Self-adaptive gain and noise reduction hearing aid based on artificial intelligence

Abstract

The invention relates to the technical field of adaptive signal processing, in particular to an adaptive gain and noise reduction hearing aid based on artificial intelligence. The method comprises the following steps: receiving an acoustic signal through a double-microphone array, separating target voice from environmental noise through time delay inequality, and establishing a space sound field model; converting the audio signal into a time-frequency domain, and extracting an MFCC feature; a CNN-LSTM hybrid network model is constructed, and voice and noise are classified in real time; dynamically adjusting gain values of different frequency bands according to the hearing loss curve and the real-time signal-to-noise ratio; noise in a non-target direction is suppressed, and the spatial selectivity of the voice signal is enhanced; and optimizing model parameters according to the hearing archive of the user, and supporting individual difference adaptation. Acoustic signals are captured through a dual-microphone array, and target voice and environmental noise are distinguished by using delay inequality. A target voice signal is effectively separated, noise from a non-target direction is suppressed, and voice definition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of adaptive signal processing, and more specifically, to a hearing aid with adaptive gain and noise reduction based on artificial intelligence. Background Art

[0002] A hearing aid with adaptive gain and noise reduction based on artificial intelligence is a high-tech hearing aid device. It uses artificial intelligence technology to automatically analyze the characteristics of environmental sounds and dynamically adjust volume enhancement and background noise suppression according to the user's hearing loss situation and personal needs. This hearing aid can intelligently identify sound patterns in different scenarios. For example, it can highlight the conversation sound in a noisy environment and provide a more natural sound experience in a quiet environment, thus effectively improving the wearer's hearing experience and quality of daily life. By continuously learning and adapting to the user's preferences, it can further personalize the service to ensure the best auditory effect for the user in any environment.

[0003] Traditional hearing aids rely on a fixed amplification gain strategy to amplify environmental sounds. Similar to a speaker, they just blindly amplify the sound and are difficult to cope with the noisy environment. Users often have difficulty distinguishing human voices from background noise, resulting in communication difficulties. Moreover, although some current hearing aids have an automatic gain control and an environmental noise detection module, they need to be preset. Similar to having a button, different levels of noise reduction can be selected, but the adaptability is relatively poor because people have individual differences, especially those with different ages, different degrees of hearing loss, and different types of hearing loss.

[0004] In summary, a hearing aid with adaptive gain and noise reduction based on artificial intelligence is provided. Summary of the Invention

[0005] The purpose of the present invention is to provide a hearing aid with adaptive gain and noise reduction based on artificial intelligence to solve the problems in the above-mentioned background art that traditional hearing aids rely on a fixed amplification gain strategy to amplify environmental sounds. Similar to a speaker, they just blindly amplify the sound and are difficult to cope with the noisy environment. Users often have difficulty distinguishing human voices from background noise, resulting in communication difficulties. Moreover, although some current hearing aids have an automatic gain control and an environmental noise detection module, they need to be preset. Similar to having a button, different levels of noise reduction can be selected, but the adaptability is relatively poor because people have individual differences, especially those with different ages, different degrees of hearing loss, and different types of hearing loss.

[0006] To achieve the above object, the present invention aims to provide a hearing aid with adaptive gain and noise reduction based on artificial intelligence, including the following steps:

[0007] S1. Receive acoustic signals through a dual-microphone array, separate the target speech and environmental noise through the time-delay difference, and establish a spatial sound field model;

[0008] S2. Convert the audio signal into the time-frequency domain and extract MFCC features;

[0009] S3. Construct a CNN-LSTM hybrid network model to classify speech and noise in real time;

[0010] S4. Dynamically adjust the gain values of different frequency bands according to the hearing loss curve and the real-time signal-to-noise ratio. Considering individual hearing differences and environmental noise changes, perform time-domain smoothing processing on the gain to optimize the gain values of different frequency bands;

[0011] S5. Suppress the noise in non-target directions to enhance the spatial selectivity of the speech signal;

[0012] S6. Optimize the model parameters according to the user's hearing profile to support adaptation to individual differences.

[0013] As a further improvement of this technical solution, in S1, the specific steps of receiving acoustic signals through a dual-microphone array, separating the target speech and environmental noise through the time-delay difference, and establishing a spatial sound field model are as follows;

[0014] S1.1. The two microphones respectively receive acoustic signals x1 and x2, covering the target sound source and environmental noise;

[0015] S1.2. Calculate the cross-correlation function R 12 (τ) of the two signals, and find the time-delay difference τ corresponding to the maximum correlation value max , which is used to determine the azimuth angle θ of the target sound source;

[0016] S1.3. Based on the time-delay difference, construct a spatial sound field model to distinguish the target speech direction from the environmental noise direction;

[0017] S1.4. S1.4. Use the noise covariance matrix R nn and the steering vector d(θ) to calculate the optimal beam weight w, which is used to enhance the main direction signal and suppress the sidelobe noise.

[0018] As a further improvement of this technical solution, in S1.3, the specific steps of constructing a spatial sound field model based on the time-delay difference to distinguish the target speech direction from the environmental noise direction are as follows:

[0019] S1.31. According to the estimated azimuth angle θ, divide the sound field into the target direction and the noise direction;

[0020] S1.32. Perform frequency-domain analysis on the signal and construct the sound field distribution P(f,θ) in the frequency-angle domain;

[0021] S1.33. Mark the frequency-domain energy in non-target directions as noise components using the time-delay estimation results.

[0022] As a further improvement of this technical solution, in S1.4, using the noise covariance matrix R nn and the steering vector d(θ), calculate the optimal beamforming weight w for enhancing the main-direction signal and suppressing sidelobe noise. The specific steps are as follows:

[0023] S1.41. During the noise-dominated period, calculate the noise covariance matrix R nn ;

[0024] S1.42. Define the steering vector d(θ) according to the target direction θ and the microphone array geometry;

[0025] S1.43. Solve the constrained optimization problem to calculate the optimal beamforming weight w;

[0026] S1.44. Apply the weight to the input signal x(t) to enhance the signal in the target direction.

[0027] As a further improvement of this technical solution, in S2, the specific steps for converting the audio signal into the time-frequency domain and extracting MFCC features are as follows:

[0028] S2.1. Generate the time-frequency matrix Y(f, l) through short-time Fourier transform;

[0029] S2.2. Extract MFCC features based on the time-frequency matrix.

[0030] As a further improvement of this technical solution, the specific steps for constructing a CNN-LSTM hybrid network model to classify speech and noise in real time in S3 are as follows:

[0031] S3.1. Normalize the time-frequency matrix Y(f, l) and convert it into a tensor suitable for network input;

[0032] S3.2. Use the convolutional layer to extract local spectral features;

[0033] S3.3. Reduce the dimension through the ReLU activation function and the pooling layer to retain key features;

[0034] S3.4. Input the output of the CNN into the LSTM layer in time series to learn the temporal dependence relationship between speech and noise;

[0035] S3.5. Control the information flow through the LSTM layer;

[0036] S3.6. The fully connected layer maps the output of the LSTM to class probabilities, and the Softmax layer outputs the speech / noise label P(y = K|x);

[0037] S3.7. Use the cross-entropy loss function L to backpropagate and optimize the parameters.

[0038] As a further improvement of this technical solution, in S4, according to the hearing loss curve and the real-time signal-to-noise ratio, the gain values of different frequency bands are dynamically adjusted. Considering individual hearing differences and environmental noise changes, the specific steps for optimizing the gain values of different frequency bands by performing time-domain smoothing on the gain are as follows:

[0039] S4.1. Use the classification result of S3 to mark the noise-dominated frames and update the noise power spectrum Φ nn (f);

[0040] S4.2. Dynamically adjust the gain according to the difference ΔG(f) between the current signal-to-noise ratio SNR(f) and the target.

[0041] S4.3. Perform wide dynamic range compression.

[0042] As a further improvement of this technical solution, in S4.3, the specific process of performing wide dynamic range compression is as follows:

[0043] S4.31. Map the input signal strength I in (f) to the output gain G out (f);

[0044] S4.32. Perform time-domain smoothing on the gain to avoid sudden changes.

[0045] As a further improvement of this technical solution, in S5, suppress the noise in non-target directions to enhance the spatial selectivity of the speech signal;

[0046] S5.1. Input the noise power spectrum Φ nn (f);

[0047] S5.2. Calculate the frequency-domain filtering coefficient W(f) to suppress the noise components;

[0048] S5.3. Combine the beamforming result of S1 to further attenuate the residual noise in non-target directions.

[0049] As a further improvement of this technical solution, in S6, optimize the model parameters according to the user's hearing profile to support adaptation to individual differences.

[0050] S6.1. Based on the CNN-LSTM model, use the user-specific data D user Fine-tune the classifier parameters

[0051] S6.2. According to the user's real-time feedback on the gain adjustment;

[0052] S6.3. Periodically collect the user's usage environment data, update the noise library and model weights, and achieve long-term personalized adaptation.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. In the hearing aid based on artificial intelligence for adaptive gain and noise reduction, the acoustic signal is captured by a dual microphone array, and the time delay difference is used to distinguish the target speech and environmental noise. The target speech signal is effectively separated, and at the same time, the noise from non-target directions is suppressed, improving speech clarity. The audio signal is transformed from the time domain to the time-frequency domain through the short-time Fourier transform to generate a time-frequency matrix. The extracted MFCC features provide important input data for subsequent classification and processing.

[0055] 2. In the hearing aid based on artificial intelligence for adaptive gain and noise reduction, by constructing a hybrid model combining a convolutional neural network and a long short-term memory network, the deep learning model can accurately distinguish speech and background noise and provide more accurate sound classification results. By integrating the reference gain and the dynamic correction amount, the device can automatically adjust the sound amplification effect according to the user's specific hearing needs and the changes in the surrounding environment, improving the auditory comfort. The speech clarity is significantly improved, the background noise interference is reduced, and a more natural and clear auditory experience is provided. The adaptability of the device to individual hearing loss is improved, and it can also long-term track the user's hearing changes and provide continuous optimized hearing support, greatly enhancing the user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is the overall method flow chart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] Embodiment: Please refer to Figure 1 As shown, this embodiment provides a hearing aid based on artificial intelligence for adaptive gain and noise reduction, including the following steps:

[0059] S1. Receive the acoustic signal through a dual microphone array, separate the target speech and environmental noise through the time delay difference, and establish a spatial sound field model;

[0060] In this example, acoustic signals are received by a dual-microphone array, and the target speech and environmental noise are separated by the time delay difference. The specific steps for establishing a spatial sound field model are as follows;

[0061] S1.1. The two microphones respectively receive acoustic signals x1 and x2, covering the target sound source and environmental noise;

[0062] S1.2. Calculate the cross-correlation function R 12 (τ) of the two signals, and find the time delay difference τ corresponding to the maximum correlation value max for determining the azimuth angle θ of the target sound source;

[0063] Specifically, the calculation process of the cross-correlation function R 12 (τ) is as follows:

[0064]

[0065] where N represents the signal frame length; τ represents the time delay offset; n represents the index of the time sample;

[0066] The time delay difference τ corresponding to the maximum correlation value max is:

[0067]

[0068] The azimuth angle θ of the target sound source is:

[0069]

[0070] where c represents the propagation speed of sound waves in air; d represents the physical distance between the two microphones.

[0071] S1.3. Based on the time delay difference, construct a spatial sound field model to distinguish the target speech direction from the environmental noise direction;

[0072] In this example, the specific steps for constructing a spatial sound field model based on the time delay difference to distinguish the target speech direction from the environmental noise direction are as follows:

[0073] S1.31. According to the estimated azimuth angle θ, divide the sound field into the target direction and the noise direction;

[0074] S1.32. Conduct frequency-domain analysis on the signal and construct the sound field distribution P(f, θ) in the frequency-angle domain;

[0075] Specifically, the construction of the sound field distribution P(f, θ) in the frequency-angle domain is:

[0076]

[0077] where X m (f) represents the spectrum of the mth microphone; τm (θ) represents the time delay of microphone m relative to the reference point; f represents the frequency; j represents the imaginary unit; m represents the microphone number index variable; M represents the number of microphones;

[0078] S1.33. Mark the frequency-domain energy in non-target directions as noise components using the time-delay estimation result.

[0079] S1.4. Use the noise covariance matrix R nn and the steering vector d(θ) to calculate the optimal beamforming weight w for enhancing the signal in the main direction and suppressing the sidelobe noise.

[0080] In this example, the specific steps for calculating the optimal beamforming weight w using the noise covariance matrix R nn and the steering vector d(θ) are as follows:

[0081] S1.41. Calculate the noise covariance matrix R nn ;

[0082] Specifically, the noise covariance matrix R nn is:

[0083]

[0084] where K represents the number of noise frames; n(k) represents the noise spectrum vector of the k-th frame; k represents the frame number index variable; H represents the conjugate transpose;

[0085] S1.42. Define the steering vector d(θ) according to the target direction θ and the geometry of the microphone array;

[0086] Specifically, the steering vector d(θ) is:

[0087]

[0088] where T represents the transpose;

[0089] S1.43. Solve the constrained optimization problem to calculate the optimal beamforming weight w;

[0090] Specifically, the optimal beamforming weight w is:

[0091]

[0092] S1.44. Apply the weight to the input signal x(t) to enhance the signal in the target direction.

[0093] Specifically, the output signal is:

[0094] y(t) = w H x(t);

[0095] Where y(t) represents the output signal after beamforming processing.

[0096] Specifically, in a hearing aid with artificial intelligence-based adaptive gain and noise reduction, the process of receiving acoustic signals through a dual-microphone array and using the time delay difference to separate the target speech from environmental noise to establish a spatial sound field model is crucial. First, the two microphones respectively receive the acoustic signals containing the target sound source and environmental noise, then calculate the cross-correlation function of the two signals, find the time delay difference corresponding to the maximum correlation value, and thus determine the azimuth angle of the target sound source. Next, based on the estimated azimuth angle, the sound field is divided into the target direction and the noise direction, and frequency-domain analysis is performed to construct a sound field distribution map in the frequency-angle domain, identify and mark the noise components in the non-target direction. Finally, the optimal beam weights are calculated using the noise covariance matrix and the steering vector, and these weights are applied to the input signal to enhance the signal in the target direction while suppressing the sidelobe noise. This method can significantly improve speech clarity, effectively reduce background noise interference, provide a more natural and comfortable listening experience for users, and is especially suitable for complex and changing listening environments. Through this precise spatial sound field modeling and beamforming technology, the hearing aid can dynamically adjust in various environments to ensure that users obtain the best hearing support.

[0097] S2. Convert the audio signal into the time-frequency domain and extract MFCC features

[0098] In this example, the specific steps for converting the audio signal into the time-frequency domain and extracting MFCC features are as follows:

[0099] S2.1. Generate the time-frequency matrix Y(f, l) through the short-time Fourier transform;

[0100] Specifically, the time-frequency matrix Y(f, l) is:

[0101]

[0102] Where Y(f, l) represents the time-frequency matrix; θ represents the number of samples in a frame; l represents the time frame index; x(n) represents the time series data of the original signal; represents the window function; represents the hop number;

[0103] S2.2. Extract MFCC features according to the time-frequency matrix.

[0104] Specifically, the MFCC feature extraction is:

[0105] MFCC = DCT[log(|Y(f, l)| 2 ·H mel )];

[0106] Where DCT represents the discrete cosine transform; Hmel Represents a Mel filter bank.

[0107] Specifically, in hearing aids with artificial intelligence-based adaptive gain and noise reduction, the process of converting audio signals into the time-frequency domain and extracting MFCC features is crucial. First, the audio signal is converted into a time-frequency matrix through the short-time Fourier transform to capture the characteristics of the signal at different times and frequencies. Then, the Mel filter bank is used to process the time-frequency matrix, and MFCC features are extracted through taking logarithms and discrete cosine transform. These features can effectively represent the way the human auditory system perceives sound. This method not only simplifies complex audio signals but also highlights the information most useful for speech recognition and noise suppression. Finally, the extracted MFCC features are used for subsequent model training and real-time classification, helping the device more accurately distinguish speech from noise, optimize the gain adjustment strategy, and provide a clearer and more comfortable auditory experience. This enables the hearing aid to dynamically adapt to changes in various environments, significantly improving the user's hearing quality and quality of life.

[0108] S3. Build a CNN-LSTM hybrid network model to classify speech and noise in real time;

[0109] In this example, the specific steps to build a CNN-LSTM hybrid network model to classify speech and noise in real time are as follows:

[0110] S3.1. Normalize the time-frequency matrix Y(f, l) and convert it into a tensor suitable for network input;

[0111] Specifically, the normalization of the time-frequency matrix Y(f, l) is as follows:

[0112]

[0113] In the formula, represents the normalized time-frequency matrix; μ f,l represents the mean of the time-frequency matrix; σ f,l represents the standard deviation of the time-frequency matrix;

[0114] The normalized time-frequency matrices are concatenated into a three-dimensional tensor T * :

[0115] T * ∈R T×F×C ;

[0116] In the formula, T represents the number of time frames; F represents the number of frequency points; C represents the number of channels;

[0117] S3.2. Use the convolutional layer to extract local spectral features;

[0118] Specifically, using the convolutional layer to extract local spectral features is as follows:

[0119]

[0120] Wherein, Z(i, I) represents the value of the feature map of the convolution output at the position (i, I); W represents the convolution kernel weight matrix; b conv represents the bias term of the convolutional layer; P represents the height of the convolution kernel; Q represents the width of the convolution kernel; p represents the index variable of the height of the convolution kernel; q represents the index variable of the width of the convolution kernel; (i, I) represents the position of the feature map;

[0121] S3.3. Reduce the dimension through the ReLU activation function and the pooling layer, and retain the key features;

[0122] Specifically, the ReLU activation function is:

[0123] A(i, I) = max(0, Z(i, I));

[0124] Wherein, A(i, I) represents the value at the position (i, I) after passing through the ReLU activation function;

[0125] The dimensionality reduction of the pooling layer is:

[0126] D(i, I) = max u∈[0,U),v∈[0,V) A(i·S + u, I·S + v);

[0127] Wherein, D(i, I) represents the value at the position (i, I) after dimensionality reduction by the pooling layer; S represents the stride; U represents the height of the pooling window; V represents the width of the pooling window; u represents the index variable of the height of the pooling window; v represents the index variable of the width of the pooling window;

[0128] S3.4. Input the CNN output into the LSTM layer according to the time series to learn the temporal dependence relationship between speech and noise;

[0129] S3.5. Control the information flow through the LSTM layer;

[0130] S3.6. The fully connected layer maps the LSTM output to class probabilities, and the Softmax layer outputs the speech / noise label P(y = K|x);

[0131] Specifically, the output speech / noise label P(y = K|x) is:

[0132]

[0133] Wherein, K represents the category; d K represents the result after the output of the LSTM layer is transformed by the fully connected layer; O represents the total number of categories; d o represents the output value of the fully connected layer; o represents the category index variable; x represents the input data; y represents the target label;

[0134] S3.7. Use the cross-entropy loss function L to backpropagate and optimize the parameters.

[0135] Specifically, the cross-entropy loss function L is:

[0136]

[0137] Specifically, in the hearing aid with artificial intelligence-based adaptive gain and noise reduction, constructing a CNN-LSTM hybrid network model to classify speech and noise in real time is crucial. First, by normalizing the time-frequency matrix and converting it into a three-dimensional tensor suitable for network input, the consistency and processability of the data are ensured. Then, the convolutional layer is used to extract local spectral features, and the ReLU activation function and pooling layer are used to further refine the key features, reducing the data dimension while retaining important information. Then, these features are input into the LSTM layer in time series to learn the temporal dependence relationship between speech and noise, enhancing the ability to understand dynamic environmental changes. Subsequently, the fully connected layer combines the Softmax function to map the output of the LSTM layer into class probabilities, accurately identifying the speech or noise label. Finally, the cross-entropy loss function is used to optimize the model parameters to improve the classification accuracy. This series of steps enables the hearing aid to intelligently distinguish speech from background noise, dynamically adjust the gain, and provide a clearer and more comfortable listening experience, especially suitable for complex and changing listening environments. This method not only improves the clarity of the speech signal but also effectively reduces background noise interference, bringing more personalized and efficient hearing support to users.

[0138] S4. Dynamically adjust the gain values of different frequency bands according to the hearing loss curve and real-time signal-to-noise ratio, taking into account individual hearing differences and environmental noise changes, and perform time-domain smoothing processing on the gain to optimize the gain values of different frequency bands;

[0139] In this example, the specific steps of dynamically adjusting the gain values of different frequency bands according to the hearing loss curve and real-time signal-to-noise ratio, taking into account individual hearing differences and environmental noise changes, and performing time-domain smoothing processing on the gain to optimize the gain values of different frequency bands are as follows:

[0140] S4.1. Use the classification result of S3 to mark the noise-dominated frames and update the noise power spectrum Φ nn (f);

[0141] Specifically, the noise power spectrum Φ nn (f) is:

[0142]

[0143] In the formula, Φ nn (f) represents the noise power spectrum; β represents the smoothing factor; Indicates the noise power of the current frame;

[0144] S4.2. Dynamically adjust the gain according to the difference ΔG(f) between the current signal-to-noise ratio SNR(f) and the target;

[0145] Specifically, the signal-to-noise ratio SNR(f) is:

[0146]

[0147] In the formula, P f Indicates the signal power at frequency f;

[0148] The difference ΔG(f) between the signal-to-noise ratio and the target is:

[0149] ΔG(f) = α·max(0, SNR target - SNR(f));

[0150] In the formula, SNR target Indicates the target signal-to-noise ratio; α indicates the adaptive factor;

[0151] The combined reference gain and dynamic correction amount is:

[0152] G total (f) = G base (f) + ΔG(f);

[0153] In the formula, G total (f) indicates the total gain finally applied to frequency f; G base (f) indicates the reference gain;

[0154] S4.3. Perform wide dynamic range compression.

[0155] Specifically, in a hearing aid with artificial intelligence-based adaptive gain and noise reduction, the process of dynamically adjusting the gain values of different frequency bands according to the user's hearing loss curve and real-time signal-to-noise ratio is crucial. First, use the classification results to identify and mark the noise-dominated audio frames, and update the noise power spectrum to ensure an accurate understanding of the environmental noise. Then, calculate the current signal-to-noise ratio and compare it with the target signal-to-noise ratio, and dynamically adjust the gain of each frequency band according to the difference to optimize speech clarity and auditory comfort. Finally, determine the final gain by combining the reference gain and the dynamic correction amount, so that the device can automatically adjust the sound amplification effect according to the user's specific hearing needs and the surrounding environment changes. This method not only improves the quality of the speech signal, but also effectively reduces the background noise interference, providing a more personalized and comfortable listening experience for the user, and helping them communicate and enjoy the sound more easily in various environments.

[0156] In this example, the specific process of performing wide dynamic range compression is:

[0157] S4.31. Map the input signal strength I in (f) to the output gain G out (f);

[0158] Specifically, mapping the input signal strength I in (f) to the output gain G out (f) is as follows:

[0159]

[0160] In the formula, CR represents the compression ratio; I ref represents the reference strength;

[0161] S4.32. Perform temporal smoothing on the gain to avoid mutations.

[0162] G final (f,t) = γ·G final (f,t - 1)+(1 - γ)·G out (f);

[0163] In the formula, G final (f,t) represents the final gain of the current frame; G final (f,t - 1) represents the gain value of the previous frame; γ represents the first-order smoothing factor;

[0164] Furthermore, in a hearing aid with artificial intelligence-based adaptive gain and noise reduction, the process of wide dynamic range compression aims to optimize the sound output quality of input signals with different intensities. Specifically, first, the output gain is adjusted according to the intensity of the input signal. For signals below the reference intensity, the original gain remains unchanged, while for signals above the reference intensity, the gain is reduced according to the set compression ratio, ensuring a comfortable listening experience even in high-volume environments and protecting the user's hearing. Then, by performing temporal smoothing on the gain, the discomfort or sound distortion caused by sudden changes in gain is avoided, making the transition more natural and smooth. This mechanism not only improves speech clarity and sound quality but also enhances the user's auditory comfort in various environments, achieving personalized and efficient hearing assistance. Finally, this meticulous gain management strategy helps the user better understand and enjoy the surrounding sound world.

[0165] S5. Suppress the noise in non-target directions to enhance the spatial selectivity of the speech signal;

[0166] In this example, suppress the noise in non-target directions to enhance the spatial selectivity of the speech signal;

[0167] S5.1. Input the noise power spectrum Φ nn (f);

[0168] S5.2. Calculate the frequency-domain filtering coefficient W(f) to suppress noise components;

[0169] Specifically, the frequency-domain filtering coefficient W(f) is:

[0170]

[0171] In the formula, Φ ss (f) represents the target signal;

[0172] S5.3. Combine the beamforming result of S1 to further attenuate the residual noise in non-target directions.

[0173] Specifically, by suppressing the noise in non-target directions, the spatial selectivity of the speech signal is enhanced. This process significantly improves speech clarity and the auditory experience, enabling the hearing aid to more precisely focus on the sounds the user wishes to hear while reducing background noise interference, providing a more natural and clear auditory sensation for the user. This not only enhances the performance of the device but also improves the user's comfort and satisfaction.

[0174] S6. Optimize the model parameters according to the user's hearing profile to support individual difference adaptation.

[0175] In this example, the model parameters are optimized according to the user's hearing profile to support individual difference adaptation.

[0176] S6.1. Based on the CNN-LSTM model, use the user-specific data D user to fine-tune the classifier parameters

[0177] S6.2. According to the user's real-time feedback on gain adjustment;

[0178] S6.3. Periodically collect the user's usage environment data to update the noise library and model weights to achieve long-term personalized adaptation.

[0179] Specifically, in a hearing aid with artificial intelligence-based adaptive gain and noise reduction, the process of optimizing the model parameters according to the user's hearing profile to support individual difference adaptation is crucial. By using the user-specific data (S6.1) to fine-tune the classifier parameters of the CNN-LSTM model, combined with the user's real-time feedback on gain adjustment (S6.2), and periodically collecting the user's usage environment data to update the noise library and model weights (S6.3), the hearing aid can achieve long-term personalized adaptation. This method not only improves speech clarity and auditory comfort but also enables the device to dynamically adapt to the user's hearing changes and different environmental noise conditions, thus providing a more customized and effective hearing assistance experience. Ultimately, this personalized adaptation mechanism ensures that each user can obtain the best hearing support in their daily lives.

[0180] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and the descriptions in the specification are only preferred examples of the present invention, which are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed.

Claims

1. A hearing aid based on artificial intelligence for adaptive gain and noise reduction, characterized in that: It includes the following steps: S1. Receive acoustic signals through a dual-microphone array, separate the target speech and environmental noise through time-delay difference, and establish a spatial sound field model; S2. Convert the audio signal into the time-frequency domain and extract MFCC features; S3. Construct a CNN-LSTM hybrid network model to classify speech and noise in real time; S4. Dynamically adjust the gain values of different frequency bands according to the hearing loss curve and real-time signal-to-noise ratio, taking into account individual hearing differences and environmental noise changes, and perform time-domain smoothing processing on the gain to optimize the gain values of different frequency bands; S5. Suppress the noise in non-target directions to enhance the spatial selectivity of the speech signal; S6. Optimize the model parameters according to the user's hearing profile to support individual difference adaptation.

2. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 1, wherein: In the above S1, the specific steps of receiving acoustic signals through a dual-microphone array, separating the target speech and environmental noise through time-delay difference, and establishing a spatial sound field model are as follows; S1.

1. The two microphones respectively receive acoustic signals x1 and x2, covering the target sound source and environmental noise; S1.

2. Calculate the cross-correlation function R 12 (τ) of the two signals and find the time-delay difference τ corresponding to the maximum correlation value max for determining the azimuth angle θ of the target sound source; S1.

3. Based on the time-delay difference, construct a spatial sound field model to distinguish the target speech direction and the environmental noise direction; S1.

4. Use the noise covariance matrix R nn and the steering vector d(θ) to calculate the optimal beam weight w for enhancing the signal in the main direction and suppressing the sidelobe noise.

3. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 2, wherein: In the above S1.3, the specific steps of constructing a spatial sound field model based on the time-delay difference to distinguish the target speech direction and the environmental noise direction are: S1.

31. According to the estimated azimuth angle θ, divide the sound field into the target direction and the noise direction; S1.

32. Perform frequency-domain analysis on the signal and construct the sound field distribution P(f,θ) in the frequency-angle domain; S1.

33. Use the time-delay estimation result to mark the frequency-domain energy in non-target directions as noise components.

4. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 3, characterized in that: In the step S1.4, the noise covariance matrix R nn and the steering vector d(θ) are used to calculate the optimal beam weight w for enhancing the signal in the main direction and suppressing the sidelobe noise. The specific steps are as follows: S1.

41. During the noise-dominated period, calculate the noise covariance matrix R nn ; S1.

42. According to the target direction θ and the geometric structure of the microphone array, define the steering vector d(θ); S1.

43. Solve the constrained optimization problem and calculate the optimal beam weight w; S1.

44. Apply the weight to the input signal x(t) to enhance the signal in the target direction.

5. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 1, characterized in that: In the above S2, the specific steps of converting the audio signal into the time-frequency domain and extracting MFCC features are: S2.

1. Generate the time-frequency matrix Y(f,l) through short-time Fourier transform; S2.

2. Extract MFCC features according to the time-frequency matrix.

6. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 1, characterized in that: In the above S3, the specific steps of constructing a CNN-LSTM hybrid network model to classify speech and noise in real time are: S3.

1. Normalize the time-frequency matrix Y(f,l) and convert it into a tensor suitable for network input; S3.

2. Use the convolutional layer to extract local spectral features; S3.

3. Reduce the dimension through the ReLU activation function and the pooling layer to retain key features; S3.

4. Input the output of the CNN into the LSTM layer according to the time series to learn the temporal dependence relationship between speech and noise; S3.

5. Control the information flow through the LSTM layer; S3.

6. The fully connected layer maps the output of the LSTM to class probabilities, and the Softmax layer outputs the speech / noise label P(y = K|x); S3.

7. Use the cross-entropy loss function L for backpropagation to optimize the parameters.

7. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 6, characterized in that: In S4, according to the hearing loss curve and the real-time signal-to-noise ratio, the gain values of different frequency bands are dynamically adjusted. Considering individual hearing differences and environmental noise changes, the specific steps for optimizing the gain values of different frequency bands by performing time-domain smoothing processing on the gain are as follows: S4.

1. Mark the noise-dominated frames using the classification results of S3 and update the noise power spectrum Φ nn (f); S4.2: Dynamically adjust the gain according to the difference ΔG(f) between the current signal-to-noise ratio SNR(f) and the target; S4.3: Perform wide dynamic range compression.

8. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 7, characterized in that: In S4.3, the specific process of performing wide dynamic range compression is as follows: S4.

31. Map the input signal strength I in (f) to the output gain G out (f); S4.32: Perform time-domain smoothing on the gain to avoid sudden changes.

9. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 1, characterized in that: In S5, suppress the noise in non-target directions to enhance the spatial selectivity of the speech signal; S5.1, Input noise power spectrum Φ nn (f); S5.2: Calculate the frequency-domain filtering coefficient W(f) to suppress the noise components; S5.3: Combine the beamforming result of S1 to further attenuate the residual noise in non-target directions.

10. The hearing aid based on artificial intelligence for adaptive gain and noise reduction according to claim 1, characterized in that: In S6, optimize the model parameters according to the user's hearing profile to support adaptation to individual differences; S6.

1. On the basis of the CNN-LSTM model, use the user-specific data D user Fine-tune the classifier parameters; S6.2: According to the user's real-time feedback on the gain adjustment; S6.3: Periodically collect the user's usage environment data to update the noise library and model weights to achieve long-term personalized adaptation.

Images