Filter coefficient matching model training method, active noise reduction method and system

Abstract

The application relates to the technical field of audio processing, in particular to a filter coefficient matching model training method, an active noise reduction method and a system. The training method comprises the following steps: obtaining multiple tasks, each task comprising a support set and a query set, the support set comprising multiple support samples from multiple noise categories and corresponding optimal filter coefficient labels, and the query set comprising multiple query samples from the same multiple noise categories; constructing a target training function by taking the probability of matching each query sample to a target label as a training target, wherein the target label is an optimal filter coefficient label corresponding to a support sample belonging to the same category as the query sample; and iteratively training a matching network according to the multiple tasks and the target training function to obtain a filter coefficient matching model. The method can improve the noise reduction effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of audio processing, in particular to a filter coefficient matching model training method, an active noise reduction method and system. BACKGROUND

[0002] Active noise cancellation (ANC) is a noise cancellation method based on the principle of acoustic signal superposition, which aims to generate an anti-noise signal with the same amplitude and opposite phase as the noise signal to eliminate or attenuate the noise signal.

[0003] A selective fixed-filtered active noise cancellation (SFANC) method based on a convolutional neural network is proposed in the related technology. The SFANC method adopts a traditional deep learning method, performs targeted training based on a large number of training samples, and finally realizes matching of the best filter coefficient according to the noise type, thereby solving the noise problem in a specific scene. The premise for its application is that the sample distribution can well represent the overall distribution.

[0004] However, the actual application scenario is very complex, and the noise signals generated in different application scenarios are various, for example, in the vehicle active noise cancellation scenario, the noise signal depends on different influencing factors such as vehicle conditions, driving habits and road conditions. Therefore, when there is a significant difference between the online noise and the training samples, the SFANC method may not be able to match the best filter coefficient, resulting in poor noise reduction effect. SUMMARY

[0005] Therefore, a filter coefficient matching model training method, an active noise reduction method and system are provided to improve the problem that the related SFANC method cannot match the best filter coefficient when there is a significant difference between the online noise and the training samples, so as to improve the noise reduction effect.

[0006] In a first aspect, a filter coefficient matching model training method is provided, and the method comprises:

[0007] Obtaining a plurality of tasks, each task comprising a support set and a query set, the support set comprising a plurality of support samples from a plurality of noise categories and corresponding best filter coefficient labels, and the query set comprising a plurality of query samples from the same plurality of noise categories;

[0008] Maximizing the probability of matching each query sample to a target label as a training target, constructing a target training function, and the target label is the best filter coefficient label corresponding to the support sample belonging to the same category as the query sample;

[0009] iteratively training a matching network according to the plurality of tasks and the target training function, to obtain a filter coefficient matching model.

[0010] With reference to the first aspect, in a first implementation form of the first aspect, before the iteratively training a matching network according to the plurality of tasks and the target training function, the method further includes:

[0011] constructing a first feature extraction sub-network, the first feature extraction sub-network being configured to extract a hidden feature of each of the support samples;

[0012] constructing a second feature extraction sub-network, the second feature extraction sub-network being configured to extract a hidden feature of each of the query samples;

[0013] constructing an attention score calculation sub-network, the attention score calculation sub-network being configured to calculate a probability that each of the query samples matches to the target label according to the hidden feature of each of the query samples and the hidden features of the support samples;

[0014] obtaining the matching network based on the first feature extraction sub-network, the second feature extraction sub-network and the attention score calculation sub-network.

[0015] With reference to the first implementation form of the first aspect, in a second implementation form of the first aspect, the constructing an attention score calculation sub-network includes:

[0016] constructing a similarity calculation function, the similarity calculation function being configured to calculate a similarity between each of the query samples and each of the support samples according to the hidden feature of each of the query samples and the hidden features of the support samples;

[0017] constructing an attention weight calculation function, the attention weight calculation function being configured to assign a corresponding attention weight according to the similarity between each of the query samples and each of the support samples;

[0018] constructing an attention score calculation function, the attention score calculation function being configured to obtain the probability that each of the query samples matches to the target label according to a weighted sum of the attention weights and the best filter coefficient labels;

[0019] obtaining the attention score calculation sub-network according to the similarity calculation function, the attention weight calculation function and the attention score calculation function.

[0020] With reference to the first aspect, in a third implementation form of the first aspect, before the obtaining the plurality of tasks, the method further includes:

[0021] The training database is constructed according to a preset sound effect library and a preset noise database, the preset sound effect library including a plurality of non-speech environmental noises, and the preset noise database including background noises in a plurality of environments;

[0022] The construction step of each task includes:

[0023] A plurality of noise categories are selected from the training database, and a plurality of noise samples are selected from noise samples corresponding to the plurality of noise categories as corresponding support samples to form the support set, and a plurality of noise samples are selected from remaining samples corresponding to the plurality of noise categories as corresponding query samples to form the query set;

[0024] Each support sample is queried in a preset filter coefficient database to obtain corresponding optimal filter coefficients and generate corresponding optimal filter coefficient labels;

[0025] The task is constructed according to the support set and the query set.

[0026] In the fourth implementation manner of the first aspect, before each support sample is queried in the preset filter coefficient database, the method further includes:

[0027] In the simulation simulation experiment, a plurality of Gaussian noises of different frequencies are obtained;

[0028] The interference signals generated by each Gaussian noise through a main path are determined, the main path representing a path from a reference microphone to an error microphone;

[0029] Based on each Gaussian noise and each interference signal, an adaptive filtering algorithm is used to obtain the optimal filter coefficients corresponding to each Gaussian noise;

[0030] The filter coefficient database is constructed according to the optimal filter coefficients corresponding to each Gaussian noise.

[0031] In the fourth implementation manner of the first aspect, in the fifth implementation manner of the first aspect, determining the interference signals generated by each Gaussian noise through the main path includes:

[0032] In a given noise reduction environment, acoustic signal tests are respectively performed at a reference microphone and an error microphone to obtain corresponding noise source signals and a first ear signal;

[0033] The main path transfer function is determined according to the noise source signal and the first ear signal.

[0034] The corresponding interference signal is obtained according to the main path transfer parameter and each Gaussian noise.

[0035] In a sixth implementation form of the first aspect, in combination with the fourth implementation form of the first aspect, the adaptive filtering algorithm is configured to obtain the optimal filter coefficients corresponding to each of the Gaussian noises based on each of the Gaussian noises and each of the interference signals, including:

[0036] In a given noise reduction environment, an acoustic signal test is performed at the cancellation speaker and the error microphone respectively to obtain a corresponding noise reduction signal and a second human ear signal;

[0037] A secondary path transfer function is determined according to the noise reduction signal and the second human ear signal, the secondary path representing a path between the cancellation speaker and the error microphone;

[0038] According to each of the interference signals, a corresponding anti-noise signal is determined;

[0039] Based on each of the Gaussian noises, the secondary path transfer function, a preset cancellation speaker transfer function, and each of the anti-noise signals, an adaptive filtering algorithm is used to obtain the optimal filter coefficients corresponding to each of the Gaussian noises.

[0040] In a second aspect, an active noise reduction method is provided, the method comprising:

[0041] An online support set is obtained, the online support set including a plurality of real noises from an actual vehicle environment and corresponding optimal filter coefficients thereof;

[0042] An online noise in a current vehicle environment is collected, and the online support set and the online noise are input into a trained filter coefficient matching model for processing to obtain a current optimal filter coefficient, the trained filter coefficient matching model being obtained according to the training method of the filter coefficient matching model of any one of the first aspect or the first implementation form to the sixth implementation form of the first aspect;

[0043] The online noise is reduced according to the current optimal filter coefficient.

[0044] In a third aspect, a training system of a filter coefficient matching model is provided, the system comprising:

[0045] An acquisition module is configured to acquire a plurality of tasks, each of the tasks including a support set and a query set, the support set including a plurality of support samples from a plurality of noise categories and corresponding optimal filter coefficient labels thereof, and the query set including a plurality of query samples from the same plurality of noise categories;

[0046] The constructing module is configured to maximize the probability of matching each query sample to a target label as a training target, and construct a target training function, wherein the target label is a best filter coefficient label corresponding to a support sample belonging to the same category as the query sample;

[0047] The training module is configured to iteratively train the matching network according to the plurality of tasks and the target training function, to obtain a filter coefficient matching model.

[0048] In a fourth aspect, an active noise reduction system is provided, and the system comprises:

[0049] The preprocessing module is configured to obtain an online support set, wherein the online support set comprises a plurality of real noises from an actual vehicle environment and corresponding best filter coefficients;

[0050] The matching module is configured to collect online noises in a current vehicle environment, input the online support set and the online noises into the trained filter coefficient matching model for processing, to obtain a current best filter coefficient, wherein the trained filter coefficient matching model is obtained according to the training method of the filter coefficient matching model of any one of the first aspect or the first to sixth implementable manners of the first aspect;

[0051] The noise reduction module is configured to reduce the online noises according to the current best filter coefficient.

[0052] The training method of the filter coefficient matching model, the active noise reduction method and the system, by obtaining a plurality of tasks, each task including a support set and a query set, the support set including a plurality of support samples from a plurality of noise categories and their corresponding optimal filter coefficient labels, and the query set including a plurality of query samples from the same plurality of noise categories; maximize the probability of each query sample matching to a target label as a training target, construct a target training function, and the target label is the optimal filter coefficient label corresponding to the support sample belonging to the same category as the query sample; iteratively train the matching network according to the plurality of tasks and the target training function, and obtain the filter coefficient matching model. Applying the filter coefficient matching model, the current optimal filter coefficient with the best online noise reduction effect can be matched according to a plurality of real noises from the actual driving environment and the online noise from the current driving environment, which have matched the optimal filter coefficient. The training method of the filter coefficient matching model of the present application uses the matching network for training, and does not need to collect training samples covering all actual active noise reduction scenes when training the model. The matching network calculates the similarity between the query sample and each support sample in the support set, so as to classify the query samples and the support samples of the same category into the same category, and then use the same optimal filter coefficient for the query sample, which makes the trained model can be applied to the noise signal that has not been encountered before, without additional training for each new category of noise signal. Therefore, compared with the related art, the present application can match the optimal filter coefficient when there is a significant difference between the actual noise and the training sample, and improve the noise reduction effect. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 An application environment diagram of the training method of the filter coefficient matching model in one embodiment;

[0054] Figure 2 A flowchart of the training method of the filter coefficient matching model in one embodiment;

[0055] Figure 3 A flowchart of the step of extracting implicit features by the one-dimensional convolutional recurrent network in one embodiment;

[0056] Figure 4 A structure block diagram of a typical active noise reduction system in one embodiment;

[0057] Figure 5 A flowchart of the active noise reduction method in one embodiment;

[0058] Figure 6 A structure block diagram of the training system of the filter coefficient matching model in one embodiment;

[0059] Figure 7 A structure block diagram of the active noise reduction system in one embodiment;

[0060] Figure 8 FIG. 1 is a diagram of an internal structure of a computer device according to an embodiment;

[0061] Figure 9 FIG. 1 is a diagram of an internal structure of a computer device according to an embodiment. DETAILED DESCRIPTION

[0062] In order to make the purposes, technical solutions, and advantages of the present application clearer, the present application will be described in further detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present application, and are not intended to limit the present application.

[0063] It should be noted that the diagrams provided in the embodiments only schematically illustrate the basic concepts of the present application, and only show the components related to the present application in the diagrams, rather than being drawn according to the number, shape, and size of the components in actual implementation. The shapes, number, and proportions of the components in actual implementation can be arbitrarily changed, and the layout pattern of the components can also be more complex.

[0064] The structures, proportions, sizes, and the like shown in the diagrams attached to the present specification are only used to cooperate with the content disclosed in the present specification, to enable those skilled in the art to understand and read, and are not used to limit the defined conditions under which the present application can be implemented. Therefore, any modification of the structure, change of the proportional relationship, or adjustment of the size, without affecting the effects that can be generated by the present application and the purposes that can be achieved, should still fall within the scope of the technical content disclosed by the present application.

[0065] The orientations or positional relationships indicated by terms such as "upper", "lower", "left", "right", "intermediate", "vertical", "horizontal", "inner", "outer", "radial", "circumferential", and the like as used in this specification are based on the orientations or positional relationships shown in the drawings, and are only used to facilitate the description, and thus cannot be understood as indicating or implying that a referred device or element must have a particular orientation, be constructed and operated in a particular orientation, and thus cannot be understood as limiting the present application. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance.

[0066] In a feed-forward active noise cancellation (ANC) system, in order to adjust and optimize the noise cancellation effect in real time, an adaptive filter is usually used to minimize the error signal. The FxLMS algorithm is a widely used adaptive filtering algorithm, which compensates for the secondary path delay and achieves optimal noise control around the error microphone. However, the adaptive filtering method represented by the FxLMS algorithm has the following limitations: when the secondary path changes greatly or the system output is saturated due to high noise level, this kind of method may lose stability; the slow convergence of the adaptive algorithm limits the effect of dynamic noise and overall noise reduction; the traditional adaptive filtering algorithm needs to run at a high sampling frequency, which requires higher performance of the processor.

[0067] On the contrary, the active noise cancellation method using fixed filter does not have the above problems, and becomes another kind of widely used active noise cancellation method. Among them, the selective fixed-filtered active noise cancellation (SFANC) method can select the filter coefficient with the best noise reduction effect for different noises. The related SFANC method uses a general scene noise data set to purposefully train a neural network, thereby solving the filter coefficient matching problem of a specific scene; but this method has a major limitation, that is, in a new, complex, and different from the training sample distribution scene, the accuracy of selecting the filter coefficient decreases, resulting in poor noise reduction effect, and the premise of this method is that the samples meet the independent and identically distributed assumption, that is, the sample distribution can well represent the overall distribution.

[0068] However, the actual application scene is very complex, and the noise signals generated in different application scenes are also different. Taking the vehicle active noise cancellation scene as an example, the actual driving environment is very complex, and the noise signal depends on different influencing factors such as vehicle conditions, driving habits and road conditions, and it is very likely to encounter a new, complex, and completely different from the training sample distribution scene in the actual noise reduction scene. At this time, this method is difficult to match the filter coefficient with the best noise reduction effect, resulting in poor noise reduction effect.

[0069] Therefore, the present application proposes a filter coefficient matching model training method, which can be applied to, for example Figure 1In the application environment shown, the vehicle end 102 communicates with the server 104 through a network. The server 104 obtains a plurality of tasks, each task including a support set and a query set, the support set including a plurality of support samples from a plurality of noise categories and corresponding optimal filter coefficient labels thereof, and the query set including query samples from the same plurality of noise categories; constructs a target training function by taking the probability of matching each query sample to a target label as a training target, the target label being an optimal filter coefficient label corresponding to a support sample belonging to the same category as the query sample; iteratively trains the matching network according to the plurality of tasks and the target training function to obtain a filter coefficient matching model. The server 104 can be implemented by an independent server or a server cluster composed of a plurality of servers.

[0070] Unlike traditional deep learning methods, the present application uses a matching network for training, without the need for retraining using a large amount of data for each new category. It finds the most similar support samples in the support set to the query samples and classifies them according to the noise categories to which these similar samples belong, and uses the same optimal filter coefficient for samples of the same category. This mechanism helps to improve the generalization ability of the model to unknown noise, even in the case of changes in noise category distribution such as the generation of new noise categories, and maintains good filter coefficient matching performance and ensures good noise reduction effect.

[0071] Next, the training method of the filter coefficient matching model of the present application will be described in detail through the following embodiments.

[0072] In one embodiment, as shown in Figure 2 , a training method of a filter coefficient matching model is provided. This method is applied to the server 104 in Figure 1 for example, and includes the following steps:

[0073] Step 202, obtaining a plurality of tasks, each task including a support set and a query set, the support set including a plurality of support samples from a plurality of noise categories and corresponding optimal filter coefficient labels thereof, and the query set including a plurality of query samples from the same plurality of noise categories.

[0074] Since the practical application scenarios of active noise reduction are very complex, the noises generated by different application scenarios are also different. For example, in the active noise reduction scenario of a vehicle, the noise signal depends on different influencing factors such as vehicle conditions, driving habits, and road conditions. Moreover, due to the difficulty in sample collection and high testing cost, it is difficult to collect noise that can cover all practical application scenarios as training samples when training the model. Therefore, the support samples and the query samples are not limited to the noise of certain specific scenarios. In some embodiments, multiple non-speech environmental noises from a preset sound effect library and / or background noises in multiple environments from a preset noise database are used.

[0075] In a specific embodiment, a training data set is constructed according to a preset sound effect library and a preset noise database; multiple categories are randomly selected from the training data set, multiple noises are randomly selected for each category, and these noises are respectively used as support samples to construct a support set; multiple noises are randomly selected from the remaining noises of the same multiple categories, and these noises are respectively used as query samples to construct a query set; since the matching network is trained in units of tasks, the support set and the query set are used as a whole to construct a task; the above step of constructing a task is repeated multiple times to obtain multiple tasks, and the multiple tasks form a task set.

[0076] For any noise, it can be determined through testing which best filter coefficient can achieve the best noise reduction effect. Therefore, for each support sample, in a simulation test environment such as a Matlab simulation test environment, the support sample is input as a reference signal, the corresponding best filter coefficient is queried from the pre-constructed filter coefficient database, and the best filter coefficient label corresponding to the support sample is generated according to the queried best filter coefficient. The filter coefficient database includes the best filter coefficients corresponding to multiple different frequency Gaussian noises, and the best filter coefficient refers to the filter coefficient with the best noise reduction effect for the support sample.

[0077] Step 204: A target training function is constructed by maximizing the probability of matching each query sample to a target label, and the target label is the best filter coefficient label corresponding to the support sample belonging to the same category as the query sample.

[0078] The support samples are of different noise categories and are labeled with the best filter coefficient of the noise, the query samples are of different noise categories but are not labeled with the best filter coefficient of the noise, and the matching network works as follows: by calculating the similarity between the query sample and each support sample, calculating the attention score of the best filter coefficient label corresponding to the query sample and each support sample according to the similarity, the attention score representing the probability of matching the best filter coefficient label corresponding to the query sample and the support sample to each other, comparing the attention scores to obtain the highest attention score, and matching the best filter coefficient corresponding to the support sample corresponding to the highest attention score as the target label of the query sample.

[0079] In the above process, the higher the similarity between the query sample and the support sample of the same category, the higher the corresponding attention score, that is, the higher the probability of the query sample matching to the target label, for example, continuously approaching 1; on the contrary, the lower the similarity between the query sample and the support sample of the different category, the lower the corresponding attention score, that is, the lower the probability of the query sample matching to the non-target label, for example, continuously approaching 0, and the higher the matching accuracy of the best filter coefficient of the model to the noise. Therefore, the application maximizes the probability of each query sample matching to the target label as the training target, constructs a target training function, and iteratively trains the matching network with the target training function to obtain the optimal filter coefficient matching model.

[0080] Step 206, iteratively training the matching network according to the plurality of tasks and the target training function to obtain a filter coefficient matching model.

[0081] The support samples and query samples of each task are input into the matching network, and the target training function is used to iteratively train the matching network to obtain optimal model parameters, and the optimal model parameters are assigned to the matching network to obtain a trained filter coefficient matching model. The filter coefficient matching model is used to match the current best filter coefficient with the best online noise reduction effect according to the plurality of real noises from the actual noise reduction scene and the online noise from the current noise reduction scene.

[0082] Taking the active noise reduction scene of a vehicle as an example, although a large number of real noises from different actual vehicle environments cannot be collected, a small amount of real noises from different categories can be collected in the process of test driving or internal testing; for these small amounts of real noises, the best filter coefficient with the best noise reduction effect is selected from the pre-constructed filter coefficient database through artificial matching or automatic matching; according to these real noises and the corresponding best filter coefficients, a support set in an online environment (hereinafter referred to as an online support set) is constructed; the online noise in the current vehicle environment and each real noise in the online support set and the corresponding best filter coefficient are input into the best filter coefficient for processing, and the current best filter coefficient with the best noise reduction effect on the online noise is obtained; the filter is configured according to the current best filter coefficient, and the online noise is reduced according to the configured filter.

[0083] The training method of the filter coefficient matching model trains by using a matching network, and does not need to collect training samples covering all active noise reduction scenes during training, but finds support samples with the same category as the query samples from the support set, so that the same best filter coefficient is used for the query samples, which enables the trained model to be applied to new categories of noise signals immediately without additional training for each new category of noise signal. Therefore, compared with the prior art, the present application can match the best filter coefficient when there is a significant difference between the actual noise and the training samples, and ensure the noise reduction effect.

[0084] Next, a complete filter coefficient matching model training process will be described through the following examples.

[0085] Constructing a data set: constructing a data set according to a preset sound effect library and a preset noise database, the preset sound effect library including a plurality of non-speech environmental noises, and the preset noise database including background noises in a plurality of environments; for example, the sound effect library disclosed at http: / / www.sound-ideas.com records more than 10,000 non-speech environmental noises, and NOISE-92 records engine noise, factory noise, speech spectrum noise (SSN) and babble noise (BN); the data set is divided into a training data set, a verification data set and a test data set according to a ratio of 70%, 15% and 15%; in a Matlab simulation test environment, each noise data in the data set is input as a reference signal, and a query is performed in a pre-constructed filter coefficient database to obtain the best filter coefficient with the best noise reduction effect and generate the corresponding best filter coefficient label, and finally a supervised data set is constructed.

[0086] Constructing task set: randomly sampling C noise categories from the training dataset, m noise samples for each noise category, a total of C*m noise samples, taking the C*m noise samples as corresponding support samples to form a support set; from the remaining noise samples corresponding to the same C noise categories, randomly sampling n noise samples for each noise category, a total of C*n noise samples, taking the C*n noise samples as corresponding query samples to form a query set; taking the support set and the query set as a whole to construct a task; repeating the above steps N times to obtain N tasks, and the N tasks constitute a task set.

[0087] Constructing matching network: constructing a first feature extraction subnetwork for extracting the hidden features g(x i ) of each support sample; constructing a second feature extraction subnetwork for extracting the hidden features Constructing an attention score calculation subnetwork for calculating the attention score of each query sample to the target label according to the hidden features of each query sample and the hidden features g(x i ) of each support sample. The attention score represents: the probability of the query sample matching to the target label given the support set, where y represents the target label, and S represents the support set formed by the support samples.

[0088] For the step of constructing the first feature extraction subnetwork and the second feature extraction subnetwork, exemplarily, a one-dimensional convolutional recurrent network (1D CNN) is used to construct the first feature extraction subnetwork and the second feature extraction subnetwork; the one-dimensional convolutional recurrent network can directly process normalized original noise waveforms, and a balance is made between computational complexity and model performance, such as Figure 3As shown, the process of extracting the implicit features includes: the input layer receives the noise signal collected by the reference microphone; the noise signal is subjected to minimum or maximum normalization processing; the normalized noise signal is input into a 1x5 convolution layer with a channel number of 10; the output of the upper layer is connected to a 1x32 convolution layer with a channel number of 20, and then connected to a maximum pooling layer with a size of 512; the output of the upper layer is connected to a flatten layer, which is converted into a one-dimensional array form; the output of the upper layer is connected to a batch normalization layer, and then connected to a fully connected layer with 32 neurons, and then connected to a hyperbolic tangent (Tanh) activation function; the output of the upper layer is connected to a batch normalization layer, and then connected to a fully connected layer with 64 neurons, and then connected to a Tanh activation function; the output of the upper layer is connected to a batch normalization layer, and then connected to a fully connected layer with 128 neurons, and then connected to a Tanh activation function; the output of the upper layer is connected to a batch normalization layer, and then connected to a fully connected layer with 256 neurons, and connected to a Tanh activation function; the output of the upper layer is connected to a fully connected layer with 256 neurons, to obtain the implicit features of the noise signal.

[0089] For the step of constructing the attention score calculation subnetwork, specifically includes: constructing a similarity calculation function, which is used to calculate the similarity between each query sample and each support sample according to the implicit features of each query sample and the implicit features of each support sample g(x i ), for example, cosine similarity Constructing an attention weight calculation function for assigning corresponding attention weights according to the similarity between each query sample and each support sample For example, the attention weight is calculated as follows: k represents the total number of support samples, i and j represent the index of the support sample; constructing an attention score calculation function for obtaining the probability of each query sample matching to the target label according to the weighted sum of each attention weight and each optimal filter coefficient label y i The mathematical expression used includes: Comparing the attention scores to obtain the highest attention score, and matching the optimal filter coefficient label corresponding to the support sample corresponding to the highest attention score to the target label of the query sample.

[0090] wherein y i ​For matrix, in the case of a given support set, each bit value in the matrix depends on the attention score of the best filter coefficient label of the query sample with the support sample in the support set. For example, assuming that the given support set includes four support samples, when calculating the attention score of the best filter coefficient label of the query sample with the first support sample, y i =(1 0 0 0), when calculating the attention score of the best filter coefficient label of the query sample with the second support sample, y i =(0 1 0 0), and so on, which is not listed here.

[0091] Constructing the target training function: we expect to train a model that can maximize the probability of matching each query sample to the target label in the overall task set. Therefore, the target training function is constructed according to the training goal of maximizing the probability of matching each query sample to the target label.

[0092] For example: randomly select N tasks from the task set; randomly select several different noise categories from the support set for each task, and select one support sample for each category; select the same number of categories from the query set, and select one query sample for each category; several support samples and several query samples constitute a batch; extract M times for each task to form M batches; for each batch, the matching network can only match one query sample with each support sample at a time, therefore, poll each query sample, and input each query sample and each support sample into the matching network respectively to obtain the probability of matching each query sample to the target label; add the probabilities of each query sample matching to the target label in the same batch to obtain the probability corresponding to one batch; calculate the mathematical expectation based on the probabilities corresponding to each batch of a task to obtain the probability corresponding to one task; calculate the mathematical expectation based on the probabilities corresponding to several tasks to obtain the final probability; in order to maximize the probability of matching the query sample to the target label in the overall task set, the target training function is constructed with the maximization of the final probability as the training goal, and the expression thereof includes:

[0093]

[0094] Wherein, argmax represents the maximum operator symbol, E represents the mathematical expectation operator symbol, L represents the task, T represents the task set, S represents the support set, N represents the total number of tasks extracted in the task set, M represents the batch number, B represents a batch, E L～T represents the probability expectation of randomly selecting N tasks from the task set, E S～L,B～L represents the probability expectation of randomly selecting the support set and the batch from the task.

[0095] In some embodiments, the automatic gradient calculation function provided by the Pytorch framework is used, the Adam optimizer is used, the learning rate is set to 0.001, the model gradient is automatically updated, the epochs are set to 50, and if the current epochs reach 50, the model is considered to be converged and the training is completed.

[0096] In some embodiments, the steps of constructing the filter coefficient database include: in the simulation experiment, obtaining a plurality of Gaussian noises of different frequencies; determining the interference signals generated by each Gaussian noise through the main path, the main path representing the path from the reference microphone to the error microphone; based on each Gaussian noise and each interference signal, using an adaptive filtering algorithm to obtain the optimal filter coefficient corresponding to each Gaussian noise; and constructing the filter coefficient database according to the optimal filter coefficient corresponding to each Gaussian noise.

[0097] An exemplary, Figure 4 The structure block diagram of a typical feedforward ANC system is shown, which includes a reference microphone, an error microphone, an active noise reduction system and a cancellation speaker. In the Matlab simulation experiment, it is assumed that the frequency interval of the Gaussian noise is [0.02, 7.08] kHz, a Gaussian noise of a frequency is randomly extracted from the interval, the Gaussian noise is transmitted through the main path to the error microphone as the reference signal x(n), and the interference signal d(n) is generated; the Gaussian noise is adaptively filtered by the active noise reduction system, and then transmitted through the secondary path to the error microphone to generate the anti-noise signal a(n);

[0098] Wherein, the interference signal d(n) is obtained by convolution operation according to the reference signal and the main path transfer function p(n), and the mathematical expression includes: d(n) = x(n) * p(n); the anti-noise signal is collected by the cancellation speaker after the Gaussian noise is filtered by the active noise reduction system, and then the convolution operation is performed with the secondary path transfer function s(n), and the mathematical expression includes: a(n) = s(n) * f LS {w(n) T x(n)},f LS represents the transfer function symbol of the cancellation speaker, w(n) T represents the matrix transpose of the filter coefficient of the active noise reduction system; e(n) is the residual noise measured by the error microphone, e(n) = d(n) - a(n);

[0099] When the anti-noise signal and the interference signal completely cancel out, that is, the anti-noise signal is a signal with the same amplitude and opposite phase as the interference signal, the noise reduction performance of the active noise reduction system is the best; therefore, through this relationship, the corresponding anti-noise signal can be determined according to the interference signal; based on the Gaussian noise, the secondary path transfer function, the cancellation speaker transfer function and the anti-noise signal, an adaptive filtering algorithm such as the LMS algorithm is used to back-propagate the mathematical expression for calculating the anti-noise signal, and the best filter coefficient for the best noise reduction effect of the Gaussian noise can be obtained; repeating the above steps multiple times, multiple best filter coefficients corresponding to the Gaussian noise are obtained, and a filter coefficient database is constructed in this way.

[0100] In the above process, the primary path transfer function and the secondary path transfer function need to be determined in advance, combined with Figure 4 The structure diagram shown in the figure is used to explain that the primary path transfer function is determined by the physical characteristics of the path between the reference microphone and the error microphone, and in a given noise reduction environment, such as a vehicle noise reduction environment, acoustic signal tests are performed at the reference microphone and the error microphone respectively, and the corresponding noise source signal and the first ear signal are obtained; according to the relationship that the first ear signal is obtained after the noise source signal is convolved with the primary path transfer function, the primary path transfer function is inversely deduced.

[0101] Similarly, the secondary path is determined by the physical characteristics of the path between the cancellation speaker and the error microphone, and in a given noise reduction environment, such as a vehicle noise reduction environment, acoustic signal tests are performed at the cancellation speaker and the error microphone respectively, and the corresponding noise reduction signal and the second ear signal are obtained, and according to the relationship that the second ear signal is obtained after the noise reduction signal is convolved with the secondary path transfer function, the secondary path transfer function is inversely deduced.

[0102] In one embodiment, as Figure 5 shown, an active noise reduction method is provided, and the method is applied to the vehicle end 102 in Figure 1 for example, and includes the following steps:

[0103] Step 302, obtaining an online support set, the online support set including multiple real noises and corresponding best filter coefficients from an actual vehicle environment.

[0104] In the active noise reduction scene of the vehicle, the noise signal depends on different influencing factors such as vehicle conditions, driving habits and road conditions, although a large amount of real noise from different actual driving environments cannot be collected, a small amount of real noise from different categories can be collected in the test drive or internal test process; For these small amounts of real noise, the best filter coefficient with the best noise reduction effect is selected from the pre-constructed filter coefficient database through artificial matching or automatic matching; According to these real noises and their corresponding best filter coefficients, an online support set is constructed.

[0105] Step 304, collecting online noise in the current driving environment, inputting the online support set and the online noise into the trained filter coefficient matching model for processing to obtain the current best filter coefficient.

[0106] Among them, the trained filter coefficient matching model is obtained according to the training method of the filter coefficient matching model described in the foregoing embodiments. Since each real noise in the online support set has matched the best filter coefficient, inputting each real noise in the online support set and the online noise into the filter coefficient matching model can match the same category of real noise for the online noise, so as to use the same best filter coefficient for the online noise.

[0107] In some embodiments, when there are more real noises in the online support set, for example, there is more than one real noise applicable to each best filter coefficient, the real noises in the online support set are classified according to the categories of the best filter coefficients, and the real noises using the same best filter coefficient are divided into the same category; When used, a real noise and its corresponding best filter coefficient are randomly extracted from each category in the online support set, and the real noises and their corresponding best filter coefficients from different categories and the online noise are input into the filter coefficient matching network for processing, so as to screen out the real noise most similar to the online noise, and then use the same best filter coefficient for the online noise, avoid matching the online noise with multiple real noises in the same category, and avoid the phenomenon of not matching the online noise with real noises in different categories, improve the sufficiency of noise matching, and thus improve the accuracy of filter coefficient matching and ensure the noise reduction effect.

[0108] Step 306, according to the current best filter coefficient, the online noise is reduced.

[0109] The filter coefficient of the filter is configured as the current best filter coefficient, and the online noise is denoised according to the configured filter. Wherein, the above online support set is only issued to the vehicle end, only a small amount of noise samples and their labels are added or updated, without any update of the parameters of the filter coefficient matching model, and without additional training of the model, which not only can alleviate the overfitting phenomenon of the model, greatly expand the actual application scenarios of the model, but also enable the model to be immediately applied to new categories of noise signals. Therefore, compared with the prior art, the best filter coefficient can be matched when there is a significant difference between the actual noise and the training samples, and the denoising effect is improved.

[0110] It should be understood that, although Figure 2 、 5 The steps in the flowchart are displayed in sequence according to the direction of the arrow, but these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 2 、 5 At least part of the steps in 、

[0111] In one embodiment, as shown in Figure 6 A training system of a filter coefficient matching model is provided, comprising an acquisition module, a construction module and a training module, wherein:

[0112] The acquisition module is configured to acquire a plurality of tasks, each of the tasks comprising a support set and a query set, the support set comprising a plurality of support samples from a plurality of noise categories and corresponding best filter coefficient labels, and the query set comprising a plurality of query samples from the same plurality of noise categories;

[0113] The construction module is configured to maximize the probability of matching each of the query samples to a target label as a training target, and to construct a target training function, the target label being a best filter coefficient label corresponding to a support sample belonging to the same category as the query sample;

[0114] The training module is configured to iteratively train a matching network according to the plurality of tasks and the target training function to obtain a filter coefficient matching model.

[0115] In some embodiments, the constructing module is further configured to: construct a first feature extraction subnetwork for extracting hidden features of each of the support samples; construct a second feature extraction subnetwork for extracting hidden features of each of the query samples; construct an attention score calculation subnetwork for calculating a probability of each of the query samples matching to the target label according to the hidden features of each of the query samples and the hidden features of each of the support samples; and obtain the matching network based on the first feature extraction subnetwork, the second feature extraction subnetwork, and the attention score calculation subnetwork.

[0116] In some embodiments, the constructing module is configured to construct the attention score calculation subnetwork, and specifically includes: constructing a similarity calculation function for calculating a similarity between each of the query samples and each of the support samples according to the hidden features of each of the query samples and the hidden features of each of the support samples; constructing an attention weight calculation function for assigning a corresponding attention weight according to the similarity between each of the query samples and each of the support samples; constructing an attention score calculation function for obtaining a probability of each of the query samples matching to the target label according to a weighted sum of each of the attention weights and each of the best filter coefficient labels; and obtaining the attention score calculation subnetwork according to the similarity calculation function, the attention weight calculation function, and the attention score calculation function.

[0117] In some embodiments, before obtaining a plurality of tasks, the obtaining module is further configured to: construct a training database according to a preset sound effect library and a preset noise database, the preset sound effect library including a plurality of non-speech environmental noises, and the preset noise database including background noises in a plurality of environments; and construct a task, a construction step of each of the tasks including: selecting a plurality of noise categories from the training database, and selecting a plurality of noise samples corresponding to the noise categories as support samples to form a support set; selecting a plurality of noise samples from remaining samples corresponding to the noise categories as query samples to form a query set; querying each of the support samples in a preset filter coefficient database to obtain a corresponding best filter coefficient and generate a corresponding best filter coefficient label; and constructing the task according to the support set and the query set.

[0118] In some embodiments, before querying each of the support samples in the preset filter coefficient database, the acquisition module is further configured to: in a simulation experiment, acquire a plurality of Gaussian noises with different frequencies; determine an interference signal generated by each of the Gaussian noises through a main path, the main path representing a path from a reference microphone to an error microphone; based on each of the Gaussian noises and each of the interference signals, obtain the optimal filter coefficient corresponding to each of the Gaussian noises by using an adaptive filtering algorithm; and construct the filter coefficient database according to the optimal filter coefficient corresponding to each of the Gaussian noises.

[0119] In some embodiments, the acquisition module is configured to determine the interference signal generated by each of the Gaussian noises through the main path, specifically comprising: in a given noise reduction environment, performing acoustic signal tests at the reference microphone and the error microphone respectively to obtain a corresponding noise source signal and a first human ear signal; determining a main path transfer function according to the noise source signal and the first human ear signal; and obtaining each of the corresponding interference signals according to the main path transfer parameter and each of the Gaussian noises.

[0120] In some embodiments, the acquisition module is configured to obtain the optimal filter coefficient corresponding to each of the Gaussian noises based on each of the Gaussian noises and each of the interference signals by using an adaptive filtering algorithm, specifically comprising: in a given noise reduction environment, performing acoustic signal tests at a cancellation speaker and the error microphone respectively to obtain a corresponding noise reduction signal and a second human ear signal; determining a secondary path transfer function according to the noise reduction signal and the second human ear signal, the secondary path representing a path from the cancellation speaker to the error microphone; determining a corresponding anti-noise signal according to each of the interference signals; and obtaining the optimal filter coefficient corresponding to each of the Gaussian noises based on each of the Gaussian noises, the secondary path transfer function, a preset cancellation speaker transfer function, and each of the anti-noise signals by using the adaptive filtering algorithm.

[0121] In one embodiment, as shown in Figure 7 An active noise reduction system is provided, comprising a preprocessing module, a matching module and a noise reduction module, wherein:

[0122] The preprocessing module is configured to acquire an online support set, the online support set comprising a plurality of real noises collected in an actual vehicle environment and corresponding optimal filter coefficients;

[0123] The matching module is configured to collect an online noise in a current vehicle environment, input the online support set and the online noise into a trained filter coefficient matching model for processing to obtain a current optimal filter coefficient, the trained filter coefficient matching model being obtained according to the training method of the filter coefficient matching model in the foregoing embodiments;

[0124] a noise reduction module, configured to reduce the online noise according to the current optimal filter coefficient.

[0125] The specific definitions of the training system of the filter coefficient matching model can refer to the definitions of the training method of the filter coefficient matching model, and the specific definitions of the active noise reduction system can refer to the definitions of the active noise reduction method, which are not described here again. Each module in the training system of the filter coefficient matching model and the active noise reduction system can be realized by software, hardware, and combinations thereof, in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor calls and executes the operations corresponding to each module.

[0126] In one embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 8 The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is configured to store data of different categories of noise and corresponding optimal filter coefficients. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement a training method of a filter coefficient matching model.

[0127] In one embodiment, a computer device, which can be a terminal, is provided, and an internal structure diagram of the computer device can be as shown in Figure 9 The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement an active noise reduction method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball, or touchpad arranged on the shell of the computer device, or can be an external keyboard, touchpad, or mouse, etc.

[0128] Those skilled in the art can understand that Figure 8 , 9 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0129] In an embodiment, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the steps of the training method of the filter coefficient matching model.

[0130] In an embodiment, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program. The computer program is executed by a processor to implement the steps of the active noise reduction method.

[0131] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0132] Each technical feature of the above embodiments can be combined arbitrarily. In order to make the description simple, not all possible combinations of each technical feature in the above embodiments are described, however, as long as the combination of technical features does not exist contradictory, it should be considered as the scope of the present application.

[0133] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A training method for a filter coefficient matching model, characterized in that, The method comprises the following steps: obtaining a plurality of tasks, each of the tasks comprising a support set and a query set, the support set comprising a plurality of support samples and corresponding optimal filter coefficient labels from a plurality of noise categories, and the query set comprising a plurality of query samples from the same plurality of noise categories; constructing a target training function by taking a probability of matching each of the query samples to a target label as a training target, the target label being an optimal filter coefficient label corresponding to a support sample belonging to a same category as the query sample; constructing a first feature extraction subnetwork for extracting hidden features of each of the support samples; constructing a second feature extraction subnetwork for extracting hidden features of each of the query samples; constructing an attention score calculation subnetwork for calculating the probability of matching each of the query samples to the target label according to the hidden features of each of the query samples and the hidden features of each of the support samples; obtaining a matching network based on the first feature extraction subnetwork, the second feature extraction subnetwork and the attention score calculation subnetwork; iteratively training the matching network according to the plurality of tasks and the target training function to obtain a filter coefficient matching model. 2.The method of Claim 1, wherein, The method of constructing the attention score calculation subnetwork comprises the following steps: constructing a similarity calculation function for calculating a similarity between each of the query samples and each of the support samples according to the hidden features of each of the query samples and the hidden features of each of the support samples; constructing an attention weight calculation function for assigning a corresponding attention weight according to the similarity between each of the query samples and each of the support samples; constructing an attention score calculation function for obtaining the probability of matching each of the query samples to the target label according to a weighted sum of each of the attention weights and each of the optimal filter coefficient labels; obtaining the attention score calculation subnetwork according to the similarity calculation function, the attention weight calculation function and the attention score calculation function. 3.The method of Claim 1, wherein, Before the step of obtaining a plurality of tasks, the method further comprises the following steps: constructing a training database according to a preset sound effect library and a preset noise database, the preset sound effect library comprising a plurality of non-speech environmental noises, and the preset noise database comprising background noises in a plurality of environments; the step of constructing each of the tasks comprises the following steps: selecting a plurality of noise categories from the training database, and selecting a plurality of noise samples as corresponding support samples from noise samples corresponding to the plurality of noise categories to form the support set, and selecting a plurality of noise samples as corresponding query samples from remaining samples corresponding to the plurality of noise categories to form the query set; querying each of the support samples in a preset filter coefficient database to obtain a corresponding optimal filter coefficient and generate a corresponding optimal filter coefficient label; constructing the task according to the support set and the query set. 4.The method of Claim 3, wherein, Before querying each of the support samples in a preset filter coefficient database, the method further comprises: In a simulation experiment, a plurality of Gaussian noises of different frequencies are obtained; An interference signal generated by each of the Gaussian noises through a main path is determined, the main path representing a path from a reference microphone to an error microphone; Based on each of the Gaussian noises and each of the interference signals, an adaptive filtering algorithm is used to obtain a best filter coefficient corresponding to each of the Gaussian noises; According to the best filter coefficient corresponding to each of the Gaussian noises, the filter coefficient database is constructed. 5.The method of Claim 4, wherein, The determination of the interference signal generated by each of the Gaussian noises through the main path comprises: In a given noise reduction environment, acoustic signal tests are respectively performed at the reference microphone and the error microphone to obtain corresponding noise source signals and a first human ear signal; A main path transfer function is determined according to the noise source signals and the first human ear signal; According to the main path transfer parameter and each of the Gaussian noises, a corresponding interference signal is obtained.

6. The method of Claim 4, wherein, Based on each of the Gaussian noises and each of the interference signals, an adaptive filtering algorithm is used to obtain a best filter coefficient corresponding to each of the Gaussian noises, comprising: In a given noise reduction environment, acoustic signal tests are respectively performed at a cancellation speaker and an error microphone to obtain corresponding noise reduction signals and a second human ear signal; A secondary path transfer function is determined according to the noise reduction signals and the second human ear signal, the secondary path representing a path from the cancellation speaker to the error microphone; According to each of the interference signals, a corresponding anti-noise signal is determined; Based on each of the Gaussian noises, the secondary path transfer function, a preset cancellation speaker transfer function, and each of the anti-noise signals, an adaptive filtering algorithm is used to obtain a best filter coefficient corresponding to each of the Gaussian noises.

7. An active noise reduction method, characterized by, Comprise: An online support set is obtained, the online support set comprising a plurality of real noises from an actual vehicle environment and corresponding best filter coefficients; An online noise in a current vehicle environment is collected, and the online support set and the online noise are input into a trained filter coefficient matching model for processing to obtain a current best filter coefficient, the trained filter coefficient matching model being obtained according to the training method of the filter coefficient matching model in any one of claims 1-6; The online noise is reduced according to the current best filter coefficient.

8. A training system of a filter coefficient matching model, characterized by, Comprise: An acquisition module is configured to acquire a plurality of tasks, each of the tasks comprising a support set and a query set, the support set comprising a plurality of support samples from a plurality of noise categories and corresponding best filter coefficient labels, and the query set comprising a plurality of query samples from the same plurality of noise categories; A construction module is configured to maximize a probability of matching each of the query samples to a target label as a training target, and to construct a target training function, the target label being a best filter coefficient label of a support sample belonging to the same category as the query sample; A training module is configured to iteratively train a matching network according to the plurality of tasks and the target training function to obtain a filter coefficient matching model. The construction module is further used for: constructing a first feature extraction subnetwork for extracting hidden features of each support sample; constructing a second feature extraction subnetwork for extracting hidden features of each query sample; constructing an attention score calculation subnetwork for calculating a probability of each query sample matching to the target label according to the hidden features of each query sample and the hidden features of each support sample; obtaining the matching network based on the first feature extraction subnetwork, the second feature extraction subnetwork and the attention score calculation subnetwork.

9. An active noise reduction system characterized by, comprise: a preprocessing module configured to acquire an online support set, the online support set comprising a plurality of real noises collected in an actual driving environment and corresponding optimal filter coefficients; a matching module configured to collect online noises in a current driving environment, input the online support set and the online noises into a trained filter coefficient matching model for processing to obtain current optimal filter coefficients, the trained filter coefficient matching model being obtained according to the training method of the filter coefficient matching model in any one of claims 1-6; a noise reduction module configured to reduce the online noises according to the current optimal filter coefficients.

Images