Intelligent Noise Reduction Method for Hearing Aids Based on Topological Data Analysis

Through topological data analysis and personalized parameter models, hearing aids achieve accurate identification and suppression of noise in complex environments, improving user experience and speech clarity, and solving the problems of poor noise reduction and personalized needs in existing technologies.

CN123123798APending Publication Date: 2026-11-17ZUODIAN IND (HUBEI) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510622554.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing hearing aid noise reduction technologies are ineffective in complex and variable non-steady-state noise environments and cannot meet the personalized needs of users, resulting in a poor user experience.

Method used

The intelligent noise reduction method for hearing aids based on topological data analysis constructs the topological space of the sound signal and combines it with the user's personalized parameter model to achieve accurate noise identification and dynamic suppression. It also uses adaptive filters and multi-scale fusion algorithms to enhance the speech signal.

Benefits of technology

It improves noise reduction accuracy in complex environments, enhances user personalization and voice clarity, and reduces computational load and power consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123798A_ABST
    Figure CN123123798A_ABST
Patent Text Reader

Abstract

This invention belongs to the technical field of intelligent audio processing and hearing aids, specifically a smart noise reduction method for hearing aids based on topological data analysis. It collects ambient sound through a microphone array, extracts topological features such as Betti number and persistence graph based on the Vietoris-Rips complex after preprocessing, and combines this with a multi-dimensional personalized parameter model. Multiple classifiers are used to achieve accurate speech / noise classification, and the filter is dynamically adjusted for noise reduction based on the topological features. The output is then fused across multiple scales. This method overcomes the limitations of traditional noise reduction, achieving precise noise reduction driven by topology, dynamic environmental adaptation, and deep personalized adjustment. It significantly improves noise reduction performance and user auditory experience in complex scenarios such as subways and exhibitions, while also possessing advantages such as low power consumption and high generalization ability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent audio processing and hearing aid technology, specifically a method for intelligent noise reduction in hearing aids based on topological data analysis. Background Technology

[0002] With the continued growth of the global population with hearing impairments, hearing aids, as key devices for improving hearing, have made noise reduction performance a core factor affecting user experience. Current mainstream hearing aid noise reduction technologies are mainly divided into traditional signal processing methods, machine learning-based intelligent methods, and hybrid technology solutions, but all have significant limitations.

[0003] Traditional signal processing methods, such as Wiener filtering and spectral subtraction, rely on pre-defined fixed-parameter models to reduce noise through simple analysis of the signal's time or frequency domain characteristics. While these methods can achieve basic noise reduction in steady-state noise environments, they struggle with complex and variable non-steady-state noise (such as the intermingling of subway rumble and station announcements, or the multi-source noise at an exhibition). This is because they cannot capture the dynamic changes in noise in real time, easily leading to either over-noise reduction causing speech distortion or insufficient noise reduction leaving the noise interference still severe. For example, traditional spectral subtraction, based on the assumption that noise is spectrally stable over a short period, may misjudge and filter out parts of the speech signal as noise in scenarios with sudden noise changes, resulting in a significant decrease in speech clarity.

[0004] Machine learning-based noise reduction methods, such as deep neural networks (DNNs) and recurrent neural networks (RNNs), train models using large amounts of labeled data to identify and suppress noise. While these methods demonstrate good noise reduction performance on specific training datasets, they face significant generalization challenges in practical applications. When the acoustic characteristics of the user's environment differ from those of the training data (e.g., transitioning from a quiet indoor environment to a noisy outdoor environment), the model's noise reduction effectiveness decreases significantly. Furthermore, the reliance on labeled data leads to high development costs, and the complex model structures result in massive computational demands, making them difficult to adapt to the limited hardware resources and low-power requirements of hearing aid devices.

[0005] While existing hybrid technologies attempt to combine the advantages of traditional methods with machine learning, they still fail to fully utilize the essential differences in the characteristics of sound signals. Most current technologies analyze only the local time-domain, frequency-domain, or time-frequency-domain features of the signal, ignoring the significant differences in the global topological structure between speech and noise signals. Speech signals possess semantic coherence and prosodic periodicity, exhibiting specific structural features in the topological space; while noise signals typically exhibit a disordered, discrete topological structure. Furthermore, existing noise reduction systems generally employ uniform parameter settings, failing to meet the personalized needs of different users in terms of hearing loss levels, frequency sensitivity, and noise reduction preferences. This leads to a large number of users abandoning hearing aids due to unsatisfactory noise reduction effects. According to the World Health Organization, approximately 30% of hearing aid users discontinue use due to inadequate noise reduction, indicating an urgent need for an intelligent noise reduction method that can effectively utilize signal topological features, adapt to complex environments, and support personalized adjustments. Summary of the Invention

[0006] To address the problems in existing technologies, this invention provides an intelligent noise reduction method for hearing aids based on topological data analysis. By constructing a topological space for sound signals and combining it with a user-personalized parameter model, it achieves accurate identification and dynamic suppression of noise, while enhancing the speech signal and improving the user's auditory experience in various complex environments.

[0007] The technical solution adopted by this invention to solve its technical problem is: a hearing aid intelligent noise reduction method based on topological data analysis, comprising the following steps: Signal acquisition and preprocessing: Ambient sound signals are acquired using the built-in microphone array of the hearing aid, and then normalized, DC component removed, and pre-emphasized after anti-aliasing filtering; Topological feature extraction: Short-time Fourier transform is performed on the preprocessed signal to obtain the time-frequency matrix. The topological space is constructed based on the Vietoris-Rips complex, and topological features such as Betti number, persistence graph, persistence value and topological entropy are calculated. Personalized model construction: By combining user hearing test data, hearing loss settings and usage feedback, a multi-dimensional personalized parameter model is established, which includes frequency sensitivity coefficient, noise reduction intensity weight and environmental adaptability coefficient. Noise Classification and Recognition: Input topological features and personalized parameters into one or more combined classifiers such as Support Vector Machine, Random Forest, Neural Network, K-Nearest Neighbors, Naive Bayes, or AdaBoost to perform speech / noise classification; Adaptive noise reduction: For the classified noise signal, the Wiener filter parameters are dynamically adjusted according to its topological characteristics for noise reduction. The filter parameter adjustment formula is as follows: ,in For speech power spectrum, The noise power spectrum, Adjust the noise topology complexity within the range of 0.1-1.0; Speech enhancement and output: A multi-scale fusion algorithm is used to fuse the noise-reduced signal with the original speech components, and then output it to the hearing aid speaker after digital-to-analog conversion.

[0008] Specifically, the normalization formula is as follows: ; The pre-weighting process formula is: ; Where x(n) is the original signal and y(n) is the pre-emphasized signal. This is the pre-emphasis coefficient. The value ranges from 0.9 to 1. By standardizing the signal amplitude and enhancing the high frequency, it provides a standardized data basis for topological feature extraction and is suitable for sound signals with different amplitude distributions.

[0009] Specifically, the frame length of the short-time Fourier transform ranges from 200 to 320 sampling points, the frame shift ranges from 100 to 160 sampling points, and the radius parameter of the Vietoris-Rips complex is adjusted from 0.1 to 0.5. By adjusting these parameters, the time-frequency resolution and computational complexity of the signal are balanced, adapting to the topological characteristics of different types of sound signals.

[0010] Specifically, the frequency sensitivity coefficient in the personalized parameter model is set for the frequency band of the user's hearing loss, the noise reduction intensity weight is adjusted according to the user's noise reduction preference, and the environmental adaptation coefficient is dynamically adjusted according to real-time environmental changes; the user's hearing loss is obtained through hearing test data and hearing aid adaptive learning, and the usage feedback includes the number of volume adjustments and noise reduction mode switching records, realizing a deep personalized noise reduction for each individual.

[0011] Specifically, when the classifier uses a support vector machine combined with a dynamic time warping algorithm, the support vector machine maps topological features to a high-dimensional space for classification through a kernel function, and the dynamic time warping algorithm calculates signal similarity; when a neural network is used, the mapping relationship between speech and noise topological features and categories is learned through training, and the combination of multiple classifiers ensures that the classification accuracy is not less than 85% in different environments.

[0012] Specifically, when dynamically adjusting the filter parameters based on noise topology characteristics, if the noise signal topology shows energy concentration in a certain frequency band, the filtering intensity of that frequency band is enhanced; when the topology characteristics of transient noise are detected, the filter parameters are quickly adjusted to enhance the filtering at the moment the noise appears, effectively suppressing non-stationary noise.

[0013] Specifically, the multi-scale fusion algorithm divides the signal into at least three scales according to frequency and calculates fusion weights for signals of different scales. The fusion weights are determined based on the energy ratio of the denoised signal to the original speech and the preservation of speech components. Compared with the traditional weighted fusion algorithm, the speech clarity is improved by more than 20%.

[0014] Specifically, it also includes periodically updating the personalized parameter model based on user feedback. The updating methods include: automatically optimizing the frequency sensitivity coefficient when the user frequently adjusts the volume; and adjusting the current coefficient based on historical data of the environmental adaptation coefficient when the environment changes, so that the noise reduction adaptability of the system is improved by 40% within one month of use.

[0015] Specifically, the anti-aliasing filter uses a low-pass filter with a cutoff frequency half that of the Nyquist frequency to prevent high-frequency signal aliasing; the removal of DC components is achieved by removing the signal mean to avoid the impact of DC offset on subsequent processing.

[0016] Specifically, the noise reduction process further includes using morphological filtering and sparse representation to further process the residual noise after adaptive filtering; the morphological filtering eliminates noise impulses through opening and closing operations, and the sparse representation performs noise reduction by analyzing the sparse characteristics of noise, thereby further improving the noise reduction effect.

[0017] The beneficial effects of this invention are: Topology-driven precision noise reduction: By extracting the topological features of the sound signal, it achieves precise classification by utilizing the structural differences between speech and noise in the topological space. Compared with traditional methods, the noise reduction accuracy is improved by 27% in complex environments, effectively reducing speech distortion problems.

[0018] Dynamic environment adaptation: Based on real-time updated topology features and personalized parameters, the system can respond to environmental changes within 300ms. Whether in complex scenarios such as subways and exhibitions, or when switching between indoor and outdoor environments, it can maintain stable noise reduction performance.

[0019] Deeply personalized customization: Through a multi-dimensional user-personalized model, adjustments to frequency sensitivity, noise reduction intensity, and other aspects are made tailored to each individual user. User satisfaction surveys show that personalized solutions improve user comfort by 42%, meeting the specific needs of different users.

[0020] Low-resource, high-efficiency operation: The topology feature extraction and classification algorithm adopts a lightweight design, which reduces the computational load by 68% compared with deep learning solutions. It can be adapted to mainstream low-power hearing aid chips, reducing device power consumption while ensuring noise reduction effect.

[0021] Significantly improved generalization ability: The geometric invariance of topological features enables it to maintain more than 89% noise reduction performance in cross-environment testing, breaking through the scenario limitations of traditional data-driven methods and adapting to new environments without extensive retraining. Attached Figure Description

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Figure 1 The flowchart illustrates the intelligent noise reduction method for hearing aids based on topological data analysis provided by this invention. Detailed Implementation

[0024] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0025] like Figure 1 As shown, the intelligent noise reduction method for hearing aids based on topology data analysis according to the present invention includes the following steps: Signal Acquisition and Preprocessing: Ambient sound signals are acquired using the built-in microphone array of the hearing aid. First, an anti-aliasing filter is used to eliminate high-frequency interference. Then, the signal is normalized using the following formula:

[0026] This formula eliminates signal amplitude differences, providing a standardized data foundation for subsequent processing. Simultaneously, it performs DC component removal and pre-emphasis operations on the signal; the pre-emphasis formula is as follows: Where x(n) is the original signal and y(n) is the pre-emphasized signal. The pre-emphasis coefficient is usually between 0.9 and 1. It enhances the amplitude of high-frequency signals and strengthens the high-frequency components in speech signals by pre-emphasing.

[0027] Topological feature extraction: Short-time Fourier Transform (STFT) is performed on the preprocessed signal to obtain the time-frequency matrix S(t,f). The frame length is set to a range of 200-320 sampling points, and the frame shift is set to a range of 100-160 sampling points to balance the signal's time-frequency resolution and computational complexity. Based on the time-frequency matrix, the topological space of the signal is constructed using a Vietoris-Rips complex, and the topological representation is optimized by adjusting the radius parameter of the Vietoris-Rips complex (adjustment range 0.1-0.5). Morse theory is used to calculate the topological features of the signal, including the Betti number. The persistent graph Dgm is calculated, and parameters such as the persistence value and topological entropy of topological features are further calculated. The Betti number formula is: These parameters comprehensively describe the topological structure of the signal, capturing the essential differences between speech signals and noise signals in the topological space.

[0028] Personalized Model Construction: This involves combining the user's hearing test data, hearing loss information entered in the hearing aid settings (such as the degree of hearing loss in different frequency bands), and real-time usage feedback (including volume adjustment counts, noise cancellation mode switching records, etc.) to build a user-personalized parameter model containing multi-dimensional parameters. Specifically, this model includes the user's sensitivity coefficients to sounds of different frequencies and intensities. Noise Reduction Intensity Weight In addition, there is an environmental adaptability coefficient. Among them, the frequency sensitivity coefficient is set for the frequency band of the user's hearing loss, the noise reduction intensity weight is adjusted according to the user's noise reduction preference, and the environmental adaptability coefficient is dynamically adjusted according to real-time environmental changes, thereby realizing a deeply personalized noise reduction solution customization.

[0029] Noise Classification and Recognition: The extracted topological features are combined with a user-personalized parameter model and input into a classifier for speech / noise classification. The classifier can employ one or more combinations of Support Vector Machine (SVM), Random Forest, Neural Network, K-Nearest Neighbors, Naive Bayes, and AdaBoost. Taking SVM combined with Dynamic Time Warping (DTW) as an example, SVM maps topological features to a high-dimensional space for classification using a kernel function, while DTW calculates signal similarity to achieve accurate classification of sound signals, determining whether the signal belongs to speech, noise, or a mixture.

[0030] Adaptive noise reduction: For the noise component in the classified noise signal or mixed signal, an adaptive filtering algorithm based on topological features is used for noise reduction. This algorithm dynamically adjusts the filter parameters according to the topological features of the noise signal, using the formula:

[0031] in Adjustable within the range of 0.1-1.0 based on noise topology complexity. When the signal topology shows concentrated energy in a certain frequency band, the filtering strength of that frequency band is enhanced to achieve precise noise suppression.

[0032] Speech Enhancement and Output: A multi-scale fusion algorithm is employed to fuse the denoised signal with the speech components of the original signal. This algorithm divides the signal into multiple scales based on its frequency components and calculates fusion weights for each scale separately. Compared to traditional weighted fusion algorithms, this effectively improves speech clarity. The fused digital signal is converted into an analog signal via a digital-to-analog converter and output to the hearing aid's speaker for the user to listen to.

[0033] Example 1: Noise Reduction in Subway Environment The hearing aid's microphone array collects ambient sound signals during subway operation, including the rumble of trains, station announcements, and passenger conversations. First, the raw signal undergoes anti-aliasing filtering to remove high-frequency interference, and then the formula is used... Normalization is performed to eliminate signal amplitude differences. Next, DC component removal and pre-emphasis are applied, with the pre-emphasis coefficient... Take 0.95.

[0034] A short-time Fourier transform was performed on the preprocessed signal, with a frame length of 256 sampling points and a frame shift of 128 sampling points to obtain the time-frequency matrix. Based on the time-frequency matrix, a Vietoris-Rips complex was used to construct the topological space, and the topological representation was optimized by adjusting the complex radius parameter (gradually from 0.1 to 0.3). The 0-2 Vietoris number was calculated to detect the train's rumbling sound. The value is significantly higher than that of the speech signal in the low-frequency band (50-200Hz), and its persistence graph shows a long-lived topological feature in the low-frequency band.

[0035] Based on the user's hearing loss data (40dB at 4kHz) and noise reduction preference (a preference for stronger noise reduction) entered in the hearing aid settings, a personalized parameter model is established. Frequency sensitivity coefficients are set. Noise reduction intensity weight Based on the characteristics of the subway environment, the environmental adaptability coefficient was initially set at 0.6.

[0036] The extracted topological features and personalized parameter models are input into a support vector machine classifier, and combined with a dynamic time warping algorithm, the sound signals are classified, accurately identifying the train roar as a noise signal.

[0037] Based on the identified train noise, the Wiener filter parameters are dynamically adjusted according to its topological characteristics. According to the formula... Due to its high topological complexity, The parameter was adjusted to 0.8 to enhance the noise reduction strength in the low-frequency band.

[0038] A multi-scale fusion algorithm is used to fuse the denoised signal with the speech component of the original signal, and the result is output to the hearing aid speaker after digital-to-analog conversion. Tests show that speech clarity is improved by 35%, and train noise is reduced from 78dB to 42dB, allowing users to clearly hear station announcements and conversations.

[0039] Example 2: Noise Reduction in Noisy Exhibition Environments The system collects mixed audio signals from the exhibition venue, including numerous human voices, equipment operating sounds, and background music. After anti-aliasing filtering, the raw signal undergoes normalization, DC component removal, and pre-emphasis processing. The pre-emphasis coefficient is... Take 0.92.

[0040] Short-time Fourier transform was performed with a frame length of 320 samples and a frame shift of 160 samples to obtain the time-frequency matrix, after which a topological space was constructed. During the construction of the topological space, the radius parameter of the Vietoris-Rips complex was carefully adjusted. By observing the changes in the persistence graph, the optimal radius parameter was determined to be 0.4 to better reflect the signal's topological structure. The topological features of the signal were extracted, and analysis revealed that background music exhibits a unique topological structure in the high-frequency range (above 8kHz), with its betty number... The value is relatively high in this frequency band, and the persistent graph shows stable topological characteristics.

[0041] The user is a business professional with good hearing, but wants to be able to clearly hear important conversations in noisy environments while preserving some of the ambiance. Therefore, noise reduction intensity weights are set. Because it is relatively sensitive to high-frequency sounds, a frequency sensitivity coefficient is set. Based on the dynamic changes in the exhibition environment, the environmental adaptability coefficient was set to 0.7.

[0042] Using extracted topological features and personalized parameter models, a random forest classification algorithm is employed to classify the signal. By constructing multiple decision trees and combining their results, the random forest algorithm accurately identifies background music as noise.

[0043] To address the topological characteristics of the background music, an adaptive filtering algorithm based on these characteristics is employed for noise reduction. According to its topological structure in the high-frequency band, the filter parameters are dynamically adjusted, increasing the filtering strength above 8kHz, thereby reducing the background music intensity by 28% while preserving vocals.

[0044] The noise-reduced signal is fused with the original speech components using a multi-scale fusion algorithm, and then converted into an analog signal output. User feedback indicates that they can communicate clearly with others during the exhibition, while retaining 82% of the naturalness of the ambient sound, thus reducing noise interference without excessively disrupting the atmosphere.

[0045] Example 3: Outdoor Wind Noise Suppression Outdoor environmental sound signals, including wind sounds, birdsong, and people exercising and talking, were collected. After anti-aliasing filtering, the raw signals were normalized, DC component removed, and pre-emphasized. The pre-emphasis coefficients were... Take 0.9.

[0046] A short-time Fourier transform was performed with a frame length of 280 sampling points and a frame shift of 140 sampling points to obtain the time-frequency matrix and construct the topological space. Analysis revealed that the persistence map of wind sounds exhibits a long, strip-shaped structure in the low-frequency range (20-100Hz), which is significantly different from the topological structures of birdsong and human voices. The Betty number was also calculated. The value is more prominent in the low-frequency range.

[0047] The user is an elderly person with hearing impairment. The audiogram shows significant low-frequency hearing loss, so low-frequency noise reduction weights are set. At the same time, we hope to preserve natural sounds such as birdsong as much as possible to enhance the outdoor experience. Given the rapidly changing nature of the outdoor environment, the environmental adaptability coefficient is set to 0.8.

[0048] Topological features and personalized parameters are input into a neural network classifier for classification. By learning from a large amount of speech and noise topological feature data, the neural network accurately identifies wind sound as a noise signal.

[0049] To address the topological characteristics of wind noise, an adaptive filtering algorithm based on topological features combined with morphological filtering is employed for noise reduction. According to the low-frequency topology, the Wiener filter parameters are dynamically adjusted to enhance the filtering intensity in the 20-100Hz frequency range, while morphological filtering further eliminates residual noise. After noise reduction, the wind noise intensity is reduced from 65dB to 38dB, while the clarity of birdsong is improved by 40%, allowing users to enjoy natural sounds while reducing wind noise interference.

[0050] A multi-scale fusion algorithm is used to fuse the denoised signal with the original speech components, and the result is output to the hearing aid speaker after digital-to-analog conversion. When exercising outdoors, users can clearly hear reminders and conversations from others, while also experiencing comfortable natural environmental sounds.

[0051] Example 4: Restaurant Environment Noise Reduction Sound signals from within the restaurant were collected, including the sounds of clinking cutlery, conversations, and kitchen equipment operation. After anti-aliasing filtering, the raw signals were normalized, DC component removed, and pre-emphasized. The pre-emphasis coefficients were... Take 0.93.

[0052] A short-time Fourier transform was performed with a frame length of 240 sampling points and a frame shift of 120 sampling points to obtain the time-frequency matrix, which was then used to construct a topological space. The topological features of the signal were extracted to identify the sounds of tableware colliding. The value is exceptionally prominent in the mid-to-high frequency range (1-5kHz). Its persistence plot in this frequency band exhibits rapidly changing and short-lived topological characteristics, which is clearly distinguishable from the topological structure of conversation.

[0053] The user has set their preferences to be more sensitive to mid-range (2-3kHz) sounds, and has adjusted the frequency sensitivity coefficient. At the same time, it is desirable to be able to clearly hear conversations between people opposite in a restaurant environment, so noise reduction intensity weights are set. Based on the fact that the restaurant environment is relatively stable but has a variety of sounds, the environmental adaptability coefficient is set to 0.75.

[0054] The K-Nearest Neighbors (KNN) classification algorithm is used to classify signals. The KNN algorithm determines the category of the sample to be classified based on the distance between the sample to be classified and samples of known categories. By calculating the distance between topological features, the sound of tableware colliding is accurately identified as a noise signal.

[0055] To address the topological characteristics of cutlery clattering sounds, an adaptive filtering algorithm based on these characteristics is employed for noise reduction. According to the topological structure in the mid-to-high frequency band, filter parameters are dynamically adjusted, increasing noise reduction intensity to 65% in the 1-5kHz frequency range. Simultaneously, by optimizing the filter design, the clarity of spoken conversation is preserved to the greatest extent possible. Testing shows that the speech signal-to-noise ratio has improved from 12dB to 21dB, allowing users to easily communicate with others in restaurants without being overly disturbed by cutlery clattering sounds.

[0056] A multi-scale fusion algorithm is used to fuse the denoised signal with the original speech components, converting it into an analog signal output to the hearing aid speaker. User feedback indicates that communication becomes smoother and the overall auditory experience is significantly improved when dining in restaurants.

[0057] Example 5: Noise Reduction in Meeting Scenarios The system collects audio signals from the conference room, including projector fan noise, speaker voice, and page-turning sounds. After anti-aliasing filtering, the raw signals undergo normalization, DC component removal, and pre-emphasis processing. The pre-emphasis coefficients are... Take 0.94.

[0058] A short-time Fourier transform was performed with a frame length of 200 sampling points and a frame shift of 100 sampling points to obtain the time-frequency matrix and construct the topological space. Analysis of the signal's topological characteristics revealed that the projector fan noise exhibits a stable topological structure in the low-frequency range (30-150Hz), with a relatively high Betti number. It maintains a high and stable value in this frequency band, and the persistence plot also shows long-lived topological characteristics, which are significantly different from the topological structure of spoken speech.

[0059] The user is sensitive to low-frequency noise, so low-frequency noise reduction weights are set. At the same time, we hope to be able to hear the speaker clearly. Given that the meeting environment is relatively quiet and requires high speech clarity, the environmental adaptability coefficient is set to 0.85.

[0060] The Naive Bayes classification algorithm is used to classify the signal. Based on Bayes' theorem, the Naive Bayes algorithm assumes that the features are independent of each other. By calculating the probability of topological features in different categories, it accurately identifies the sound of the projector fan as a noise signal.

[0061] To address the topological characteristics of the projector's fan noise, an adaptive filtering method combined with spectral subtraction was employed for noise reduction. Based on the low-frequency topology, the Wiener filter parameters were dynamically adjusted to enhance the filtering intensity in the 30-150Hz frequency range, while spectral subtraction was used to further eliminate residual noise. After processing, the fan noise was reduced from 55dB to 32dB, while the distortion of spoken speech was controlled within 3%, ensuring the clarity and integrity of the speech.

[0062] A multi-scale fusion algorithm is employed to fuse the denoised signal with the original speech components. The algorithm first divides the signal into three scales based on frequency: low frequency (20-200Hz), mid frequency (200-2000Hz), and high frequency (2000-20000Hz). For the low-frequency scale, since a small amount of useful low-frequency speech components are still retained in the denoised signal, the fusion weight is calculated to be 0.3 based on the ratio of its energy to the low-frequency energy of the original speech. The mid-frequency scale is dominated by the original speech, and the fusion weight is set to 0.7. For the high-frequency scale, the fusion weight is set to 0.6 based on the preservation of high-frequency speech details in the denoised signal. The fused digital signal is output to the hearing aid speaker via a digital-to-analog converter. User feedback indicates that during the meeting, the speaker's content can be heard clearly, noise interference from the projector fan is largely eliminated, and overall meeting communication efficiency is significantly improved.

[0063] Example 6: Noise Reduction for Home Television Environment The system collects a mixed signal from a home environment, including television sound and family conversations. After applying anti-aliasing filtering to the original signal, a normalization formula is used. Amplitude normalization is performed, followed by DC component removal and pre-emphasis processing. The pre-emphasis coefficient is... Take 0.91.

[0064] Short-time Fourier transform was performed with a frame length of 300 sampling points and a frame shift of 150 sampling points to obtain the time-frequency matrix and construct the topological space. By calculating the Betti number and analyzing the persistence graph, it was found that the topological structure of low-frequency background noise in the low-frequency band (50-300Hz) exhibits stable and long-lived characteristics. The value remains at a high level in this frequency band, clearly distinguishing the topological structure of conversations with family members.

[0065] Users want to effectively suppress TV background noise while preserving the sound of family conversations, therefore, noise reduction intensity weights are set for the TV background audio range (50-300Hz). Considering that family conversations may contain mid-to-high frequency components, a frequency sensitivity coefficient is set for the mid-to-high frequency range (500-5000Hz). Considering the relatively stable nature of the home environment, the environmental adaptability coefficient is set at 0.7.

[0066] The AdaBoost classification algorithm is used to classify the signal. The AdaBoost algorithm iteratively trains multiple weak classifiers and combines them into a strong classifier. Based on the signal topological features and personalized parameters, it accurately identifies the TV background sound as a noise signal.

[0067] To address the topological characteristics of television background noise, a sparse representation method based on topological features is employed for noise reduction. Based on its stable topological structure in the low-frequency band, its sparse representation coefficients are analyzed, and the filter parameters in the 50-300Hz frequency band are dynamically adjusted. This reduces the television background noise by 25% while preserving the sound of family conversations. Simultaneously, by optimizing the filter's time-frequency response characteristics, interference with the mid-to-high frequency components of the conversations is avoided.

[0068] A multi-scale fusion algorithm is used to fuse the denoised signal with the original speech components. The signal is first subdivided into four scales based on frequency: very low frequency (20-100Hz), low frequency (100-500Hz), mid frequency (500-2000Hz), and high frequency (2000-20000Hz). For the very low frequency and low frequency scales, fusion weights of 0.2 and 0.4 are calculated respectively based on the ratio of the residual energy of the TV background noise after denoising to the energy of the original speech. For the mid frequency scale, the fusion weight is set to 0.8 to ensure the clarity of conversation. For the high frequency scale, the fusion weight is set to 0.7 based on the preservation of speech details in the denoised signal. The fused digital signal is then converted from digital to analog and output to the hearing aid speaker. User feedback indicates that they can now converse clearly with family members at home, the TV background noise no longer interferes with communication, and the overall family communication atmosphere is more comfortable.

[0069] 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 to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent noise reduction in hearing aids based on topological data analysis, characterized in that, Includes the following steps: The ambient sound signal is collected using the built-in microphone array of the hearing aid, and then normalized, DC component removed, and pre-emphasized after anti-aliasing filtering. Short-time Fourier transform is performed on the preprocessed signal to obtain the time-frequency matrix. The topological space is constructed based on the Vietoris-Rips complex, and topological features such as Betti number, persistence graph, persistence value and topological entropy are calculated. By combining user hearing test data, hearing loss settings, and usage feedback, a multi-dimensional personalized parameter model is established, which includes frequency sensitivity coefficient, noise reduction intensity weight, and environmental adaptability coefficient. Input topological features and personalized parameters into one or more combined classifiers of support vector machine, random forest, neural network, K-nearest neighbor, Naive Bayes or AdaBoost to classify speech / noise; For the classified noise signals, the Wiener filter parameters are dynamically adjusted based on their topological characteristics for noise reduction. The filter parameter adjustment formula is as follows: ,in For speech power spectrum, The noise power spectrum, Adjust the noise topology complexity within the range of 0.1-1.0; A multi-scale fusion algorithm is used to fuse the noise-reduced signal with the original speech components, and then the signal is output to the hearing aid speaker after digital-to-analog conversion.

2. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: The normalization formula is: ; The pre-weighting process formula is: ; Where x(n) is the original signal and y(n) is the pre-emphasized signal. This is the pre-emphasis coefficient. The value ranges from 0.9 to 1. By standardizing the signal amplitude and enhancing the high frequency, it provides a standardized data basis for topological feature extraction and is suitable for sound signals with different amplitude distributions.

3. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: The frame length of the short-time Fourier transform ranges from 200 to 320 sampling points, the frame shift ranges from 100 to 160 sampling points, and the radius parameter of the Vietoris-Rips complex is adjusted from 0.1 to 0.

5. By adjusting these parameters, the time-frequency resolution and computational complexity of the signal are balanced, adapting to the topological characteristics of different types of sound signals.

4. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: The frequency sensitivity coefficient in the personalized parameter model is set for the frequency band of the user's hearing loss, the noise reduction intensity weight is adjusted according to the user's noise reduction preference, and the environmental adaptation coefficient is dynamically adjusted according to real-time environmental changes. The user's hearing loss status is obtained through hearing test data and hearing aid adaptive learning. The usage feedback includes the number of volume adjustments and noise reduction mode switching records, achieving deeply personalized noise reduction for each individual.

5. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: When the classifier uses a support vector machine combined with a dynamic time warping algorithm, the support vector machine maps topological features to a high-dimensional space for classification through a kernel function, and the dynamic time warping algorithm calculates signal similarity. When using neural networks, the mapping relationship between speech and noise topological features and categories is learned through training.

6. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: When dynamically adjusting filter parameters based on noise topology characteristics, if the noise signal topology shows energy concentration in a certain frequency band, the filtering strength of that frequency band is enhanced. When the topological characteristics of transient noise are detected, the filter parameters are quickly adjusted to enhance the filter at the moment the noise appears, effectively suppressing non-stationary noise.

7. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: The multi-scale fusion algorithm divides the signal into at least three scales according to frequency and calculates the fusion weight for signals of different scales respectively. The fusion weights are determined based on the energy ratio of the denoised signal to the original speech and the preservation of speech components.

8. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: This also includes regularly updating personalized parameter models based on user feedback; The update methods include: Automatically optimize frequency sensitivity coefficient when the user frequently adjusts the volume; When the environment changes, the current coefficient is adjusted based on historical data of the environmental adaptability coefficient.

9. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: The anti-aliasing filter uses a low-pass filter with a cutoff frequency of half the Nyquist frequency to prevent high-frequency signal aliasing. The removal of the DC component is achieved by removing the signal mean, thus avoiding the impact of DC offset on subsequent processing.

10. The intelligent noise reduction method for hearing aids based on topological data analysis according to claim 1, characterized in that: The noise reduction process also includes further processing of the residual noise after adaptive filtering using morphological filtering and sparse representation. The morphological filtering eliminates noise impulses through opening and closing operations; The sparse representation method reduces noise by analyzing the sparse characteristics of noise.