A state-driven dynamic virtual sensing active noise reduction method, device and medium

Abstract

The application discloses a state-driven dynamic virtual sensing active noise reduction method and device, and a medium, relates to the technical field of active noise reduction and audio signal processing, and comprises the following steps: constructing a mapping relationship, calibrating a filter bank and a lookup table; obtaining real-time low-frequency energy ratio based on a reference microphone signal and a physical error microphone signal collected in real time, obtaining a current leakage state factor through the mapping relationship and an asymmetric time smoothing strategy, linearly interpolating to dynamically generate a current virtual error reconstruction filter coefficient vector and a current virtual secondary path model; performing dot product operation on the historical vector and the current virtual error reconstruction filter coefficient vector to obtain a virtual eardrum error signal estimate value, combining the current virtual secondary path model to execute a normalized FxLMS control algorithm, and outputting an active noise reduction control signal. The application realizes continuous tracking of the leakage state and model smoothing switching, and significantly improves the noise reduction robustness and eardrum noise reduction performance under complex wearing conditions.

Description

Technical Field

[0001] This application relates to the field of active noise reduction and audio signal processing technology, and in particular to a state-driven dynamic virtual sensing active noise reduction method, device and medium. Background Technology

[0002] Existing in-ear active noise cancellation systems typically employ a single reference microphone plus a physical error microphone inside the ear canal, using a filter-x least mean square algorithm to adaptively eliminate error signals. However, the following problems exist in actual products: 1. Due to factors such as wearing depth, ear canal shape and user movement, the leakage state of the ear canal changes randomly, which alters the acoustic characteristics of the ear canal and leads to a severe mismatch of the noise reduction controller calibrated based on the sealing state.

[0003] 2. Due to limitations in safety, comfort, and structural complexity, it is impossible to place a microphone on the eardrum for an extended period. Therefore, the actual product can only rely on a reference microphone outside the shell, and the real eardrum error can only be measured briefly during experimental calibration, making it difficult to obtain online.

[0004] 3. Most existing virtual sensing solutions are based on a fixed mapping of the outer error to the diaphragm error under a single sealed state. When the leakage state changes significantly, this fixed model will produce obvious mismatch, resulting in a decrease in noise reduction performance or even instability. Furthermore, some solutions that support multi-state switching can only handle discrete scenarios and are difficult to smoothly adapt to continuously changing leakage levels.

[0005] Therefore, there is an urgent need in this field for an active noise reduction technology that can estimate the leakage state online and dynamically schedule the virtual error and secondary path model according to the leakage state, so as to improve the robustness and noise reduction performance in complex leakage scenarios. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this application provides a state-driven dynamic virtual sensing active noise reduction method, device, and medium.

[0007] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a state-driven dynamic virtual sensing active noise reduction method, including: Multiple leakage states were selected to construct a mapping relationship between the leakage state factor and the low-frequency energy ratio; For each leakage state, the virtual error reconstruction filter coefficient vector and the virtual secondary path FIR model are calibrated and stored as a filter bank and a lookup table. Real-time acquisition of reference microphone signal and physical error microphone signal, low-frequency preprocessing and energy estimation, and determination of real-time low-frequency energy ratio; Based on the mapping relationship, the real-time low-frequency energy ratio is mapped to the instantaneous leakage state estimate, and the current leakage state factor is obtained by using an asymmetric time smoothing strategy. Based on the current leakage state factor, linear interpolation is performed in the filter bank and lookup table to dynamically generate the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model; The virtual tympanic membrane error signal estimate is obtained by performing a dot product operation between the historical vector of the physical error microphone signal and the coefficient vector of the current virtual error reconstruction filter. The historical vector of the reference microphone signal is filtered through the current virtual secondary path model to obtain the filter reference vector. The estimated value of the virtual tympanic membrane error signal is used as the error input. The normalized FxLMS control algorithm is executed to update the control filter coefficients, and an active noise reduction control signal is generated and output.

[0008] Secondly, this application provides a state-driven dynamic virtual sensing active noise cancellation device, which is installed in an in-ear headphone, hearing aid, or wearable hearing device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the steps of the state-driven dynamic virtual sensing active noise cancellation method provided above.

[0009] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the state-driven dynamic virtual sensing active noise reduction method described above.

[0010] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a state-driven dynamic virtual sensing active noise reduction method, device, and medium. By mapping the real-time low-frequency energy ratio to an instantaneous leakage state estimate based on a mapping relationship, and employing an asymmetric time smoothing strategy to obtain the current leakage state factor, it achieves continuous and robust online estimation of the ear canal leakage state, overcoming the abrupt problem of discrete state switching. By performing linear interpolation in the filter bank and lookup table based on the current leakage state factor, it dynamically generates the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model, ensuring real-time matching between the virtual sensing model and the acoustic leakage degree, significantly reducing performance degradation caused by fixed model mismatch. The asymmetric smoothing strategy balances the stability of leakage detection with the speed of state recovery, avoiding model scheduling jitter. Using the virtual tympanic membrane error signal estimate as the control target, combined with a state-driven normalized FxLMS algorithm, it achieves stable convergence of active noise reduction with tympanic membrane position as the optimization target even when tympanic membrane error cannot be directly measured, significantly improving the actual listening experience and noise reduction robustness in complex wearing scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a state-driven dynamic virtual sensing active noise reduction method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the mapping relationship between the leakage state factor and the low-frequency energy ratio provided in an embodiment of this application; Figure 3 A schematic diagram illustrating virtual error reconstruction relationships in different scenarios provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the smooth change of low-frequency energy ratio during switching between different leakage scenarios, provided in an embodiment of this application. Figure 5 A schematic diagram illustrating the smooth change of leakage factors during switching between different leakage scenarios, provided in an embodiment of this application; Figure 6 A historical vector of physical error microphone signal provided in an embodiment of this application And virtual tympanic membrane error signal estimation value A comparison diagram; Figure 7 This is a schematic diagram illustrating the implementation process of a state-driven dynamic virtual sensing active noise reduction method provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] In one exemplary embodiment, this application provides a state-driven dynamic virtual sensing active noise reduction method. This method is executed by a computer device, specifically by a terminal or server, or by both. In this embodiment, the method is described using a server as an example. Figure 1 As shown, the method includes: Step 100: Select multiple leakage states and construct the mapping relationship between the leakage state factor and the low-frequency energy ratio; Step 101: For each leakage state, calibrate the virtual error reconstruction filter coefficient vector and the virtual secondary path FIR model, and store them as a filter bank and a lookup table; Step 102: Real-time acquisition of reference microphone signal and physical error microphone signal, low-frequency preprocessing and energy estimation, and determination of real-time low-frequency energy ratio; Step 103: Based on the mapping relationship, the real-time low-frequency energy ratio is mapped to the instantaneous leakage state estimate, and the current leakage state factor is obtained by using an asymmetric time smoothing strategy; Step 104: Based on the current leakage state factor, perform linear interpolation in the filter bank and lookup table to dynamically generate the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model; Step 105: Perform a dot product operation between the historical vector of the physical error microphone signal and the coefficient vector of the current virtual error reconstruction filter to obtain the estimated value of the virtual tympanic membrane error signal; Step 106: Filter the historical vector of the reference microphone signal through the current virtual secondary path model to obtain the filter reference vector, and use the estimated value of the virtual tympanic membrane error signal as the error input to execute the normalized FxLMS control algorithm to update the control filter coefficients, and generate and output the active noise reduction control signal.

[0016] By implementing steps 100-106 above, this application can achieve continuous tracking of leakage status and smooth model switching, thereby significantly improving noise reduction robustness and tympanic membrane noise reduction performance under complex wearing conditions.

[0017] In an exemplary embodiment of this application, steps 100 and 101 correspond to the offline calibration stage. The purpose of offline calibration is to first create a batch of stable and known typical leakage states in a controlled environment, and then establish the relationship between the leakage states and several key model parameters in advance, so as to facilitate direct querying / interpolation for dynamic adaptation during online runtime. Based on this, the implementation process of steps 100-101 can be described as follows: To establish a mapping relationship or lookup table between low-frequency energy ratio and leakage state factor, we first select a set of typical leakage state factors corresponding to multiple typical leakage states, such as: .

[0018] In the formula, This is a typical leakage state factor. This represents the total number of typical leakage states, i.e., the total number of calibrated states. The larger the value, the finer the final mapping relationship or lookup table, resulting in higher accuracy, but also higher calibration costs and storage requirements.

[0019] A common setting is: Where 0 corresponds to the ideal sealing state, 1 corresponds to the maximum leakage state, and several intermediate values ​​(such as...) This corresponds to different degrees of leakage or improper wearing.

[0020] During the experiment, the headphones remained stable under each leakage condition, for example, by constructing representative wearing scenarios through different sizes of ear tips, deliberate gaps, and variations in insertion depth.

[0021] For each common setting The speaker is controlled in an open-loop state (i.e., no active noise cancellation control signal is sent), that is... , here This refers to the control signal that the speaker would normally emit for noise reduction. Without the speaker emitting any control signals, it simultaneously begins acquiring signals from two microphones: the reference microphone signal is denoted as... The signal from the physical error microphone (the one inside or near the ear canal) is recorded as... The reference microphone signal can be used. It can be understood as a reference input for external noise, and the physical error microphone signal is used as a reference input. This can be understood as the result of external noise being sealed / leaked through the earplug and forming inside the ear.

[0022] Because leakage has the most significant impact on low frequencies (low frequencies are more likely to leak into the ear canal when the seal deteriorates), in order to extract low-frequency energy characteristics that are sensitive to changes in leakage, both signals (reference microphone signal and physical error microphone signal) are first preprocessed through an identical first-order infinite impulse response low-pass filter to suppress rapid high-frequency changes and retain only slow-changing / low-frequency components. The filter coefficients... By the set cutoff frequency (e.g., 400 Hz) and system sampling frequency Decision. The filtered low-frequency signals (i.e., the low-frequency signal from the reference microphone and the low-frequency signal from the physical error microphone) are denoted as follows: and (subscript) lfThe English word for low frequency is "low frequency". There are two purposes for doing this: first, to highlight the energy characteristics of the low frequency band that change significantly with leakage; second, to suppress the more random fluctuations in the mid-to-high frequency range that are more easily affected by changes in the noise field, so that the characteristics calculated later are more stable.

[0023] The filter itself is defined by the following two recursive formulas: the current low-frequency output is not exactly equal to the current input, but rather a weighted average of a small portion of the current input and a large portion of the output from the previous time step. If the memory coefficients (i.e., the filter coefficients) If the value is close to 1, the output relies more on past outputs, changes slowly, and the filtering is stronger (low-pass filtering is more powerful). If the memory factor is smaller, the output follows the current input more closely, changes faster, and the filtering is weaker. Using the same filter for two signals ensures consistent processing and avoids identifying differences in characteristics as caused by the filter rather than leakage.

[0024] .

[0025] In the formula, This is the low-frequency signal from the reference microphone at the previous moment.

[0026] Among them, the memory coefficients are the filter coefficients. The specific formula is as follows: .

[0027] Based on this, a cutoff frequency is first specified. (e.g., 400 Hz), combined with the sampling frequency This allows us to calculate a suitable memory factor, ensuring that the first-order low-pass filter is approximately within the cutoff frequency range. The frequency attenuation begins to noticeably decrease to higher frequencies in the vicinity. Intuitively, this is the cutoff frequency. The larger the cutoff frequency, the wider the desired frequency range, and the slower the filtering needs to be. Therefore, the calculated memory coefficient will be smaller, and the output will more easily follow the input. Conversely, the cutoff frequency... The smaller the value, the more effective the smoothing is needed to retain lower frequencies. This results in a calculated memory coefficient closer to 1, making the output more dependent on past performance and changing more slowly. (Sampling frequency) The reason it appears in the denominator is that the faster the sampling, the more sampling points the same time scale will be divided into, and the discrete parameters of the filter need to be adjusted accordingly to maintain the same actual cutoff frequency effect.

[0028] For the two low-frequency signals after low-pass filtering (i.e., the low-frequency signal from the reference microphone and the low-frequency signal from the physical error microphone), they are first converted into sequences of energy magnitude (essentially focusing on intensity, not instantaneous positive or negative values), and then first-order smoothing is performed to obtain a more stable low-frequency energy estimate. The smoothed energy sequences are denoted as follows: and ,have: .

[0029] In the formula, To update speed or smooth intensity, The reference microphone energy sequence before smoothing, The microphone energy sequence before smoothing is the physical error. This is affected by changes in update speed or smoothing intensity. Larger values, new They are more willing to follow the current trend. Changes are faster but fluctuations are also greater; if the update speed or smoothing strength... Smaller It relies more on the past. The changes are slower but smoother. (Update speed or smoothing intensity) Based on sampling frequency and the expected smoothing time constant Decisions, such as: .

[0030] When the desired smoothing time constant Sampling times ranging from tens to hundreds of milliseconds provide a relatively stable low-frequency energy estimate, suitable for identifying wear leakage conditions. Sampling frequency The higher the value, the more sampling points per second, which is important if a smoothing time constant is desired. If the value remains unchanged, then the contribution of each sampling point to the average should be smaller, so the update speed or smoothing strength should be adjusted accordingly. It will vary with the sampling frequency It increases and decreases; similarly, if we want the energy estimate to be more stable and the averaging time to be longer (the desired smoothing time constant), it will also decrease. If it increases, then the update speed or smoothing intensity will be affected. It should also be smaller, making system updates slower. The desired smoothing time constant. The time constant is taken as tens to hundreds of milliseconds because the leakage state is essentially a slowly changing variable caused by the wearer, and does not require millisecond-level tracking; conversely, if a smooth time constant is desired... If the time is too short, the energy estimate will follow the random fluctuations of noise, and the characteristics will be unstable, making it unsuitable for identifying leakage conditions.

[0031] When in a certain fixed leakage state (that is, a typical leakage state) Data was collected over a period of time, and two stable low-frequency energy estimates (i.e., smoothed energy sequences) were obtained. and After that, we can construct the actual characteristics used to infer the leakage, namely the low-frequency energy ratio. ,have: .

[0032] Low frequency energy ratio This represents the ratio of the low-frequency energy of the physical error microphone to the low-frequency energy of the reference microphone; where, A small constant is used to prevent numerical instability caused by an excessively small denominator. The aforementioned low-frequency energy ratio... In the formula, when the overall external noise increases, both the numerator and denominator tend to increase together, making the ratio relatively more stable and better reflecting the relative changes caused by leakage. Then, using the ratio again... The conversion to dB (decibels) is because energy ratios can span a wide range; using dB is more compact, easier to compare, and also better aligns with the common expression of intensity ratios in acoustics and signal processing. (Small constants) Its function is not to change the physical meaning, but to ensure the robustness of numerical calculations, such as when... When the value is very small (e.g., in a quiet environment), avoid the denominator being too small, which could lead to an explosion in the ratio or a division by zero problem, so that the entire feature can be calculated normally under all working conditions.

[0033] Then, average this energy ratio over the steady-state range to obtain a calibration point corresponding to the leakage state. ,have: .

[0034] Typical leakage conditions With calibration point Create discrete point pairs and plot the leakage state factor. With low frequency energy ratio The relationship curve (combining the point pairs of energy ratio and leakage state factor, such as...) Figure 2 (As shown). In software implementation, these point pairs can be stored as filter banks and lookup tables, or obtained as continuous function forms through polynomial fitting, spline fitting, etc. , This represents the fitted function. When running online, this function... This provides a mapping from real-time low-frequency energy ratio to leakage state factor. During the online phase, this relationship is used to infer the leakage state from the real-time calculated energy ratio. The aforementioned steady-state interval emphasizes averaging over a period where the leakage state remains consistently stable, avoiding transient states such as immediate onset, adjustment, or accidental friction / collision. This results in the calibration point. Only then can the current level of leakage be considered a typical leakage condition. The low-frequency energy ratio must be high; otherwise, the calibration point will have transient deviations, leading to errors in the lookup table.

[0035] Whether it's lookup table interpolation or function fitting, they are logically the same thing: in the offline stage, measurable features (i.e., low-frequency energy ratios) are first established. ) and leakage state factor The correspondence between them; in the online stage, there is no need to conduct complex experiments, only to calculate the low-frequency energy ratio in real time. This mapping relationship allows us to infer the current leakage state, thus providing a unified state input for subsequent model selection and parameter adaptation.

[0036] It should be noted that the leakage state factor is not limited to monotonic changing scenarios. In actual wear, the leakage state factor... The leakage state factor may fluctuate randomly, drift slowly, or change periodically. The above calibration only provides a few reference points for subsequent interpolation and state tracking. In practical applications, the leakage state factor... The changes can occur in any manner over time, including random, periodic, or irregular changes. The leakage state observation and state-driven virtual sensing method provided in this application is also applicable to any of the above-mentioned changes.

[0037] In one exemplary embodiment of this application, the implementation process of step 102 described above may include: The reference microphone signal and the physical error microphone signal are preprocessed using the same first-order IIR low-pass filter to obtain the low-frequency signal. The instantaneous energy of the low-frequency signal is calculated and first-order exponential smoothing is performed to obtain a smoothed low-frequency energy estimate. Based on the smoothed low-frequency energy estimate, the ratio of the low-frequency energy estimate of the physical error microphone to the low-frequency energy estimate of the reference microphone is determined and converted into a decibel value to obtain the real-time low-frequency energy ratio.

[0038] In one exemplary embodiment of this application, in establishing the low-frequency energy ratio Leakage state factor Based on the mapping relationship, it is necessary to further establish a mapping model from the physical error point to the sound pressure at the tympanic membrane position, namely, a virtual error reconstruction filter. The signal measured by a physical error microphone (i.e., the physical error microphone signal) that is relatively easy to place in the ear canal is used to estimate the error in the tympanic membrane position; this estimated error is called the virtual error. To enable real-time calculation during online operation, this mapping is represented by an FIR filter. Essentially, it involves using a past physical error signal, weighted and summed through a set of fixed coefficients, to obtain an estimate of the current tympanic membrane error. Based on this, the implementation process of step 105 above can be described as follows: In each leakage state, the signal from the physical error microphone is first organized into a history vector, that is, the values ​​of the current moment and the previous few sampling points are packed into a column in sequence, as follows: .

[0039] In the formula, It is not a single number, but a vector of numbers, representing the physical error at the current sampling point. This represents the physical error of the previous sampling point, and so on, until the error is reached. There are several points. Acoustic systems have propagation delays and cavity effects. The sound pressure at the eardrum is not determined by this instant, but is related to the changes in the sound field over a short period of time. Only by using historical data as input can the model capture this memory-based relationship. This represents the number of historical points taken, which is the number of points in the FIR filter. A value that is too short may affect fitting accuracy, while a value that is too long increases computational burden and the risk of overfitting. For each state, the values ​​of the physical error microphone at the current moment and several previous sampling points are combined to form a historical vector / historical data block, which serves as the input to the FIR filter.

[0040] Suppose there exists a set of FIR coefficients We can perform a linear combination of this historical vector to obtain the best approximation of the tympanic membrane target signal, so that the estimated value of the virtual tympanic membrane error signal is expressed as: : .

[0041] Past Multiply each physical error sampling point by its corresponding coefficient, and then sum them all up to obtain the current estimated value of the virtual tympanic membrane error signal. Here, the estimated value of the virtual tympanic membrane error signal is... Not the actual value obtained directly from measurement; FIR coefficient It is a length of The coefficient vector, for example . It is the FIR coefficient. The transpose (from column to row) is needed to match the historical vector of the microphone signal with physical error. Perform the inner product operation. The actual expansion of the inner product operation is: This is the typical form of an FIR filter, whose input is a series of historical samples and whose output is a weighted sum of these samples. It is called a virtual error reconstruction filter because it uses the measurable signal from the physical error microphone to reconstruct (estimate) the error signal of the tympanic membrane position.

[0042] The least squares algorithm is applied to a large number of samples collected under this condition to solve for the estimated value of the virtual tympanic membrane error signal. With tympanic membrane error The optimal filter coefficient vector with the minimum mean square error This makes the virtual tympanic membrane error signal estimate value The closest approximation to the true tympanic membrane error in the least squares sense This process aims to find a set of FIR coefficients that minimizes the average difference (minimum average squared difference) between the predicted tympanic membrane error and the actual measured tympanic membrane error across the entire dataset. This statistically approximates the linear mapping between the physical error point and the tympanic membrane position, assuming the relationship can be approximated by a linear system—that is, the relationship where each historical sampling point is multiplied by a coefficient and then summed (an FIR filter uses this linear model). Least squares does not guarantee that the world is linear; however, if a linear model is used to approximate the true relationship, then least squares can provide the best set of linear coefficients for this dataset. Least squares not only aims to match at a specific time step but also to minimize the overall error across all time steps; moreover, to avoid positive and negative errors canceling each other out, it squares the error and then averages it, resulting in the mean squared error (MSE). Repeat this process for k typical leakage states to obtain a filter bank for each state. And organize them into a matrix for storage, with . This is the matrix for storage.

[0043] When running online, based on the leak status estimated in real time. Interpolation is performed between these filter banks to dynamically generate a virtual sensing model adapted to the current leakage. Only offline operations were performed. There are discrete states (e.g., 0, 1 / 3, 2 / 3, 1), but the changes in online wearing are continuous and won't fall exactly on these few points. Therefore, we use... The process involves determining which two filters the current input is closer to, then interpolating their coefficients to obtain an intermediate filter. The intuition behind interpolation is like smoothly transitioning between two different settings: if... Closer to 1 / 3, it resembles the filter used in the 1 / 3 state; closer to 2 / 3, it resembles the filter used in the 2 / 3 state; in between, it's a mixture of both. This allows the virtual error model to smoothly adjust as leakage changes slowly, preventing sudden shifts in estimation.

[0044] In one exemplary embodiment of this application, the acoustic transmission relationship of the control sound output from the speaker, after propagating through the headphone structure, ear canal cavity, and leakage gaps, reaches the eardrum. This transmission relationship is commonly referred to as the secondary path. Based on this, a filter bank and a lookup table are used to construct virtual secondary paths. This prepares an acoustic path model from the speaker to the eardrum for the FxLMS algorithm, i.e., provides a crucial prior model for the FxLMS active noise reduction algorithm. The virtual secondary path can switch / interpolate according to the leakage state. In each typical leakage state... Under certain conditions, a known excitation signal (which can be broadband noise, an MLS sequence, or a swept frequency signal, etc.) is input to the loudspeaker to measure the error signal of the tympanic membrane position. As the output, by combining the known input and the measured output, the virtual secondary path under this leakage state can be obtained using an identification method, and represented by a finite-length FIR filter, we have: .

[0045] In the formula, Represents the secondary path FIR vector. The second path FIR represents the first Tap coefficient , The length of the selected FIR filter.

[0046] Complete all After identifying the paths under typical leakage states, the virtual secondary paths corresponding to each state are stored in pairs with the leakage factors to form a secondary path lookup table (LUT), which is as follows: .

[0047] In the formula, This forms a secondary path lookup table.

[0048] This lookup table (belonging to filter banks and lookup tables) can be understood as a set of labeled models: each leakage calibration point Each corresponds to a specific secondary path FIR vector. During online operation, the current leakage state factor is first obtained. Since the leakage state during actual wear is usually continuously changing, the current leakage state factor... The current leakage state factor often doesn't fall exactly on the discrete points of the offline calibration. Therefore, interpolation is needed between adjacent calibration points to generate a secondary path model that more closely reflects the current state. Because the current leakage state factor changes continuously and may fall between two calibration points, linear interpolation is used to obtain the current virtual secondary path model when the current leakage state factor is between two calibration points. The interpolation weights are normalized to ensure smooth interpolation and numerical stability. Between two calibration points and Between these, the corresponding current virtual secondary path model can be represented as: .

[0049] In the formula, Current leakage state factor The corresponding current virtual secondary path model is the calibration point. The corresponding virtual secondary path model, For calibration point The corresponding virtual secondary path model, and All are factors related to the current leakage state. Adjacent calibration points. To normalize the interpolation weights, i.e., the current leakage state factor In the interval The positional proportion within, therefore when hour ,when hour And when the current leakage state factor Located at the calibration point and Between The resulting interpolation weights naturally satisfy the normalization requirement, as follows: .

[0050] Based on the above description, the implementation process of step 104 in this application can be described as follows: determining whether the current leakage state factor is within the offline calibration interval; when the current leakage state factor is within the offline calibration interval, determining normalized weights between the two calibration points in the offline calibration interval; using the normalized weights to perform linear interpolation in the filter bank and lookup table to generate the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model; when the current leakage state factor is not within the offline calibration interval, directly using the filter coefficient vector corresponding to the boundary calibration state as the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model. The virtual error reconstruction relationships for different scenarios are as follows: Figure 3 As shown.

[0051] In one exemplary embodiment of this application, in order to track changes in the wearing status of headphones in real time and observe the leakage status online, the core task is to calculate the leakage status factor at the current moment using the real-time acquired reference microphone signal and the physical error microphone signal. As a unified state-driven variable for the entire system, the closer its value is to 0, the more airtight the fit; the closer it is to 1, the more severe the leakage. Based on this, the implementation process of step 103 above may include: First, continuously acquire reference microphone signals and physical error microphone signals. Specifically, acquire the reference microphone signal (i.e., the reference signal from outside the earphone). And physical error microphone signals (i.e., physical error signals from near the earpiece). To ensure consistency between online observation and offline calibration, both signals were processed using the same low-frequency preprocessing and energy estimation procedures as the offline calibration stage described above, resulting in a smoothed low-frequency energy estimate. Based on this, the real-time low-frequency energy ratio was calculated as the leakage observation metric; the calculation method is described in the formula above. .

[0052] Next, the point-to-point set constructed during the offline calibration phase will be utilized. This maps the current real-time low-frequency energy ratio to a leakage state. Specifically, at the calibration point... With calibration point Piecewise linear interpolation is performed within the defined calibration interval. Interpolation weights. The calculation is as follows: . This refers to the real-time low-frequency energy ratio.

[0053] Based on this, the instantaneous leakage state estimate is obtained as follows: .

[0054] when When the instantaneous leakage state is exceeded, the estimated value will be... Clamp to Furthermore, [0,1] interval clipping is performed to avoid non-physical interpretation.

[0055] Due to the instantaneous leakage state estimate It may be too sensitive to short-term noise fluctuations, and direct use in model scheduling could cause unnecessary jitter; therefore, time smoothing is necessary. To align with the control intuition of cautiously addressing leakage aggravation and rapidly recovering from leakage mitigation, this application employs an asymmetric smoothing strategy to obtain the current leakage state factor. Let the leakage state factor at the previous time step be... ,when At that time, it was assumed that the leakage was tending to worsen, so a smaller smoothing coefficient (i.e., the first smoothing coefficient) was used. Updated, with: .

[0056] In the formula, The leakage state factor at the current moment. A smaller positive number, such as 0.001 to 0.01, corresponds to a longer time constant, in order to suppress false alarms and enhance stability.

[0057] Conversely, when At that time, it was assumed that the leakage was decreasing, so a larger smoothing coefficient (i.e., the second smoothing coefficient) was used. Updated, with: .

[0058] In the formula, A larger positive number, such as 0.01 to 0.1, corresponds to a shorter time constant, so as to quickly track state improvement and restore performance in a timely manner.

[0059] Finally, the leakage state factor at the current moment. By pruning the [0,1] interval (i.e. the defined interval of the mapping relationship), we obtain the robust leakage state factor, which is the current leakage state factor. Figure 4 and Figure 5 The results of smooth changes in low-frequency energy ratio and leakage factor during different leakage scenario switching are presented.

[0060] In one exemplary embodiment of this application, an acoustic mapping model suitable for the current wearing conditions is dynamically constructed using the real-time estimated current leakage state factor, and the historical vector of the microphone signal, which is easily measurable, is used to represent the physical error. Real-time conversion to virtual tympanic membrane error signal estimation value . Figure 6 Give the historical vector of the microphone signal with physical error. And virtual tympanic membrane error signal estimation value Comparison results. Based on this, the implementation process of step 105 above can be described as follows: At each sampling time First, using the same method as the offline calibration stage described above, we define the historical vector of the physical error signal, which is: .

[0061] In this formula, a length of... Using historical vectors as input allows the model to more accurately reconstruct the current tympanic membrane error by utilizing short-term past information. The reason for using a historical period instead of just the current point is that acoustic propagation and ear canal cavity effects have memory; the sound pressure at the tympanic membrane is not only determined by the sound at this moment but also by the changes in the sound field over a short period of time.

[0062] Subsequently, based on the current leakage state factor, the virtual sensing filter bank obtained from offline calibration is... Perform linear interpolation to generate the filter coefficient vector corresponding to the current state. The interpolation rule is consistent with the linear interpolation form of the aforementioned virtual secondary path, that is, at adjacent calibration points... and Calculate normalized weights between And based on this, and The weighted sum is used to obtain the filter coefficient vector corresponding to the current state. When the current leakage state factor exceeds the limit, the filter corresponding to the boundary calibration state is directly used as the coefficient vector of the current virtual error reconstruction filter and the current virtual secondary path model. That is, when the current leakage state factor exceeds the offline calibration range, extrapolation is not performed; instead, the filter corresponding to the boundary calibration state is directly used as the coefficient vector of the current virtual error reconstruction filter and the current virtual secondary path model. This avoids the model becoming uncontrollable due to extrapolation, thus making it more robust.

[0063] Finally, the estimated value of the virtual tympanic membrane error signal is obtained by vector dot product, as follows: Includes There are 3 coefficients, each corresponding to a sampling point in the history vector; the dot product operation is equivalent to multiplying each history sampling point by its corresponding coefficient and then summing them. From a signal processing perspective, this is a standard FIR filter: the input is... The output is The difference lies in the fact that the FIR filter coefficients here are not fixed, but are generated online by interpolation based on the current leakage state factor. Therefore, it can dynamically adjust the mapping relationship as the leakage changes, so that the estimated value of the reconstructed virtual tympanic membrane error signal is... The virtual tympanic membrane error signal is estimated to closely approximate the actual error at the tympanic membrane under various wearing conditions. It is regarded as a real-time estimate of residual noise at the tympanic membrane and is used as the error signal input for the normalized FxLMS control algorithm.

[0064] In one exemplary embodiment of this application, the normalized FxLMS control algorithm dynamically schedules the current virtual secondary path model under the current leakage condition using the current leakage state factor, thereby ensuring that the tympanic membrane position is always the control target for noise cancellation under different leakage conditions.

[0065] The controller input is a reference microphone signal. Since the control output of the normalized FxLMS control algorithm is essentially a response to the reference microphone signal... Perform a linear filter, that is, use the reference microphone signal The current value is added to the values ​​over a past period, weighted accordingly. Therefore, the reference microphone signal must be added first. Organized into a historical vector, we have: .

[0066] In the formula, To control the length of the filters, the current virtual secondary path model consists of an offline-built filter bank and a secondary path lookup table. The reference buffer vector is obtained by linear interpolation based on the current leakage state factor. Then, it is filtered through the currently estimated virtual secondary path to obtain the filtered reference scalar. ,have: .

[0067] In the formula, Indicates transpose. This represents the dynamic interpolation result of the current leakage state factor.

[0068] Then, from this filter reference scalar Its historical value construction length is Filtering reference vector ,have: .

[0069] The vector of control filter coefficients is denoted as This indicates the weighting combination of historical reference signal values ​​that the controller currently believes should be used to generate a suitable anti-phase control signal. The control signal at the current moment. Produced by the dot product, we have: .

[0070] Based on this formula, by multiplying each component of the historical vector of the reference microphone signal by its corresponding weight and then summing them, the control signal that the speaker needs to output at this moment can be obtained. .

[0071] When controlling the vector of filter coefficients As you gradually learn to be more appropriate, This will more effectively cancel out noise at the tympanic membrane. The control objective is to minimize residual noise at the tympanic membrane position. Since the tympanic membrane error cannot be directly measured online, it can be estimated using a virtual tympanic membrane error signal. As an error signal, the vector of control filter coefficients is processed using the normalized FxLMS control algorithm. Adaptive updates are performed. The update idea of ​​the normalized FxLMS control algorithm is: if the current error is still large, adjust the vector of control filter coefficients. The direction of the next output. It can better offset errors; and the direction of this adjustment depends on the filtering reference vector. The decision is made based on the step size. To ensure stable updates under different environmental intensities, the normalized FxLMS control algorithm uses the step size to vary with the filter reference vector. The energy is automatically scaled.

[0072] Furthermore, a first-order smoothing is performed on the filtered reference energy to obtain the smoothed result. : .

[0073] In the formula, Used to control smooth speed This can be understood as the filter reference vector. The sum of the squares of each component represents the intensity of the updated input at this moment. This is the smoothed result from the previous time step.

[0074] Calculate the normalization step size factor ,have: .

[0075] In the formula, To preset the maximum step size, To adjust the normalization strength and use it for numerically stable positive numbers, in order to avoid When the value is very small, the denominator is too small, resulting in an excessively large step size. When smoothing the result... When the energy is high, the denominator increases, and the step size automatically decreases to prevent excessive divergence during updates; when the result is smoothed... When the energy is low, the denominator is small, and the step size becomes relatively large, ensuring that there is still enough renewal force.

[0076] The control filter coefficients are updated according to the following rules: .

[0077] In the formula, These are the updated control filter coefficients. Using the error magnitude... and the filter reference vector The resulting update direction collectively determines how to adjust the control filter coefficients, and uses the normalized step size factor. The adjustment range is controlled. In the formula, the negative sign indicates that the update direction is to reduce the error (i.e., update along the direction that reduces the error). With continuous iteration, the control filter coefficients will gradually converge to a set that makes... Minimize the weight to effectively cancel noise at the tympanic membrane. As the leakage changes, the actual secondary path also changes, which is obtained through dynamic interpolation of the current leakage state factor. Let the filter reference vector The calculation changes along with the leakage, so that the update direction of FxLMS remains reasonable under different wearing conditions, ultimately achieving stable convergence and stable noise reduction with the tympanic membrane position as the target.

[0078] Based on the above description, the implementation process of step 106 provided in this application can be described as follows: constructing a historical vector of a reference microphone signal; filtering the historical vector of the reference microphone signal using the current virtual secondary path model to obtain the filtered reference vector; performing first-order smoothing on the energy of the filtered reference vector to obtain a smoothing result; determining a normalization step size factor based on the smoothing result; updating the control filter coefficients using the normalization step size factor; generating and outputting the active noise reduction control signal based on the dot product of the updated control filter coefficients and the historical vector of the reference microphone signal.

[0079] In summary, the overall implementation process of the state-driven dynamic virtual sensing active noise reduction method provided in this application can be as follows: Figure 7 As shown, it can be applied to scenarios where there is ear canal leakage, such as in-ear headphones, hearing aids, and wearable hearing devices. In one exemplary embodiment, a state-driven dynamic virtual sensing active noise cancellation device is provided, disposed in an in-ear headphone, hearing aid, or wearable hearing device, comprising: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the steps of the state-driven dynamic virtual sensing active noise cancellation method provided in the above embodiment.

[0080] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0081] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0082] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0083] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (RRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0084] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A state-driven dynamic virtual sensing active noise reduction method, characterized in that, include: Multiple leakage states were selected to construct a mapping relationship between the leakage state factor and the low-frequency energy ratio; For each leakage state, the virtual error reconstruction filter coefficient vector and the virtual secondary path FIR model are calibrated and stored as a filter bank and a lookup table. Real-time acquisition of reference microphone signal and physical error microphone signal, low-frequency preprocessing and energy estimation, and determination of real-time low-frequency energy ratio; Based on the mapping relationship, the real-time low-frequency energy ratio is mapped to the instantaneous leakage state estimate, and the current leakage state factor is obtained by using an asymmetric time smoothing strategy. Based on the current leakage state factor, linear interpolation is performed in the filter bank and lookup table to dynamically generate the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model; The virtual tympanic membrane error signal estimate is obtained by performing a dot product operation between the historical vector of the physical error microphone signal and the coefficient vector of the current virtual error reconstruction filter. The historical vector of the reference microphone signal is filtered through the current virtual secondary path model to obtain the filter reference vector. The estimated value of the virtual tympanic membrane error signal is used as the error input. The normalized FxLMS control algorithm is executed to update the control filter coefficients, and an active noise reduction control signal is generated and output.

2. The state-driven dynamic virtual sensing active noise reduction method according to claim 1, characterized in that, Perform low-frequency preprocessing and energy estimation to determine the real-time low-frequency energy ratio, including: The reference microphone signal and the physical error microphone signal are preprocessed using the same first-order IIR low-pass filter to obtain the low-frequency signal. The instantaneous energy of the low-frequency signal is calculated and first-order exponential smoothing is performed to obtain a smoothed low-frequency energy estimate. Based on the smoothed low-frequency energy estimate, the ratio of the low-frequency energy estimate of the physical error microphone to the low-frequency energy estimate of the reference microphone is determined and converted into a decibel value to obtain the real-time low-frequency energy ratio.

3. The state-driven dynamic virtual sensing active noise reduction method according to claim 1, characterized in that, The current leakage state factors are obtained using an asymmetric time smoothing strategy, including: Let the leakage state factor at the previous moment be... The instantaneous leakage state estimate is ; like The leak was determined to be worsening, and the first smoothing coefficient was used. The leakage state factor is updated to obtain the current leakage state factor; at this time, the leakage state factor is expressed as: ； In the formula, The leakage state factor at the current moment; like The leakage was determined to be decreasing, and a second smoothing coefficient was applied. The leakage state factor is updated to obtain the current leakage state factor; at this time, the leakage state factor is expressed as: ； The leakage state factor at the current moment is clipped to the defined range of the mapping relationship to obtain the current leakage state factor.

4. The state-driven dynamic virtual sensing active noise reduction method according to claim 1, characterized in that, Based on the current leakage state factor, linear interpolation is performed in the filter bank and lookup table to dynamically generate the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model, including: Determine whether the current leakage state factor is within the offline calibration range; When the current leakage state factor is in the offline calibration interval, a normalized weight is determined between the two calibration points in the offline calibration interval; The normalized weights are used to perform linear interpolation in the filter bank and lookup table to generate the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model; When the current leakage state factor is not in the offline calibration range, the filter coefficient vector corresponding to the boundary calibration state is directly used as the current virtual error reconstruction filter coefficient vector and the current virtual secondary path model.

5. The state-driven dynamic virtual sensing active noise reduction method according to claim 1, characterized in that, The historical vector of the reference microphone signal is filtered through the current virtual secondary path model to obtain a filtered reference vector. Using the estimated virtual tympanic membrane error signal as the error input, the normalized FxLMS control algorithm is executed to update the control filter coefficients, generating and outputting an active noise reduction control signal, including: Construct the history vector of the reference microphone signal; The historical vector of the reference microphone signal is filtered using the current virtual secondary path model to obtain the filtered reference vector; The energy of the filtered reference vector is smoothed using a first-order smoothing process to obtain a smoothing result. Determine the normalization step size factor based on the smoothing results; The control filter coefficients are updated using the normalized step size factor. The active noise reduction control signal is generated and output based on the dot product of the updated control filter coefficients and the historical vector of the reference microphone signal.

6. The state-driven dynamic virtual sensing active noise reduction method according to claim 2, characterized in that, The memory coefficients of the first-order IIR low-pass filter are determined by the cutoff frequency and the system sampling frequency.

7. The state-driven dynamic virtual sensing active noise reduction method according to claim 2, characterized in that, The smoothing coefficient of the first-order exponential smoothing is determined by the desired smoothing time constant.

8. The state-driven dynamic virtual sensing active noise reduction method according to claim 4, characterized in that, The current virtual secondary path model is represented as follows: ； In the formula, Current leakage state factor The corresponding current virtual secondary path model, For calibration point The corresponding virtual secondary path model, For calibration point The corresponding virtual secondary path model, and All are factors related to the current leakage state. Adjacent calibration points, and satisfying ; For normalized weights.

9. A state-driven dynamic virtual sensing active noise reduction device, characterized in that, It is installed in in-ear headphones, hearing aids, or wearable hearing devices; The state-driven dynamic virtual sensing active noise reduction device includes: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the state-driven dynamic virtual sensing active noise reduction method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the state-driven dynamic virtual sensing active noise reduction method as described in any one of claims 1-8.

Images