noise signal and the residual noise signal in the ear within a preset time window before the current moment to form a training sample set, and using the recursive least squares method to update the output weights of the echo state network.

[0015] Preferably, the online retraining specifically includes: S1. Cache the environmental noise signal within the preset time window. and the residual noise signal in the ear This constitutes the training input sequence and the target error sequence; S2. Input the training input sequence frame by frame into the reservoir of the echo state network, and collect the state vector of the reservoir at each time step. Construct the state matrix ; S3. Using the target error sequence as the desired output, solve the updated output weight matrix using the recursive least squares method with a forgetting factor. : in, For the forgetting factor A diagonal weighted matrix with base 1. The regularization coefficient is . It is the identity matrix. The desired output matrix is ​​composed of the target error sequence; S4. The output weight matrix obtained by solving... Update the output layer of the echo state network to complete online retraining.

[0016] Preferred options also include: After the online retraining is completed, the updated echo state network is used to perform real-time inference on the subsequent input environmental noise signals, and the noise reduction performance indicators are continuously monitored. When the noise reduction performance index is detected to be lower than the preset performance threshold, or when a new wearing method or acoustic leakage is detected, the online retraining is triggered again to form a closed-loop adaptive noise reduction control loop.

[0017] The above-described solution of the present invention has at least the following beneficial effects: The collaborative feedforward and feedback signals are used for mixed error-driven retraining, which greatly improves the adaptive accuracy. During online retraining, the environmental noise signal in the current time window and the error signal determined by the feedforward and feedback signals are used as the training input and training target, respectively. This allows the update of the echo state network output weights to be directly based on minimizing residual noise in the ear. This can accurately compensate for acoustic path shifts caused by changes in wearing method and leakage. The noise reduction after retraining is higher and the acoustic adaptability is stronger. The recursive least squares method is used to update the output weights of the echo state network, which has fast convergence speed and strong real-time performance. The recursive least squares algorithm combined with the forgetting factor can quickly track changes in the acoustic environment, and the convergence speed is significantly improved compared with the traditional gradient descent method. It is very suitable for headphone embedded platforms to complete fast online retraining under limited computing resources. Online retraining and normal noise reduction filtering are executed in parallel, achieving seamless switching that is imperceptible to the user. During the online update of output weights, the processor continuously filters and reduces the noise of the current environment until the retraining is completed before applying the updated weights. This ensures that the noise reduction function is uninterrupted and does not generate sudden noise during the entire retraining process, thus guaranteeing the continuity of the user experience. An adaptive gain adjustment based on feedback signals is introduced to enhance system stability and auditory safety. By applying adaptive gain control based on residual noise energy in the ear to the superimposed inverse noise signal, overshoot and howling can be prevented in the early stages of retraining and under conditions of sudden changes in the acoustic environment, thereby improving the robustness of the noise reduction system and protecting the human ear from accidental high sound pressure impacts. The intelligent triggering mechanism forms a closed-loop adaptive control, maintaining good noise reduction performance over a long period of time. The system automatically monitors preset conditions such as wearing status and leakage changes. When the noise reduction performance shows a deteriorating trend, it immediately triggers online retraining. After training is completed, it continues to monitor, forming a closed-loop control loop of "monitoring-retraining-application-remonitoring", so that the headphones always maintain the best noise reduction effect throughout the entire usage cycle. Furthermore, by incorporating secondary path compensation in the preferred scheme, interference from the electroacoustic path from the speaker to the feedback microphone on the training target can be eliminated, making online retraining more accurate and further improving noise reduction performance in complex acoustic environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0019] Figure 1This is a schematic diagram of the overall structure and signal flow of the noise reduction system based on adaptive retraining of echo state network provided in this embodiment of the invention; Figure 2 This is a flowchart of the echo state network neural network filter provided in the embodiments of the present invention.

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "circumferential," and "radial," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0023] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0025] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0026] The following describes in detail, with reference to the accompanying drawings, a noise reduction system and method based on adaptive retraining of an echo state network according to an embodiment of the present invention.

[0027] Its hardware composition is as follows Figure 1 As shown, it mainly includes: a feedforward microphone, a feedback microphone, a speaker, and a processor. The processor is typically a digital signal processor (DSP) or an embedded microprocessor, internally implementing: a neural network filter module, a superposition and adaptive gain adjustment module, a retraining trigger detection module, an online retraining module, and an optional secondary path compensation module.

[0028] Design and Working Mechanism of Echo State Network (ESN) Neural Network Filters Echo-state networks are a type of recurrent neural network whose core is a randomly generated and fixed pool of neurons with a large number of sparse connections; for example... Figure 2 In this embodiment, the ESN includes an input layer, a storage pool, and an output layer; the input layer receives the ambient noise signal. By inputting the weight matrix Mapped to the reserve pool One neuron; neurons within the reservoir are connected via a connection weight matrix. The dynamic memory is formed, and its state update equation is: in, For time steps The state vector of the reserve pool, The activation function is a non-linear function, typically the hyperbolic tangent; the input weights are... and internal join weights During initialization, data is randomly generated according to a fixed distribution; scaling is then applied to adjust the distribution. The spectral radius is less than 1 to ensure echo state characteristics, and both remain permanently unchanged after initialization; the output layer linearly maps the reservoir state to a one-dimensional inverted noise signal. in To output the weight matrix, This is a bias term and can be considered as part of the output weights; It is the only parameter in ESN that is learned through training, which greatly reduces the training complexity.

[0029] Collaborative noise reduction and adaptive gain adjustment of feedforward and feedback signals Under normal noise reduction conditions, the system utilizes both feedforward and feedback microphone signals in a coordinated manner. The feedforward microphone collects the ambient noise signal. Inputting the ESN neural network filter, an inverse noise signal is obtained through inference. Meanwhile, the feedback microphone picks up residual noise signals in the ear in real time. The processor will and The following processing is performed by combining superposition: in A preset feedback fusion coefficient (e.g., 0.1–0.3) is used to introduce partial feedback signal to help suppress narrowband noise or compensate for phase deviation. Then, the signals are superimposed. Gain is applied via adaptive gain adjustment module This generates the final noise cancellation signal. And drive the speaker to play, gain Energy level based on feedback signal Dynamic adjustment, the specific rules are as follows: when Below the preset lower threshold When the current leakage is small and the noise reduction depth is large, the gain is gradually reduced to prevent high-frequency noise generated by over-cancellation; when Exceeding the preset upper limit threshold When this occurs, it indicates increased leakage and insufficient noise reduction; the gain should be appropriately increased to compensate. When between two levels, the gain remains at 1 or changes gradually; this gain adjustment effectively improves system stability and listening experience during sudden changes in the acoustic environment.

[0030] Preset trigger condition detection for retraining To achieve all-weather adaptive noise reduction, the processor continuously monitors various preset conditions to determine whether to initiate online retraining. The main conditions include: Headphone wearing method change detection: This is achieved by monitoring the change in the electroacoustic transfer function from the feedforward microphone to the feedback microphone. The processor uses the downlink pilot signal or ambient noise itself as excitation to estimate the impulse response or its low-frequency energy distribution along the path in real time. If the Euclidean distance or normalized mean square error between the current estimate and the reference model exceeds a first threshold... If so, it is determined that there has been a significant change in the wearing experience (such as the headphones moving or the seal changing).

[0031] Acoustic leakage change detection: By monitoring the energy of specific low-frequency bands (e.g., 20Hz-200Hz) of the feedback microphone signal Achieve; because the low-frequency band is extremely sensitive to leakage, when Exceeding the second threshold At that time, it was determined that the leakage amount had changed significantly.

[0032] Noise reduction performance degradation monitoring: During quiet intervals or by estimating the current noise reduction using the statistical characteristics of environmental noise, if the noise reduction index is lower than the preset performance threshold, retraining is triggered.

[0033] Triggered by a specific event: When a user issues a manual trigger command via touch or the app, or when the earbuds are removed from the charging case and powered on, the system forces an offline or online training session to update the model.

[0034] The fulfillment of any of the above conditions can trigger the retraining process independently. This multi-source triggering mechanism ensures timely capture of acoustic mutations.

[0035] Online retraining process When the preset trigger conditions are met, the processor initiates an online retraining process in the background, while the noise reduction filtering in the foreground is not interrupted; the specific steps are as follows: I. Training Data Acquisition and Caching The processor allocates a circular buffer to continuously cache the most recently used data during noise reduction operations. Ambient noise signal within a time window (e.g., 2 seconds) and the corresponding residual noise signal in the ear and the state vector of the reserve pool at the same moment. Once retraining is triggered, these data are immediately retrieved from the buffer as the training sample set. This approach eliminates the time required to reset the reserve pool and recalculate the state, fully utilizing the system's parallel recording capabilities to achieve rapid data readiness.

[0036] II. Hybrid Construction of Expected Signals (Reflecting the Synergy of Feedforward and Feedback) The core of this implementation lies in the fact that the training target integrates feedforward and feedback signals. In a preferred embodiment with secondary path compensation, the transfer function of the pre-identified secondary path is used to estimate... (The path from the speaker to the feedback microphone) filters and compensates for the ambient noise signal to obtain the estimated main noise component in the ear: Then the expected output signal at that moment Constructed by combining the main noise component with the current residual error in the ear: Its physical meaning lies in: the desired output of ESN via secondary path It can then approach This results in residual earwax. Minimize. As a training objective, it directly reflects the synergy between the feedforward environmental signal and the feedback error signal. In a simplified scheme without secondary path compensation, it can also be directly set... It participates directly in RLS updates as a reserve pool, but its performance will be lower than the above-mentioned filter-X type scheme.

[0037] III. Constructing the training matrix and solving using the recursive least squares method The state vector of the reserve pool at each moment within the time window Through secondary path estimation Perform filtering (using) (By filtering the trajectory of each neuron), we obtain the filtered state vector. Arrange them into a matrix: The corresponding expected output is: Introducing the forgetting factor Construct a diagonal weighted matrix To increase the focus on recent samples, a regularized recursive least squares method is used to solve for the new output weights in batches. : in The regularization coefficient (e.g., 0.01) is used to avoid matrix non-invertibility. Compared to stochastic gradient descent, RLS has a quadratic convergence speed and can approach the optimal solution in a very small number of iterations, making it very suitable for embedded headphone platforms with limited computing resources.

[0038] IV. Seamless Switching and Online Updates

[0039] After the update calculation is complete, the processor runs in the background to the output layer of the ESN, replacing the old output weights. During this process, the foreground noise reduction and filtering inference pipeline always runs using the old weights, without interruption or signal glitches. The replacement is smoothed through windowing (such as brief crossfading) to further avoid audible tones; at this point, a complete online retraining is completed.

[0040] Methods for obtaining secondary path estimates Secondary paths An average model under typical wearing conditions can be obtained through offline measurements before shipment, or initial online identification can be performed after the user wears the device. A simple method for online identification is as follows: in a quiet environment or while playing a specific noise signal, white noise / frequency sweep signal is played through a speaker, and feedback microphone signals are simultaneously acquired. The FIR filter coefficients are then fitted using the LMS algorithm. The model can then remain unchanged, or be coupled with slow online adaptive updates.

[0041] Closed-loop adaptive noise reduction control loop After online retraining is completed, the processor continues to monitor noise cancellation performance metrics and wearing status. When a new wearing change is detected again or the noise cancellation performance falls below the recovery threshold, a new round of online retraining is triggered. This forms a closed-loop adaptive control of "monitoring-triggering-application-remonitoring," enabling the headphones to autonomously and dynamically maintain optimal noise cancellation performance throughout the entire usage process, regardless of user adjustments to the fit or leakage changes caused by external forces. Simultaneously, the synergy of feedforward and feedback signals during training, the rapid convergence of the recursive least squares method, and the triple guarantee of adaptive gain significantly enhance the system's robustness and real-time performance in complex high sound pressure environments compared to existing technologies.

[0042] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0043] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A noise reduction system based on adaptive retraining of an echo state network, characterized in that, include: A feedforward microphone is used to collect ambient noise signals; Feedback microphone, used to collect residual noise signals in the ear; A loudspeaker, used to play noise cancellation signals; The processor is configured to: acquire the ambient noise signal collected by the feedforward microphone and the residual noise signal in the ear collected by the feedback microphone; The environmental noise signal is filtered using a trained neural network filter containing an echo state network to output a filtered inverse noise signal. The echo state network includes an input layer, a reservoir, and an output layer. The reservoir is composed of a large number of sparsely connected internal neurons. The input weights and internal connection weights of the reservoir remain unchanged after initialization, and only the output weights of the output layer are determined through training. The inverted noise signal is superimposed with the residual noise signal in the ear collected by the feedback microphone, and an adaptive gain adjustment is applied to generate the noise cancellation signal, so that the noise cancellation signal is played through the speaker to cancel the noise in the ear of the earphone; Under the condition of meeting the preset retraining triggering conditions, the echo state network is retrained online, wherein the preset retraining triggering conditions include at least detecting a change in the way the headphones are worn or a change in the amount of acoustic leakage. The online retraining includes: constructing a training sample set from the environmental noise signal and the residual noise signal in the ear within a preset time window before the current moment; and updating the output weights of the echo state network using the recursive least squares method based on the training sample set to redetermine the output weights of the echo state network.

2. The noise reduction system based on adaptive retraining of an echo state network according to claim 1, characterized in that, The online retraining further includes: using the ambient noise signal as the training input of the echo state network, using the error signal between the ambient noise signal and the residual noise signal in the ear as the training target, and using recursive least squares to minimize the following cost function: in, The output weight matrix of the echo state network is... The desired signal is determined based on the ambient noise signal and the residual noise signal in the ear. Let be the state vector of the reserve pool at time step t. Forgetting factor, This is the length of the training window.

3. The noise reduction system based on adaptive retraining of an echo state network according to claim 2, characterized in that, The preset retraining trigger conditions also include at least one of the following: the noise reduction performance index is lower than a preset threshold, a user manual trigger command is received, a significant change in the environmental noise characteristics is detected, and the earphone is taken out of the charging case or powered on.

4. The noise reduction system based on adaptive retraining of an echo state network according to claim 1, characterized in that, The detection of changes in headphone wearing method or acoustic leakage is achieved through at least one of the following methods: The change in the signal transfer function between the feedforward microphone and the feedback microphone is monitored. When the change exceeds a first preset threshold, it is determined that the wearing method has changed. Alternatively, the low-frequency energy value of the signal collected by the feedback microphone can be monitored. When the change in the low-frequency energy value exceeds a second preset threshold, it is determined that the acoustic leakage has changed.

5. A noise reduction system based on adaptive retraining of an echo state network according to claim 1, characterized in that, The processor is further configured to: during the online retraining of the echo state network, maintain the filtering of the ambient noise signal by the neural network filter until the retraining is completed, and replace the current output weights with the updated echo state network output weights to achieve seamless switching.

6. The noise reduction system based on adaptive retraining of an echo state network according to claim 1, characterized in that, The adaptive gain adjustment is dynamically adjusted based on the energy level of the residual noise signal in the ear collected by the feedback microphone. When the energy level of the residual noise signal in the ear is lower than a preset lower threshold, the gain is reduced, and when it is higher than a preset upper threshold, the gain is increased.

7. A noise reduction system based on adaptive retraining of an echo state network according to claim 1, characterized in that, The processor further includes a secondary path compensation module for processing the secondary path transfer function of the noise cancellation signal from the speaker to the feedback microphone. Estimation is performed, and during the online retraining process, the following is utilized: The environmental noise signal is filtered and compensated before being used as the training input for the echo state network.

8. A noise reduction method based on adaptive retraining of echo state networks, characterized in that, A noise reduction system based on adaptive retraining of an echo state network as described in any one of claims 1 to 7, characterized in that the method comprises: The ambient noise signal is acquired through the feedforward microphone, and the residual noise signal in the ear is acquired through the feedback microphone. The environmental noise signal is filtered using a trained neural network filter containing an echo state network to output a filtered inverse noise signal. The echo state network includes an input layer, a reservoir, and an output layer. The input weights and internal connection weights of the reservoir remain unchanged after initialization, and only the output weights are determined through training. The inverted noise signal is superimposed with the residual noise signal in the ear and adaptive gain adjustment is applied to generate a noise cancellation signal, and the speaker is driven to play the noise cancellation signal to cancel the noise in the ear. Under the condition of meeting the preset retraining triggering conditions, the echo state network is retrained online. The preset retraining triggering conditions include at least detecting a change in the way the headphones are worn or a change in the amount of acoustic leakage. The online retraining includes: using the environmental noise signal and the residual noise signal in the ear within a preset time window before the current moment to form a training sample set, and using the recursive least squares method to update the output weights of the echo state network.

9. The noise reduction method based on adaptive retraining of an echo state network according to claim 1, characterized in that, The online retraining specifically includes: S1. Cache the environmental noise signal within the preset time window. and the residual noise signal in the ear This constitutes the training input sequence and the target error sequence; S2. Input the training input sequence frame by frame into the reservoir of the echo state network, and collect the state vector of the reservoir at each time step. Construct the state matrix ; S3. Using the target error sequence as the desired output, solve the updated output weight matrix using the recursive least squares method with a forgetting factor. : in, For the forgetting factor A diagonal weighted matrix with base 1. The regularization coefficient is . It is the identity matrix. The desired output matrix is ​​composed of the target error sequence; S4. The output weight matrix obtained by solving... Update the output layer of the echo state network to complete online retraining.

10. A noise reduction method based on adaptive retraining of an echo state network according to claim 9, characterized in that, Also includes: After the online retraining is completed, the updated echo state network is used to perform real-time inference on the subsequent input environmental noise signals, and the noise reduction performance indicators are continuously monitored. When the noise reduction performance index is detected to be lower than the preset performance threshold, or when a new wearing method or acoustic leakage is detected, the online retraining is triggered again to form a closed-loop adaptive noise reduction control loop.

Images