A howling suppression method and system for hearing aids

Abstract

The present application provides a howling suppression method and system for a hearing aid, the method comprising: obtaining an input signal of a preset length; dividing the input signal into a plurality of signal frames based on a time sequence; inputting each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label comprises a voice label and a howling label; according to the label corresponding to each of the signals and the howling stage of the hearing aid, the input signal is howled suppressed using a preset howling suppression method; wherein the signal recognition model is constructed based on a neural network model and is obtained through multi-stage training based on historical signal data of each howling stage, so as to improve the howling suppression effect of the hearing aid and the user experience.

Description

Technical Field

[0001] The present application relates to the field of machine learning and signal processing technology, and in particular to a howling suppression method and system for hearing aids. Background Art

[0002] Hearing aid howling, also known as "acoustic feedback" or "self-excited oscillation", refers to the process in which the amplified sound is picked up again by the microphone and amplified again due to some reason during the use of the hearing aid, thus forming a closed loop, causing the sound signal to be continuously amplified within the loop, and ultimately producing a harsh, high-frequency howling sound. The howling sound generated by the hearing aid during use greatly affects the user experience. Therefore, in the development of hearing aid products, anti-howling is a very important function and a basic and indispensable function.

[0003] Among the existing hearing aid products, anti-howling technologies mainly include the following: 1. Digital signal processing technology. Digital hearing aids can analyze the spectrum of environmental signals and track the changes in the spectrum to distinguish the frequency bands of noise and speech. By digitally converting and filtering the sound signals received by the microphone, digital signal processing technology can accurately identify and suppress feedback signals that may cause howling. This method not only effectively reduces the occurrence of howling, but also significantly improves the audio quality and speech clarity of the hearing aid. 2. Physical compression technology. Physical compression technology is usually implemented through the built-in compressor of the hearing aid. When the input sound signal reaches a certain amplitude, the compressor will automatically reduce the volume to avoid howling caused by excessive amplification of the sound signal. Although this technology is relatively simple, it can also effectively reduce feedback problems to a certain extent. 3. Feedback suppression algorithm. Some advanced hearing aids use complex feedback suppression algorithms, such as deep edge feedback cancellation (DeepMFC) and other deep learning frameworks. These algorithms can more accurately identify and suppress howling signals, while reducing color effects and limiting excessive amplification at certain frequencies.

[0004] However, the existing feedback suppression algorithms do not take into account the changing process of howling itself. Howling is usually a gradual phenomenon. It may be a weak high-frequency feedback at first, which gradually increases over time. Therefore, in the process of training the relevant models, simulating the gradual formation of howling can help the model identify the different stages of howling, especially the early stages, so as to intervene more effectively. In this way, the model can learn to suppress howling in the early stages of formation, without having to wait until the howling is fully formed. Therefore, training according to the process of howling gradually forming will significantly improve the howling suppression effect of the model. This strategy is based on the dynamic characteristics of howling, which enables the model to better capture the occurrence, growth and maintenance process of howling, and thus perform feedback suppression more accurately. Summary of the invention

[0005] In response to the above technical problems, the present application provides a howling suppression method and system for a hearing aid, which trains a machine learning model based on the characteristics of the howling change process itself, thereby improving the howling suppression effect of the hearing aid and the user experience.

[0006] In a first aspect, an embodiment of the present application provides a howling suppression method for a hearing aid, comprising:

[0007] Obtaining an input signal of a preset length;

[0008] Dividing the input signal into a plurality of signal frames based on a time sequence;

[0009] Inputting each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label includes a speech label and a howling label;

[0010] According to the labels corresponding to the signals and the howling stage of the hearing aid, using a preset howling suppression method to suppress the input signal;

[0011] The signal recognition model is constructed based on a neural network model and is obtained through multi-stage training based on historical signal data of each howling stage.

[0012] The embodiment of the present application provides a howling suppression method for a hearing aid, which uses a signal recognition model to perform signal recognition on an input signal, divides each signal frame into a voice signal and a howling signal, and provides data support for subsequent howling suppression. In addition, the signal recognition model provided by the embodiment of the present application can further identify the current howling stage, and then select the corresponding howling suppression method according to the howling stage to perform howling suppression. Compared with the prior art, the embodiment of the present application takes into account the change process of the howling itself during the howling suppression process, performs targeted howling suppression, and avoids excessive suppression that affects normal voice signals while ensuring the howling suppression effect, thereby improving the howling suppression effect of the hearing aid and the user experience. Correspondingly, in the training process of the signal recognition model, the embodiment of the present application also performs multi-stage training based on the historical signal data of each howling stage to obtain the signal recognition model, thereby improving the accuracy of model recognition and enabling the model to have the ability to recognize the current howling stage according to the input signal.

[0013] In a possible implementation, the signal recognition model is obtained by constructing the signal recognition model based on the neural network model and performing multi-stage training based on historical signal data of each howling stage, including:

[0014] Construct an initial signal recognition model based on a neural network model;

[0015] According to different howling stages, the historical signal data are combined into training data of different stages;

[0016] According to the training data of the different stages, the initial signal recognition model is trained in multiple stages to obtain the signal recognition model, wherein in the model training process of each stage, the initial signal recognition model is trained using a multi-learning rate parallel training method.

[0017] The embodiment of the present application provides a training method for a signal recognition model. For the construction of training data, the embodiment of the present application recombines various historical signal data based on different howling stages to construct and generate training data of different stages, and introduces the characteristic of gradual change of howling intensity into the training process of the model to improve the training effect of the model. Furthermore, in the model training process of each stage, a multi-learning rate parallel training method is used to train the model, so that the training process of the model is not limited to a single learning rate, but multiple learning rates are used for parallel training at the same time to improve the recognition accuracy of the model.

[0018] Furthermore, combining the historical signal data into training data of different stages according to different howling stages includes:

[0019] According to different howling stages, each of the historical signal data is divided into first historical signal data, second historical signal data and third historical signal data, which respectively correspond to the first howling stage, the second howling stage and the third howling stage;

[0020] Using the first historical signal data as first-stage training data;

[0021] Combining the first historical signal data with the second historical signal data to construct second-stage training data;

[0022] The first historical signal data, the second historical signal data and the third historical signal data are combined to construct third-stage training data.

[0023] In a possible implementation, the method of training the initial signal recognition model by using multiple learning rate parallel training includes:

[0024] According to the current learning rate of the initial signal recognition model, a plurality of different training learning rates are generated based on a preset formula and a preset number of parallel trainings;

[0025] Generate a plurality of identical parallel training models according to the current model parameters of the initial signal recognition model, where the number of the parallel training models is the number of the training learning rates;

[0026] Setting different training learning rates for each of the parallel training models, and using the same training data for training, to determine the loss rate of each of the parallel training models on the validation set;

[0027] Determine the optimal parallel training model with the smallest loss rate according to the loss rate of each parallel training model on the validation set;

[0028] According to the model parameters and learning rate of the optimal parallel training model, the current learning rate and current model parameters of the initial signal recognition model are updated, and the next round of multi-learning rate parallel training is performed until the initial signal recognition model meets the preset training termination conditions.

[0029] The embodiment of the present application provides a multi-learning rate parallel model training method. Considering that the application scenario of the embodiment of the present application is to deploy a neural network model in a hearing aid, and the hearing aid is a very extreme embedded device, generally requiring the battery life to be at least 10 hours, the power consumption of its chip cannot be very high, and the computing power that can be provided is not very large. Therefore, the neural network model deployed in the hearing aid is destined not to have too high complexity. Under the premise that the complexity of the model itself is limited, the neural network model applied to the hearing aid should improve the accuracy of model recognition as much as possible during the training process. Traditional algorithms often only use one strategy to update a single learning rate, while the embodiment of the present application abandons the idea of ​​using only one strategy for training in the traditional way, and designs a learning rate update algorithm that matches most scenarios at the same time. Through multiple rounds of iterations and multi-learning rate parallel training, several learning rates and a corresponding number of identical models are generated in each round of iteration for parallel training, and the model parameters and corresponding learning rates with the best training effect are selected to update the initial signal recognition model, and then enter the next round of iteration. The training effect of the model can be improved by increasing the training complexity. At the same time, due to the low complexity of the model itself, this method will not significantly extend the training time, thus achieving a balance between training efficiency and training effect.

[0030] Furthermore, the using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid includes:

[0031] When the howling stage is the first stage, signal suppression is performed on the signal frame with the howling tag in the input signal.

[0032] Furthermore, the using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid includes:

[0033] When the howling stage is the second stage, determining the frequency characteristics of the speech signal and the frequency characteristics of the howling signal according to each of the labels;

[0034] Dividing a corresponding speech frequency band and a howling frequency band according to the frequency characteristics of the speech signal and the frequency characteristics of the howling signal;

[0035] Under the premise of ensuring the integrity of the voice frequency band, the howling frequency band is suppressed.

[0036] Furthermore, the using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid includes:

[0037] When the howling stage is the third stage, the signal spectrum of the hearing aid is adjusted to a preset signal spectrum.

[0038] The embodiment of the present application provides a method for howling suppression based on different howling stages. When the howling stage is the first stage, the embodiment mainly suppresses it slightly according to the howling label to protect the integrity of the voice and prevent excessive compression. When the howling stage is the second stage, the embodiment further strengthens the suppression and performs frequency band suppression on the entire howling frequency band. When the howling stage is the third stage, the howling has been fully formed and obviously affects the voice signal. At this time, in order to prevent further feedback of the howling and deterioration of the sound quality, the integrity of the voice signal is no longer considered, but the entire signal spectrum is adjusted to achieve strong howling suppression. The embodiment of the present application achieves targeted processing of different howling by judging the howling stage, further improving the howling suppression effect of the hearing aid and the user experience.

[0039] In a second aspect, accordingly, an embodiment of the present application provides a howling suppression system for a hearing aid, including an acquisition module, a signal segmentation module, a signal recognition module, a howling suppression module, and a model building module;

[0040] Wherein, the acquisition module is used to acquire an input signal of a preset length;

[0041] The signal segmentation module is used to divide the input signal into a plurality of signal frames based on a time sequence;

[0042] The signal recognition module is used to input each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label includes a speech label and a howling label;

[0043] The howling suppression module is used to suppress the howling of the input signal using a preset howling suppression method according to the labels corresponding to each of the signals and the howling stage of the hearing aid;

[0044] The model building module is used to build a neural network model and perform multi-stage training based on historical signal data of each howling stage to obtain the signal recognition model.

[0045] Further, the model construction module includes an initial model construction unit, a training data construction unit and a training unit;

[0046] Wherein, the initial model building unit is used to build an initial signal recognition model based on a neural network model;

[0047] The training data construction unit is used to combine the historical signal data into training data of different stages according to different howling stages;

[0048] The training unit is used to perform multi-stage training on the initial signal recognition model according to the training data of the different stages to obtain the signal recognition model, wherein in the model training process of each stage, the initial signal recognition model is trained using a multi-learning rate parallel training method.

[0049] In a possible implementation, the training unit trains the initial signal recognition model using a multi-learning rate parallel training method, including:

[0050] According to the current learning rate of the initial signal recognition model, a plurality of different training learning rates are generated based on a preset formula and a preset number of parallel trainings;

[0051] Generate a plurality of identical parallel training models according to the current model parameters of the initial signal recognition model, where the number of the parallel training models is the number of the training learning rates;

[0052] Setting different training learning rates for each of the parallel training models, and using the same training data for training, to determine the loss rate of each of the parallel training models on the validation set;

[0053] Determine the optimal parallel training model with the smallest loss rate according to the loss rate of each parallel training model on the validation set;

[0054] According to the model parameters and learning rate of the optimal parallel training model, the current learning rate and current model parameters of the initial signal recognition model are updated, and the next round of multi-learning rate parallel training is performed until the initial signal recognition model meets the preset training termination conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 : A flow chart of a howling suppression method for a hearing aid provided in an embodiment of the present application.

[0056] Figure 2 : A flow chart of a model construction in a howling suppression method for a hearing aid provided in an embodiment of the present application.

[0057] Figure 3 : A flowchart of multi-learning rate parallel training in a howling suppression method for a hearing aid provided in an embodiment of the present application.

[0058] Figure 4 : A structural schematic diagram of a howling suppression system for a hearing aid provided in an embodiment of the present application. DETAILED DESCRIPTION

[0059] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0060] It should be noted that the step numbers in the text are only for the convenience of explanation of the specific embodiments and do not serve to limit the order in which the steps are executed. In the description of this application, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0061] Embodiment 1:

[0062] like Figure 1 As shown, embodiment 1 provides a howling suppression method for a hearing aid, comprising steps S1-S4:

[0063] Step S1, obtaining an input signal of a preset length;

[0064] Step S2, dividing the input signal into a plurality of signal frames based on time sequence;

[0065] Step S3, inputting each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label includes a speech label and a howling label;

[0066] Step S4, using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid;

[0067] The signal recognition model is constructed based on a neural network model and is obtained through multi-stage training based on historical signal data of each howling stage.

[0068] The embodiment of the present application provides a howling suppression method for a hearing aid, which uses a signal recognition model to perform signal recognition on an input signal, divides each signal frame into a voice signal and a howling signal, and provides data support for subsequent howling suppression. In addition, the signal recognition model provided by the embodiment of the present application can further identify the current howling stage, and then select the corresponding howling suppression method according to the howling stage to perform howling suppression. Compared with the prior art, the embodiment of the present application takes into account the change process of the howling itself during the howling suppression process, performs targeted howling suppression, and avoids excessive suppression that affects normal voice signals while ensuring the howling suppression effect, thereby improving the howling suppression effect of the hearing aid and the user experience. Correspondingly, in the training process of the signal recognition model, the embodiment of the present application also performs multi-stage training based on the historical signal data of each howling stage to obtain the signal recognition model, thereby improving the accuracy of model recognition and enabling the model to have the ability to recognize the current howling stage according to the input signal.

[0069] In a possible implementation, the signal recognition model is obtained by constructing a neural network model and performing multi-stage training based on historical signal data of each howling stage, such as Figure 2 As shown, steps S501-S503 are included:

[0070] Step S501, constructing an initial signal recognition model based on a neural network model;

[0071] Step S502: combining the historical signal data into training data of different stages according to different howling stages;

[0072] Step S503, performing multi-stage training on the initial signal recognition model according to the training data of the different stages to obtain the signal recognition model, wherein in the model training process of each stage, the initial signal recognition model is trained using a multi-learning rate parallel training method.

[0073] The embodiment of the present application provides a training method for a signal recognition model. For the construction of training data, the embodiment of the present application recombines various historical signal data based on different howling stages to construct and generate training data of different stages, and introduces the characteristic of gradual change of howling intensity into the training process of the model to improve the training effect of the model. Furthermore, in the model training process of each stage, a multi-learning rate parallel training method is used to train the model, so that the training process of the model is not limited to a single learning rate, but multiple learning rates are used for parallel training at the same time to improve the recognition accuracy of the model.

[0074] Further, in step S502, the historical signal data are combined into training data of different stages according to different howling stages, including:

[0075] According to different howling stages, each of the historical signal data is divided into first historical signal data, second historical signal data and third historical signal data, which respectively correspond to the first howling stage, the second howling stage and the third howling stage;

[0076] Using the first historical signal data as first-stage training data;

[0077] Combining the first historical signal data with the second historical signal data to construct second-stage training data;

[0078] The first historical signal data, the second historical signal data and the third historical signal data are combined to construct third-stage training data.

[0079] In a preferred embodiment, in order to adapt to the gradual generation process of howling, this embodiment proposes a training method so that the label also gradually changes with the intensity of the howling. The label here is a masking curve in real or complex form.

[0080] Initial label, that is, the first historical signal data: In the early stage when howling just begins, the label signal may only need to be slightly adjusted, and the clean voice signal remains basically unchanged.

[0081] Mid-term label, i.e., the second historical signal data: As the howling signal gradually increases, the label should also reflect that more suppression work is needed, and a stronger mask may be needed to distinguish between speech and howling signals.

[0082] Complete howling label, that is, the third historical signal data: After the howling is fully formed, the label needs to clearly characterize the difference between the complete feedback signal and the clean signal to ensure that the model can strongly suppress the howling.

[0083] In other words, the training process is no longer a single label, but a serialization of labels.

[0084] First, the initial labels are used for training. At this time, the labels only represent the increase in the energy of some isolated frequencies. After the initial label training, the initial + mid-term training is performed, and finally the initial + mid-term + full training is performed to complete the entire training process. Through this gradually enhanced training label design, the network can better learn the howling characteristics and processing strategies at different stages.

[0085] In a possible implementation, in step S503, the initial signal recognition model is trained by using a multi-learning rate parallel training method, including:

[0086] According to the current learning rate of the initial signal recognition model, a plurality of different training learning rates are generated based on a preset formula and a preset number of parallel trainings;

[0087] Generate a plurality of identical parallel training models according to the current model parameters of the initial signal recognition model, where the number of the parallel training models is the number of the training learning rates;

[0088] Setting different training learning rates for each of the parallel training models, and using the same training data for training, to determine the loss rate of each of the parallel training models on the validation set;

[0089] Determine the optimal parallel training model with the smallest loss rate according to the loss rate of each parallel training model on the validation set;

[0090] According to the model parameters and learning rate of the optimal parallel training model, the current learning rate and current model parameters of the initial signal recognition model are updated, and the next round of multi-learning rate parallel training is performed until the initial signal recognition model meets the preset training termination conditions.

[0091] The embodiment of the present application provides a multi-learning rate parallel model training method. Considering that the application scenario of the embodiment of the present application is to deploy a neural network model in a hearing aid, and the hearing aid is a very extreme embedded device, generally requiring the battery life to be at least 10 hours, the power consumption of its chip cannot be very high, and the computing power that can be provided is not very large. Therefore, the neural network model deployed in the hearing aid is destined not to have too high complexity. Under the premise that the complexity of the model itself is limited, the neural network model applied to the hearing aid should improve the accuracy of model recognition as much as possible during the training process. Traditional algorithms often only use one strategy to update a single learning rate, while the embodiment of the present application abandons the idea of ​​using only one strategy for training in the traditional way, and designs a learning rate update algorithm that matches most scenarios at the same time. Through multiple rounds of iterations and multi-learning rate parallel training, several learning rates and a corresponding number of identical models are generated in each round of iteration for parallel training, and the model parameters and corresponding learning rates with the best training effect are selected to update the initial signal recognition model, and then enter the next round of iteration. The training effect of the model can be improved by increasing the training complexity. At the same time, due to the low complexity of the model itself, this method will not significantly extend the training time, thus achieving a balance between training efficiency and training effect.

[0092] In a preferred embodiment, the multi-learning rate parallel training process is as follows: Figure 3 As shown in the figure, first initialize the learning rate, then select the information passed by the learning rate strategy module and update several learning rates at the same time. Use several learning rates to train the neural network at the same time, get the loss on the validation set, and then select the appropriate learning rate and update strategy based on this loss. The specific steps are as follows:

[0093] 1. Initialize learning rate

[0094] Initially set a basic learning rate (LR). This learning rate will serve as the starting point of the algorithm. The initialization learning rate is generally in the range of (0,1).

[0095] 2. Update learning rate

[0096] The learning rate selection strategy module outputs two pieces of information to this module: one is the base learning rate baseLr, and the other is the accuracy improvement percentage ratio. This embodiment generates a set of new learning rates (LR) based on these two updated base learning rates. These learning rates will be used for the next step of training.

[0097] First, define a ratio r, which is a hyperparameter that can be set by the user in the range of (0,10]. It is generally set to increase slowly from a number less than 1 to 1 or more. This is similar to warmup during training because statistically speaking, the learning rate increases slowly. Then it slowly decreases from 1 to slightly less than 1. This has two purposes: one is to converge slowly, and the other is to prevent training from oscillation. Assume that the number of new learning rates to be generated is N, and N is generally an odd number. The learning rates are arranged from small to large on the horizontal axis, and the base learning rate baseLr is in the middle, that is, the horizontal axis is 0. There are (N-1) / 2 learning rates on each side. The learning rate Lr coordinates greater than baseLr are increased by 1, and the learning rate coordinates less than baseLr are decreased by 1. Then define a ratio rt=ratio*r. Assuming that the coordinate corresponding to a learning rate is x, the value of the learning rate is:

[0098] Lr(x)=baseLr*rt x*k

[0099] Note that for learning rates greater than baseLr, x is positive, and for learning rates less than baseLr, x is negative. The range of x is [-(N-1) / 2, (N-1) / 2]. k is a hyperparameter set by the user and defaults to 1. k is set to avoid oscillation, that is, jumping back and forth around the extreme points during training and the loss value no longer converges.

[0100] 3. Parallel training

[0101] Use the updated learning rate group to train the model in parallel. This means training the neural network with these N different learning rates at the same time to observe the performance of each learning rate on the validation set. Record the loss rate of these N learning rates on the validation set [Loss1, Loss2, ..., LossN].

[0102] 4. Choose the best learning rate

[0103] Select the branch with the smallest loss value from the above losses, lossN=min(loss1,loss2,loss3,loss4,loss5), and record the learning rate lr of the branch corresponding to lossN.

[0104] Calculate the ratio ratio = |lossP / lossN|, where lossP is the loss of the previous round.

[0105] Calculate the base learning rate baseLr = lr*ratio. This reflects the change in the learning rate adaptation loss ratio.

[0106] Save the current loss lossP = lossN.

[0107] Feed baseLr and ratio back to the learning rate update module to continue updating the learning rate and perform the next round of parallel training.

[0108] This cycle will continue until a predetermined training termination condition is reached (e.g. the loss value no longer decreases or the maximum number of iterations).

[0109] Further, in step S4, the howling suppression of the input signal is performed using a preset howling suppression method according to the labels corresponding to the respective signals and the howling stage of the hearing aid, including:

[0110] When the howling stage is the first stage, signal suppression is performed on the signal frame with the howling tag in the input signal.

[0111] When the howling stage is the second stage, the frequency characteristics of the speech signal and the frequency characteristics of the howling signal are determined according to each of the labels; the corresponding speech frequency band and the howling frequency band are divided according to the frequency characteristics of the speech signal and the frequency characteristics of the howling signal; and the howling frequency band is frequency suppressed while ensuring the integrity of the speech frequency band.

[0112] When the howling stage is the third stage, the signal spectrum of the hearing aid is adjusted to a preset signal spectrum.

[0113] The embodiment of the present application provides a method for howling suppression based on different howling stages. When the howling stage is the first stage, the embodiment mainly suppresses it slightly according to the howling label to protect the integrity of the voice and prevent excessive compression. When the howling stage is the second stage, the embodiment further strengthens the suppression and performs frequency band suppression on the entire howling frequency band. When the howling stage is the third stage, the howling has been fully formed and obviously affects the voice signal. At this time, in order to prevent further feedback of the howling and deterioration of the sound quality, the integrity of the voice signal is no longer considered, but the entire signal spectrum is adjusted to achieve strong howling suppression. The embodiment of the present application achieves targeted processing of different howling by judging the howling stage, further improving the howling suppression effect of the hearing aid and the user experience.

[0114] In a preferred embodiment, after the model training is completed, the inference phase is entered.

[0115] The model reasoning process is similar to the common reasoning process. First, the input signal is divided into several time frames and input into the model frame by frame for processing. The model will output corresponding labels to indicate which part of the signal is speech and which part is howling. Since our training process takes into account the growth process of howling intensity, the model will also automatically identify different howling according to the three stages: early, middle, and complete.

[0116] 1. Initial howling stage

[0117] Symptoms: When the howling begins, the signal shows signs of frequency increase, but it is not fully formed.

[0118] Processing: The model automatically identifies the early stage of howling and performs mild suppression to protect the integrity of the voice and prevent over-compression. At the same time, it performs preliminary identification of the howling source. If the model fails to suppress the howling in the early stage, the howling energy will further increase and enter the middle stage.

[0119] 2. Mid-term howling stage

[0120] Symptoms: The howling is quite obvious, but can still be controlled within a reasonable range.

[0121] Processing: Strengthen the suppression strength, suppress the howling, and maintain the main frequency characteristics of the voice. Similarly, if the howling is not suppressed at this stage, it will naturally enter the full howling stage.

[0122] 3. Complete howling stage

[0123] Significance: Howling has fully developed and significantly affects the speech signal.

[0124] Processing: Strongly suppress this frequency band, possibly adjust the entire signal spectrum to prevent howling feedback and avoid sound quality degradation.

[0125] Compared with the traditional reasoning process, this embodiment trains the entire howling process step by step during the reasoning process, so the model can dynamically adjust the suppression strength according to different stage labels to achieve intelligent howling processing.

[0126] Embodiment 2:

[0127] like Figure 4 As shown, accordingly, embodiment 2 provides a howling suppression system for a hearing aid, including an acquisition module 10, a signal segmentation module 20, a signal recognition module 30, a howling suppression module 40 and a model building module 50;

[0128] Wherein, the acquisition module 10 is used to acquire an input signal of a preset length;

[0129] The signal segmentation module 20 is used to divide the input signal into a number of signal frames based on a time sequence;

[0130] The signal recognition module 30 is used to input each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label includes a speech label and a howling label;

[0131] The howling suppression module 40 is used to suppress the howling of the input signal using a preset howling suppression method according to the labels corresponding to each of the signals and the howling stage of the hearing aid;

[0132] The model building module 50 is used to build a neural network model and perform multi-stage training based on historical signal data of each howling stage to obtain the signal recognition model.

[0133] Furthermore, the model construction module 50 includes an initial model construction unit 501, a training data construction unit 502 and a training unit 503;

[0134] Wherein, the initial model building unit 501 is used to build an initial signal recognition model based on a neural network model;

[0135] The training data construction unit 502 is used to combine the historical signal data into training data of different stages according to different howling stages;

[0136] The training unit 503 is used to perform multi-stage training on the initial signal recognition model according to the training data of the different stages to obtain the signal recognition model, wherein in the model training process of each stage, the initial signal recognition model is trained using a multi-learning rate parallel training method.

[0137] Furthermore, the training data construction unit 502 combines the historical signal data into training data of different stages according to different howling stages, including:

[0138] According to different howling stages, each of the historical signal data is divided into first historical signal data, second historical signal data and third historical signal data, which respectively correspond to the first howling stage, the second howling stage and the third howling stage;

[0139] Using the first historical signal data as first-stage training data;

[0140] Combining the first historical signal data with the second historical signal data to construct second-stage training data;

[0141] The first historical signal data, the second historical signal data and the third historical signal data are combined to construct third-stage training data.

[0142] Furthermore, the training unit 503 trains the initial signal recognition model using a multi-learning rate parallel training method, including:

[0143] According to the current learning rate of the initial signal recognition model, a plurality of different training learning rates are generated based on a preset formula and a preset number of parallel trainings;

[0144] Generate a plurality of identical parallel training models according to the current model parameters of the initial signal recognition model, where the number of the parallel training models is the number of the training learning rates;

[0145] Setting different training learning rates for each of the parallel training models, and using the same training data for training, to determine the loss rate of each of the parallel training models on the validation set;

[0146] Determine the optimal parallel training model with the smallest loss rate according to the loss rate of each parallel training model on the validation set;

[0147] According to the model parameters and learning rate of the optimal parallel training model, the current learning rate and current model parameters of the initial signal recognition model are updated, and the next round of multi-learning rate parallel training is performed until the initial signal recognition model meets the preset training termination conditions.

[0148] Furthermore, the howling suppression module 40 uses a preset howling suppression method to suppress the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid, including:

[0149] When the howling stage is the first stage, signal suppression is performed on the signal frame with the howling tag in the input signal.

[0150] When the howling stage is the second stage, the frequency characteristics of the speech signal and the frequency characteristics of the howling signal are determined according to each of the labels; the corresponding speech frequency band and the howling frequency band are divided according to the frequency characteristics of the speech signal and the frequency characteristics of the howling signal; and the howling frequency band is frequency suppressed while ensuring the integrity of the speech frequency band.

[0151] When the howling stage is the third stage, the signal spectrum of the hearing aid is adjusted to a preset signal spectrum.

[0152] The more detailed working principle and step flow of this embodiment can refer to, but are not limited to, the relevant records of Embodiment 1.

[0153] The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A howling suppression method for a hearing aid, characterized in that: include: Obtaining an input signal of a preset length; Dividing the input signal into a plurality of signal frames based on a time sequence; Inputting each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label includes a speech label and a howling label; According to the labels corresponding to the signals and the howling stage of the hearing aid, using a preset howling suppression method to suppress the input signal; The signal recognition model is constructed based on a neural network model and is obtained through multi-stage training based on historical signal data of each howling stage.

2. A howling suppression method for a hearing aid as claimed in claim 1, characterized in that: The signal recognition model is constructed based on the neural network model and is obtained by multi-stage training based on historical signal data of each howling stage, including: Construct an initial signal recognition model based on a neural network model; According to different howling stages, the historical signal data are combined into training data of different stages; According to the training data of the different stages, the initial signal recognition model is trained in multiple stages to obtain the signal recognition model, wherein in the model training process of each stage, the initial signal recognition model is trained using a multi-learning rate parallel training method.

3. A howling suppression method for a hearing aid as claimed in claim 2, characterized in that: Combining the historical signal data into training data of different stages according to different howling stages includes: According to different howling stages, each of the historical signal data is divided into first historical signal data, second historical signal data and third historical signal data, which respectively correspond to the first howling stage, the second howling stage and the third howling stage; Using the first historical signal data as first-stage training data; Combining the first historical signal data with the second historical signal data to construct second-stage training data; The first historical signal data, the second historical signal data and the third historical signal data are combined to construct third-stage training data.

4. The howling suppression method for a hearing aid as claimed in claim 2, characterized in that: The method of training the initial signal recognition model by using multiple learning rate parallel training includes: According to the current learning rate of the initial signal recognition model, a plurality of different training learning rates are generated based on a preset formula and a preset number of parallel trainings; Generate a plurality of identical parallel training models according to the current model parameters of the initial signal recognition model, where the number of the parallel training models is the number of the training learning rates; Setting different training learning rates for each of the parallel training models, and using the same training data for training, to determine the loss rate of each of the parallel training models on the validation set; Determine the optimal parallel training model with the smallest loss rate according to the loss rate of each parallel training model on the validation set; According to the model parameters and learning rate of the optimal parallel training model, the current learning rate and current model parameters of the initial signal recognition model are updated, and the next round of multi-learning rate parallel training is performed until the initial signal recognition model meets the preset training termination conditions.

5. The howling suppression method for a hearing aid according to claim 1, characterized in that: The step of using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid comprises: When the howling stage is the first stage, signal suppression is performed on the signal frame with the howling tag in the input signal.

6. The howling suppression method for a hearing aid according to claim 1, characterized in that: The step of using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid includes: When the howling stage is the second stage, determining the frequency characteristics of the speech signal and the frequency characteristics of the howling signal according to each of the labels; Dividing a corresponding speech frequency band and a howling frequency band according to the frequency characteristics of the speech signal and the frequency characteristics of the howling signal; Under the premise of ensuring the integrity of the voice frequency band, the howling frequency band is suppressed.

7. The howling suppression method for a hearing aid according to claim 1, characterized in that: The step of using a preset howling suppression method to suppress howling on the input signal according to the labels corresponding to the signals and the howling stage of the hearing aid includes: When the howling stage is the third stage, the signal spectrum of the hearing aid is adjusted to a preset signal spectrum.

8. A howling suppression system for a hearing aid, characterized in that: It includes an acquisition module, a signal segmentation module, a signal recognition module, a howling suppression module and a model building module; Wherein, the acquisition module is used to acquire an input signal of a preset length; The signal segmentation module is used to divide the input signal into a plurality of signal frames based on a time sequence; The signal recognition module is used to input each of the signal frames into a preset signal recognition model in sequence, so that the signal recognition model recognizes each of the signal frames, outputs a label corresponding to each of the signal frames, and determines the howling stage of the hearing aid, wherein the label includes a speech label and a howling label; The howling suppression module is used to suppress the howling of the input signal using a preset howling suppression method according to the labels corresponding to each of the signals and the howling stage of the hearing aid; The model building module is used to build a neural network model and perform multi-stage training based on historical signal data of each howling stage to obtain the signal recognition model.

9. The howling suppression system for a hearing aid according to claim 8, characterized in that: The model construction module includes an initial model construction unit, a training data construction unit and a training unit; Wherein, the initial model building unit is used to build an initial signal recognition model based on a neural network model; The training data construction unit is used to combine the historical signal data into training data of different stages according to different howling stages; The training unit is used to perform multi-stage training on the initial signal recognition model according to the training data of the different stages to obtain the signal recognition model, wherein in the model training process of each stage, the initial signal recognition model is trained using a multi-learning rate parallel training method.

10. The howling suppression system for a hearing aid according to claim 9, characterized in that: The training unit trains the initial signal recognition model using a multi-learning rate parallel training method, including: According to the current learning rate of the initial signal recognition model, a plurality of different training learning rates are generated based on a preset formula and a preset number of parallel trainings; Generate a plurality of identical parallel training models according to the current model parameters of the initial signal recognition model, where the number of the parallel training models is the number of the training learning rates; Setting different training learning rates for each of the parallel training models, and using the same training data for training, to determine the loss rate of each of the parallel training models on the validation set; Determine the optimal parallel training model with the smallest loss rate according to the loss rate of each parallel training model on the validation set; According to the model parameters and learning rate of the optimal parallel training model, the current learning rate and current model parameters of the initial signal recognition model are updated, and the next round of multi-learning rate parallel training is performed until the initial signal recognition model meets the preset training termination conditions.

Images