Hearing aid mode switching method and system

Abstract

The hearing aid mode switching method comprises the following steps: collecting a sound signal of a current environment; the sound signals are recognized, recognition results are obtained, the recognition results comprise a first recognition result and a second recognition result, the first recognition result shows that human voice exists, and the second recognition result shows that human voice does not exist; in response to the identification result, selecting an operation mode corresponding to the hearing aid for starting, and when the identification result is the first identification result, processing the sound signal by using a human voice feedback mode; and if the identification result is a second identification result, processing the sound signal by using an omnidirectional mode. According to the hearing aid, the overall power consumption of the hearing aid can be remarkably reduced, the service life of the battery is prolonged, meanwhile, the integrity of original functions is kept, a user can enjoy self-adaptive listening experience without manual intervention, the problem of frequent battery replacement is directly solved, and the use convenience and the energy efficiency ratio are improved.

Description

Technical Field

[0001] This invention relates to the field of hearing aid technology, and in particular to a hearing aid mode switching method and system. Background Technology

[0002] As a miniature amplification device, hearing aids can amplify sounds that are otherwise difficult for users to perceive, and use their residual hearing to transmit sound signals to the auditory center of the brain, thereby enabling the perception of sound and effectively helping hearing-impaired users improve their listening experience.

[0003] Currently, most hearing aids on the market support multiple operating modes, such as voice feedback mode (sound is transmitted to the speaker via a DAC and PA power amplifier) ​​and omnidirectional mode (equalizing sound signals from all directions). To ensure smooth switching between different modes and improve the user experience, even if certain functions are not activated, their corresponding modules usually remain operational. For example, in omnidirectional mode, the DAC and multiple signal channels remain continuously active, which undoubtedly increases the overall power consumption of the device significantly, leading to faster battery drain and requiring users to replace batteries frequently. Therefore, how to achieve intelligent coordination and dynamic management of different operating modes while retaining the original functions has become a critical issue that urgently needs to be addressed in current hearing aid technology. Summary of the Invention

[0004] To address the aforementioned shortcomings, the present invention aims to provide a hearing aid mode switching method and system that enables automatic switching between different operating modes and extends the battery life of the hearing aid.

[0005] To achieve this objective, the present invention adopts the following technical solution: a hearing aid mode switching method, comprising the following steps: Step S1: Collect the sound signal of the current environment; Step S2: Recognize the sound signal and obtain the recognition result, wherein the recognition result includes a first recognition result and a second recognition result, wherein the first recognition result indicates that a human voice is present and the second recognition result indicates that a human voice is not present; Step S3: In response to the recognition result, select the corresponding operating mode of the hearing aid to start it. When the recognition result is the first recognition result, the human voice feedback mode is used to process the sound signal. If the recognition result is the second recognition result, then the omnidirectional mode is used to process the sound signal.

[0006] Preferably, the specific steps of step S2 are as follows: Step S21: Perform noise reduction, filtering, and frame segmentation on the collected sound signal of the current environment to obtain preprocessed data; Step S22: Extract features from the preprocessed data to obtain the first feature; Step S23: Input the first feature into the pre-trained classification model, obtain the probability of human voice in the sound signal of the current environment, and determine whether the probability of human voice is greater than the probability threshold. If it is greater, it is judged as the first recognition result; if it is less, it is judged as the second recognition result.

[0007] Preferably, the rules for frame division processing in step S21 are as follows: The audio signal is processed in frames using different frame rates based on the current operating mode, with the frame rate of the omnidirectional mode being higher than that of the human voice feedback mode.

[0008] Preferably, the specific steps of step S22 are as follows: Step S221: For each frame of audio signal, calculate the number of times its waveform crosses the zero-level axis to obtain the short-time zero-crossing value; Step S222: When the current operating mode is omnidirectional mode, perform absolute value summation on each frame of sound signal to obtain short-time energy; When the current operating mode is human voice feedback mode, the sum of the squares of the absolute values ​​of each frame of sound signal is calculated to obtain the short-time energy. Step S223: Perform a Fast Fourier Transform on each frame of the audio signal to obtain its spectrum; The spectrum is input into a set of Mel-scale filters to obtain the first sensing; After taking the logarithm of the first perception, a discrete cosine transform is performed to obtain the Mel frequency cepstral coefficients. Step S224: Calculate the weighted average value of each frame of audio signal to obtain the spectral centroid; Step S225: Combine the time-crossing zero value, short-time energy, Mel frequency cepstral coefficients, and spectral centroid to obtain the preliminary characteristics; Step S226: Set up a lightweight gating network, obtain gating weights through the gating network, and perform weighted processing on the preliminary features through the gating weights to obtain weighted features; Step S227: Input the weighted features into the multi-scale causal convolution module, obtain short-term, medium-term and long-term multi-scale acoustic features through convolution kernels of different sizes, and output the multi-scale acoustic features in the channel dimension for concatenation to obtain the processed features; Step S228: Pool the processed features in the time dimension and extract multiple global features, including mean, variance, maximum and minimum values, and fuse the global features with the preliminary features to obtain the first feature.

[0009] Preferably, the gating weight calculation formula is as follows: ; Where σ is the Sigmoid function. and For learning parameters; The weighted features are obtained by multiplying the gating weights element-wise with the preliminary features.

[0010] Preferably, the multi-scale causal convolution module consists of three parallel convolutional branches and a feature fusion layer; The first convolutional branch has a kernel size of 1×1 and an output channel number D1 of 16~32. Its mathematical expression is as follows: ; The second convolution branch has a kernel size of 3×1, an output channel number D2 of 32~64, and is padded with two zero vectors on the left. Its mathematical expression is as follows: ; The third convolution branch has a kernel size of 5×1, an output channel number D3 of 64~128, and is padded with 4 zero vectors on the left. Its mathematical expression is as follows: ; in The weights of the first, second, and third convolutional branches are respectively. The weighted feature at time t, σ represents the constant coefficients of the first, second, and third convolution branches, respectively, and σ is the Sigmoid function.

[0011] A hearing aid mode switching system, using the aforementioned hearing aid mode switching method, includes a data collection module, a recognition module, and a mode control module; The collection module is used to collect sound signals from the current environment; The recognition module is used to recognize the sound signal and obtain the recognition result, wherein the recognition result includes a first recognition result and a second recognition result, wherein the first recognition result indicates that a human voice is present and the second recognition result indicates that a human voice is not present; The mode control module is used to select the corresponding operating mode of the hearing aid in response to the recognition result. When the recognition result is the first recognition result, the human voice feedback mode is used to process the sound signal. If the recognition result is the second recognition result, then the omnidirectional mode is used to process the sound signal.

[0012] Preferably, the recognition module includes a preprocessing submodule, a feature extraction submodule, and a model input submodule; The preprocessing submodule is used to perform noise reduction, filtering, and frame segmentation on the collected sound signals of the current environment to obtain preprocessed data; The feature extraction submodule is used to extract features from the preprocessed data to obtain the first feature; The model input submodule is used to input the first feature into the pre-trained classification model, obtain the probability of human voice in the sound signal of the current environment, and determine whether the probability of human voice is greater than the probability threshold. If it is greater, it is judged as the first recognition result; if it is less than, it is judged as the second recognition result.

[0013] Preferably, the feature extraction submodule includes a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit, a seventh unit, and an eighth unit; The first unit is used to calculate the number of times the waveform of each frame of audio signal crosses the zero-level axis to obtain the short-time zero-crossing value; The second unit is used to perform an absolute sum operation on each frame of sound signal to obtain short-time energy when the current operating mode is omnidirectional mode; When the current operating mode is human voice feedback mode, the sum of the squares of the absolute values ​​of each frame of sound signal is calculated to obtain the short-time energy. The third unit is used to perform a fast Fourier transform on each frame of audio signal to obtain its spectrum; The spectrum is input into a set of Mel-scale filters to obtain the first sensing; After taking the logarithm of the first perception, a discrete cosine transform is performed to obtain the Mel frequency cepstral coefficients. The fourth unit is used to calculate the weighted average value of each frame of audio signal and obtain the spectral centroid; The fifth unit is used to combine the time-crossing zero value, short-time energy, Mel frequency cepstral coefficients, and spectral centroid to obtain the preliminary characteristics; The sixth unit is used to set up a lightweight gating network, obtain gating weights through the gating network, and perform weighted processing on the preliminary features through the gating weights to obtain weighted features; The seventh unit is used to input weighted features into the multi-scale causal convolution module, obtain short-term, medium-term and long-term multi-scale acoustic features through convolution kernels of different sizes, and output the multi-scale acoustic features in the channel dimension for concatenation to obtain the processed features; The eighth unit is used to pool the processed features in the time dimension and simultaneously extract multiple global features such as mean, variance, maximum and minimum values. The global features are then fused with the preliminary features to obtain the first feature.

[0014] One of the above technical solutions has the following advantages or beneficial effects: This application can significantly reduce the overall power consumption of hearing aids, extend battery life, and maintain the integrity of the original functions. Users can enjoy an adaptive listening experience without manual intervention, which directly solves the problem of frequent battery replacement and improves ease of use and energy efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart of one embodiment of the method of the present invention.

[0016] Figure 2 This is a schematic diagram of the structure of one embodiment of the system of the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] In the description of embodiments of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] like Figures 1-2 As shown, a hearing aid mode switching method includes the following steps: Step S1: Collect the sound signal of the current environment; Step S2: Recognize the sound signal and obtain the recognition result, wherein the recognition result includes a first recognition result and a second recognition result, wherein the first recognition result indicates that a human voice is present and the second recognition result indicates that a human voice is not present; Step S3: In response to the recognition result, select the corresponding operating mode of the hearing aid to start it. When the recognition result is the first recognition result, the human voice feedback mode is used to process the sound signal. If the recognition result is the second recognition result, then the omnidirectional mode is used to process the sound signal.

[0021] To address the high power consumption issue of existing hearing aids, this invention first acquires and analyzes sound signals to obtain sound signal recognition results. Based on these results, it automatically selects between activating a human voice feedback mode or an omnidirectional mode. Since the human voice feedback mode is specifically optimized for amplifying human voices, while the omnidirectional mode is suitable for capturing ambient sounds, the mode switching is triggered by sound recognition. This ensures that the relevant modules for the human voice feedback mode (such as the DAC and PA power amplifier) ​​are activated only when human voices are detected. In environments without human voices, only the omnidirectional mode operates, thus avoiding unnecessary continuous operation of these modules.

[0022] This application can significantly reduce the overall power consumption of hearing aids, extend battery life, and maintain the integrity of the original functions. Users can enjoy an adaptive listening experience without manual intervention, which directly solves the problem of frequent battery replacement and improves ease of use and energy efficiency.

[0023] Preferably, the specific steps of step S2 are as follows: Step S21: Perform noise reduction, filtering and frame segmentation on the collected sound signal of the current environment to obtain preprocessed data; clean the signal and divide it into short-term stable analysis units to lay the foundation for subsequent accurate feature extraction.

[0024] Step S22: Extract features from the preprocessed data to obtain the first features; extract key information parameters that can significantly distinguish between human voices and non-human voices from the preprocessed audio frames. These first features are the basis for building an efficient classification model.

[0025] Step S23: Input the first feature into the pre-trained classification model to obtain the probability of human voice in the current environment's sound signal. Determine whether the human voice probability is greater than a probability threshold. If it is greater, it is judged as the first recognition result; if it is less, it is judged as the second recognition result. The pre-trained classification model intelligently recognizes and calculates the probability of these features, and makes a binary decision by comparing it with the probability threshold. This ensures the automation, objectivity, and high reliability of the mode switching judgment.

[0026] Preferably, the rules for frame division processing in step S21 are as follows: The audio signal is processed in frames using different frame rates based on the current operating mode, with the frame rate of the omnidirectional mode being higher than that of the human voice feedback mode.

[0027] In the use of hearing aids, in order to better recognize human voices and improve communication, this invention uses a higher frame rate to process the sound signal in omnidirectional mode by frame segmentation to obtain more data, thereby improving the sensitivity of voice recognition in omnidirectional mode and enhancing the sensitivity of mode switching.

[0028] Preferably, the specific steps of step S22 are as follows: Step S221: For each frame of audio signal, calculate the number of times its waveform crosses the zero-level axis to obtain the short-time zero-crossing value; Step S222: When the current operating mode is omnidirectional mode, perform absolute value summation on each frame of sound signal to obtain short-time energy; When the current operating mode is human voice feedback mode, the sum of the squares of the absolute values ​​of each frame of sound signal is calculated to obtain the short-time energy. Since the omnidirectional mode is already set to use a higher frame rate for frame processing, the amount of data in omnidirectional mode is relatively large. In order to reduce the amount of computation and improve the system response speed, when acquiring short-time energy, only the absolute value sum is performed on each frame of audio signal, and the absolute value square sum is not used, thus reducing the computational burden of short-time energy.

[0029] It's worth noting that in subsequent processing, only character representations of different current operating modes need to be added to distinguish them. Finally, the data needs to be input into different classification models to eliminate the impact of varying short-term energy values ​​across different operating modes. For example, there are first-class and second-class classification models. The first-class model primarily handles the first feature in the omnidirectional mode, while the second-class model handles the first feature in the human voice feedback mode.

[0030] Step S223: Perform a Fast Fourier Transform on each frame of the audio signal to obtain its spectrum; The spectrum is input into a set of Mel-scale filters to obtain the first sensing; After taking the logarithm of the first perception, a discrete cosine transform is performed to obtain the Mel frequency cepstral coefficients. Step S224: Calculate the weighted average value of each frame of audio signal to obtain the spectral centroid; Step S225: Combine the time-crossing zero value, short-time energy, Mel frequency cepstral coefficients, and spectral centroid to obtain the preliminary characteristics; Step S226: Set up a lightweight gating network, obtain gating weights through the gating network, and perform weighted processing on the preliminary features through the gating weights to obtain weighted features; Step S227: Input the weighted features into the multi-scale causal convolution module, obtain short-term, medium-term and long-term multi-scale acoustic features through convolution kernels of different sizes, and output the multi-scale acoustic features in the channel dimension for concatenation to obtain the processed features; Step S228: Pool the processed features in the time dimension and extract multiple global features, including mean, variance, maximum and minimum values, and fuse the global features with the preliminary features to obtain the first feature.

[0031] Currently, the data input into classification models is generally preliminary features, as described in CN115312067A. This preliminary feature is obtained by combining short-time zero-crossing values, short-time energy, Mel-frequency cepstral coefficients, and spectral centroids, and then inputting these preliminary features into the classification model. Although noise reduction is performed through preprocessing before obtaining the preliminary features, noise still exists within them. Direct use of these preliminary features can lead to misidentification scenarios. For example, in omnidirectional mode, if a sound in the environment resembles a human voice, it can cause misidentification, resulting in frequent switching of operating modes.

[0032] Therefore, in this invention, after obtaining the preliminary features, the preliminary features are not directly input into the classification model for recognition. Instead, the preliminary features are further processed, such as by setting a lightweight gating network. After passing through the lightweight gating network, the preliminary features can be further noise-reduced. The lightweight gating network does not use a fixed PCA transform during the processing, which can better adapt to the changing acoustic environment. After obtaining weighted features through a lightweight gating network, the weighted features are processed by a multi-scale causal convolution module. This aims to capture contextual information at different time scales from the weighted feature sequence. For example, the first convolution branch has the fewest kernels and is sensitive to abrupt changes at the level of a single sampling point. The second convolution branch has a moderate number of kernels and can be used to identify and capture rapid changes in phonemes over a short period of time. The third convolution branch has a large number of kernels and can be used to identify acoustic patterns over a longer period of time, such as changes in pitch within syllables. By capturing short-term changes at the syllable or word level through the second and third convolution branches, human voice confirmation can be performed based on its persistence, avoiding false triggering due to transient noise.

[0033] Finally, pooling and fusing global statistical features in the time dimension aim to aggregate a series of frame-level local feature sequences into a stable vector that can represent the overall characteristics of the sound within the current analysis window. This first feature, which integrates statistical measures such as mean and variance, provides the most general and discriminative input for subsequent classification models.

[0034] Preferably, the gating weight calculation formula is as follows: ; Where σ is the Sigmoid function. and For learning parameters; The weighted features are obtained by multiplying the gating weights element-wise with the preliminary features.

[0035] Preferably, the multi-scale causal convolution module consists of three parallel convolutional branches and a feature fusion layer; The first convolutional branch has a kernel size of 1×1 and an output channel number D1 of 16~32. Its mathematical expression is as follows: ; The second convolution branch has a kernel size of 3×1, an output channel number D2 of 32~64, and is padded with two zero vectors on the left. Its mathematical expression is as follows: ; The third convolution branch has a kernel size of 5×1, an output channel number D3 of 64~128, and is padded with 4 zero vectors on the left. Its mathematical expression is as follows: ; in The weights of the first, second, and third convolutional branches are respectively. The weighted feature at time t, σ represents the constant coefficients of the first, second, and third convolution branches, respectively, and σ is the Sigmoid function.

[0036] A hearing aid mode switching system, using the aforementioned hearing aid mode switching method, includes a data collection module, a recognition module, and a mode control module; The collection module is used to collect sound signals from the current environment; The recognition module is used to recognize the sound signal and obtain the recognition result, wherein the recognition result includes a first recognition result and a second recognition result, wherein the first recognition result indicates that a human voice is present and the second recognition result indicates that a human voice is not present; The mode control module is used to select the corresponding operating mode of the hearing aid in response to the recognition result. When the recognition result is the first recognition result, the human voice feedback mode is used to process the sound signal. If the recognition result is the second recognition result, then the omnidirectional mode is used to process the sound signal.

[0037] Preferably, the recognition module includes a preprocessing submodule, a feature extraction submodule, and a model input submodule; The preprocessing submodule is used to perform noise reduction, filtering, and frame segmentation on the collected sound signals of the current environment to obtain preprocessed data; The feature extraction submodule is used to extract features from the preprocessed data to obtain the first feature; The model input submodule is used to input the first feature into the pre-trained classification model, obtain the probability of human voice in the sound signal of the current environment, and determine whether the probability of human voice is greater than the probability threshold. If it is greater, it is judged as the first recognition result; if it is less than, it is judged as the second recognition result.

[0038] Preferably, the feature extraction submodule includes a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit, a seventh unit, and an eighth unit; The first unit is used to calculate the number of times the waveform of each frame of audio signal crosses the zero-level axis to obtain the short-time zero-crossing value; The second unit is used to perform an absolute sum operation on each frame of sound signal to obtain short-time energy when the current operating mode is omnidirectional mode; When the current operating mode is human voice feedback mode, the sum of the squares of the absolute values ​​of each frame of sound signal is calculated to obtain the short-time energy. The third unit is used to perform a fast Fourier transform on each frame of audio signal to obtain its spectrum; The spectrum is input into a set of Mel-scale filters to obtain the first sensing; After taking the logarithm of the first perception, a discrete cosine transform is performed to obtain the Mel frequency cepstral coefficients. The fourth unit is used to calculate the weighted average value of each frame of audio signal and obtain the spectral centroid; The fifth unit is used to combine the time-crossing zero value, short-time energy, Mel frequency cepstral coefficients, and spectral centroid to obtain the preliminary characteristics; The sixth unit is used to set up a lightweight gating network, obtain gating weights through the gating network, and perform weighted processing on the preliminary features through the gating weights to obtain weighted features; The seventh unit is used to input weighted features into the multi-scale causal convolution module, obtain short-term, medium-term and long-term multi-scale acoustic features through convolution kernels of different sizes, and output the multi-scale acoustic features in the channel dimension for concatenation to obtain the processed features; The eighth unit is used to pool the processed features in the time dimension and simultaneously extract multiple global features such as mean, variance, maximum and minimum values. The global features are then fused with the preliminary features to obtain the first feature.

[0039] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0040] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A hearing aid mode switching method, characterized by, The method comprises the following steps: Step S1: collecting a sound signal of a current environment; Step S2: identifying the sound signal to obtain an identification result, wherein the identification result comprises a first identification result and a second identification result, wherein the first identification result is that there is human voice, and the second identification result is that there is no human voice; Step S3: in response to the identification result, selecting a corresponding operation mode of a hearing aid to start, wherein when the first identification result is identified, a human voice feedback mode is used to process the sound signal; If the identification result is the second identification result, an omnidirectional mode is used to process the sound signal.

2. A hearing aid mode switching method according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S21: performing noise reduction, filtering and frame processing on the collected sound signal of the current environment to obtain preprocessed data; Step S22: performing feature extraction on the preprocessed data to obtain first features; Step S23: inputting the first features into a pre-trained classification model to obtain a human voice probability of the current sound signal of the current environment, and determining whether the human voice probability is greater than a probability threshold, if yes, determining that the first identification result is obtained, if no, determining that the second identification result is obtained.

3. A hearing aid mode switching method according to claim 2, characterized in that, The rules for frame processing in step S21 are as follows: Based on the current operation mode, the sound signal is processed using different frame rates, wherein the frame rate of the omnidirectional mode is greater than that of the human voice feedback mode.

4. A hearing aid mode switching method according to claim 3, characterized in that, The specific steps of step S22 are as follows: Step S221: for each frame of sound signal, calculating the number of times its waveform crosses the zero level axis to obtain a short-time zero-crossing value; Step S222: when the current operation mode is the omnidirectional mode, performing absolute value sum operation on each frame of sound signal to obtain a short-time energy; When the current operation mode is the human voice feedback mode, performing absolute value square sum operation on each frame of sound signal to obtain a short-time energy; Step S223: performing fast Fourier transform on each frame of sound signal to obtain its frequency spectrum; Inputting the frequency spectrum into a group of mel-scale filter banks to obtain first perceptions; Taking the logarithm of the first perceptions and then performing discrete cosine transform to obtain mel-frequency cepstral coefficients; Step S224: performing weighted average value calculation on each frame of sound signal to obtain a spectral centroid; Step S225: combining the short-time zero-crossing value, the short-time energy, the mel-frequency cepstral coefficients and the spectral centroid to obtain the preliminary features; Step S226: setting a lightweight gating network, obtaining gating weights through the gating network, and performing weighted processing on the preliminary features through the gating weights to obtain weighted features; Step S227: inputting the weighted features into a multi-scale causal convolution module, obtaining multi-scale acoustic features of short, medium and long terms through different size convolution kernels, and splicing the multi-scale acoustic features in the channel dimension to obtain processing features; Step S228: performing pooling on the processing features in the time dimension, and simultaneously extracting multiple global features of mean, variance, maximum and minimum, fusing the global features with the preliminary features to obtain the first features.

5. A hearing aid mode switching method according to claim 4, characterized in that, The gating weight calculation formula is as follows: ； where σ is a sigmoid function, and are learning parameters; The weighted feature is obtained by element-wise multiplication of the preliminary feature and the gating weight.

6. A hearing aid mode switching method according to claim 4, characterized in that, The multi-scale causal convolution module is composed of three parallel convolution branches and a feature fusion layer; The first convolution branch has a convolution kernel size of 1*1 and an output channel number D1 of 16-32, The mathematical expression thereof is as follows: ; The second convolution branch is: the convolution kernel size is 3x1, the output channel number D2 is 32-64, and 2 zero vectors are filled on the left side, and the mathematical expression is as follows: ; The third convolution branch is: the convolution kernel size is 5x1, the output channel number D3 is 64-128, and 4 zero vectors are filled on the left side, and the mathematical expression is as follows: ; wherein are weights of the first, second and third convolutional branches, respectively, is a weighted feature at time t, are constant coefficients of the first, second and third convolutional branches, respectively, and σ is a Sigmoid function.

7. A hearing aid mode switching system, characterized by The hearing aid mode switching method according to any one of claims 1-6 comprises a collection module, an identification module, and a mode control module; The collection module is configured to collect a sound signal of a current environment; The identification module is configured to identify the sound signal to obtain an identification result, wherein the identification result comprises a first identification result and a second identification result, wherein the first identification result indicates that there is human voice, and the second identification result indicates that there is no human voice; The mode control module is configured to select a corresponding operation mode of the hearing aid to start in response to the identification result, wherein the sound signal is processed in a human voice feedback mode when the first identification result is identified. If the identification result is the second identification result, the sound signal is processed in an omnidirectional mode.

8. A hearing aid mode switching system according to claim 7, characterized in that The identification module comprises a preprocessing submodule, a feature extraction submodule, and a model input submodule; The preprocessing submodule is configured to perform noise reduction, filtering, and frame processing on the collected sound signal of the current environment to obtain preprocessed data; The feature extraction submodule is configured to extract features from the preprocessed data to obtain first features; The model input submodule is configured to input the first features into a pre-trained classification model to obtain a probability of human voice of the sound signal of the current environment, and determine whether the probability of human voice is greater than a probability threshold, wherein the first identification result is determined if the probability of human voice is greater than the probability threshold, and the second identification result is determined if the probability of human voice is less than the probability threshold.

9. A hearing aid mode switching system according to claim 7, characterized in that The feature extraction submodule comprises a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit, a seventh unit, and an eighth unit; The first unit is configured to calculate a number of times that a waveform of each frame of the sound signal crosses a zero level axis to obtain a short-time zero-crossing value; The second unit is configured to perform an absolute value sum operation on each frame of the sound signal to obtain a short-time energy when the current operation mode is the omnidirectional mode. When the current operation mode is the human voice feedback mode, the second unit is configured to perform an absolute value square sum operation on each frame of the sound signal to obtain a short-time energy. The third unit is configured to perform a fast Fourier transform on each frame of the sound signal to obtain a frequency spectrum thereof. The frequency spectrum is input into a group of mel-scale filter banks to obtain a first perceptual feature. The fourth unit is configured to perform a weighted average value calculation on each frame of the sound signal to obtain a spectral centroid. The fifth unit is configured to combine the short-time zero-crossing value, the short-time energy, the mel-frequency cepstral coefficient, and the spectral centroid to obtain the preliminary feature. The sixth unit is configured to set a lightweight gating network, obtain a gating weight through the gating network, and perform a weighting processing on the preliminary feature through the gating weight to obtain a weighted feature. ​ The seventh unit is configured to input the weighted features into a multi-scale causal convolution module, obtain multi-scale acoustic features of short, medium and long terms through convolution kernels of different sizes, and output the multi-scale acoustic features in a channel dimension to be spliced to obtain processing features; The eighth unit is configured to pool the processing features in a time dimension, extract multiple global features of mean, variance, maximum and minimum values, and fuse the global features with the preliminary features to obtain the first features.

Images