Hearing assistance device, method and arrangement with adaptive hearing compensation

Abstract

This application provides a hearing aid device, method, and apparatus with adaptive hearing compensation function. The hearing aid device includes a wireless communication module, a main control chip, and an audio output module. The main control chip has a built-in pre-trained hearing compensation model. Through model inference, it outputs a personalized frequency response gain curve that is continuous across the entire frequency band and whose gain amplitude is positively correlated with hearing loss. Based on this curve, the audio output module performs full-band, frequency-point equalization linear gain compensation on the original digital audio signal played by the device itself. This process does not compress the audio dynamic range or change the original spectral structure, preserving high-fidelity sound quality. Unlike traditional hearing aids and conventional audio playback devices, this method only processes the internal audio of the device, without the need for sound pickup and noise reduction. It can accurately adapt to the hearing characteristics of hearing-impaired users, achieving clear and balanced listening, while avoiding sound quality distortion, thus balancing compensation accuracy, user comfort, and versatility.

Description

Technical Field

[0001] This application relates to the field of audio output device technology, and in particular to a hearing aid device, method and apparatus with adaptive hearing compensation function. Background Technology

[0002] Traditional audio playback devices, including Bluetooth speakers, TV speakers, and desktop speakers, can only amplify and output internal audio signals according to a uniform standard. They cannot adapt to the individual auditory differences of hearing-impaired users. When hearing-impaired people use such devices, they often experience problems such as unclear sound in some frequency bands and uneven overall loudness. They are unable to meet the multimedia audio listening needs of hearing-impaired users, and their applicable population is significantly limited.

[0003] Traditional hearing aids, as specialized devices for people with hearing impairments, rely on microphones to collect and process ambient sounds. Their processing logic focuses on improving speech intelligibility, employing methods such as dynamic compression, amplitude limiting, noise reduction, and speech enhancement. While these methods can meet daily communication needs, they easily disrupt the original audio spectrum structure and dynamic range, leading to severe distortion in multimedia audio such as music and television audio. This makes them unsuitable for home theater and multimedia playback scenarios. Furthermore, traditional hearing aids are separate products from regular audio playback devices, requiring users to wear them separately. This results in insufficient adaptability to different scenarios and ease of use, failing to achieve high-fidelity hearing compensation. Summary of the Invention

[0004] This application provides a hearing aid device, method, and apparatus with adaptive hearing compensation function. This application is not a traditional hearing aid, but is based on the architecture of a conventional audio playback device, with a built-in dedicated hearing compensation model. It only processes the internal audio signal played by the device itself. Through full-band equalization linear compensation logic, it improves the accuracy of hearing loss fitting and audio fidelity, and achieves high-fidelity hearing fitting for hearing-impaired users. It solves the dual pain points of traditional audio playback devices being unable to fit hearing-impaired users and traditional hearing aids being prone to audio distortion and limited applicable scenarios.

[0005] In a first aspect, embodiments of this application provide a hearing aid device with adaptive hearing compensation function, comprising: a wireless communication module for receiving a user's hearing loss data; a main control chip, wherein the main control chip has a pre-trained hearing compensation model built in; the main control chip is used to extract hearing features based on the hearing loss data; preprocess the hearing loss data and the hearing features to construct an input dataset; input the input dataset to the pre-trained hearing compensation model, and output a personalized frequency response gain curve adapted to the user through inference by the hearing compensation model, wherein the personalized frequency response gain curve is a continuous gain curve across the entire frequency band, the gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies fitting accuracy constraints, curve smoothness constraints, and clinical hearing safety constraints; and an audio output module for performing full-band frequency-by-frequency gain compensation on the original digital audio signal according to the personalized frequency response gain curve before outputting it.

[0006] Secondly, embodiments of this application provide an adaptive hearing compensation method applied to a hearing aid device as described in any of the first aspects. The method includes: acquiring a user's hearing loss data; extracting hearing features based on the hearing loss data; preprocessing the hearing loss data and the hearing features to construct an input dataset; inputting the input dataset into a pre-trained hearing compensation model, and outputting a personalized frequency response gain curve adapted to the user through inference by the pre-trained hearing compensation model. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band, and the gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies fitting accuracy constraints, curve smoothness constraints, and clinical hearing safety constraints; and outputting the original digital audio signal after performing full-band frequency-by-frequency gain compensation based on the personalized frequency response gain curve.

[0007] Thirdly, embodiments of this application provide an adaptive hearing compensation device, the device comprising: an acquisition unit for acquiring a user's hearing loss data; a processing unit for extracting hearing features based on the hearing loss data; preprocessing the hearing loss data and the hearing features to construct an input dataset; inputting the input dataset into a pre-trained hearing compensation model, and outputting a personalized frequency response gain curve adapted to the user through inference by the pre-trained hearing compensation model, wherein the personalized frequency response gain curve is a continuous gain curve across the entire frequency band, the gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies fitting accuracy constraints, curve smoothness constraints, and clinical hearing safety constraints; and an output unit for performing full-band frequency-by-frequency gain compensation on the original digital audio signal based on the personalized frequency response gain curve and then outputting the signal.

[0008] As can be seen from the embodiments of this application, the hearing aid device includes a wireless communication module, a main control chip, and an audio output module. The main control chip has a built-in pre-trained hearing compensation model. The wireless communication module receives the user's hearing loss data, preprocesses it to construct an input dataset, and then inputs it into the hearing compensation model. The model infers and outputs a personalized frequency response gain curve that is continuous across the entire frequency band and whose gain amplitude is positively correlated with the hearing loss. Based on this curve, the audio output module performs full-band, frequency-point equalization linear gain compensation on the original digital audio signal played by the device itself. The entire process does not compress the audio dynamic range or change the original spectral structure, thus preserving high-fidelity sound quality. This invention differs from traditional hearing aids and conventional audio playback devices. It only targets the internal audio processing of the device, without the need for sound pickup and noise reduction. It is suitable for multimedia scenarios such as home theaters and TV audio. It can accurately adapt to the hearing characteristics of hearing-impaired users, achieve clear and balanced listening, and avoid sound quality distortion, taking into account compensation accuracy, user comfort, and scenario versatility. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the structure of a hearing aid provided in an embodiment of this application;

[0011] Figure 2 A schematic diagram of the training method for the hearing compensation model provided in this application embodiment;

[0012] Figure 3 A flowchart illustrating an adaptive hearing compensation method provided in an embodiment of this application;

[0013] Figure 4 A functional unit structure block diagram of an adaptive hearing compensation device provided in this application embodiment;

[0014] Figure 5 This is a schematic diagram of the structure of a main control chip provided in an embodiment of this application. Detailed Implementation

[0015] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0016] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but in some embodiments includes steps or units not listed, or in some embodiments includes other steps or units inherent to these processes, methods, products, or apparatuses.

[0017] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0018] In the embodiments of this application, "and / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone; A and B exist simultaneously; B exists alone. Among them, A and B can be singular or plural.

[0019] In this embodiment, the symbol " / " can indicate that the preceding and following objects are in an "or" relationship. Alternatively, the symbol " / " can also represent a division sign, i.e., performing a division operation. For example, A / B can mean A divided by B.

[0020] In the embodiments of this application, "at least one item" or its similar expression refers to any combination of these items, including any combination of a single item or a plurality of items. "One or more" means one or more, while "multiple" means two or more. For example, "at least one item" of a, b, or c can represent the following seven cases: a, b, c; a and b; a and c; b and c; a, b, and c. Each of a, b, and c can be an element or a set containing one or more elements.

[0021] In the embodiments of this application, "equal to" can be used with "greater than" and is applicable to technical solutions used when "greater than" is used; it can also be used with "less than" and is applicable to technical solutions used when "less than" is used. When "equal to" is used with "greater than", it is not used with "less than"; when "equal to" is used with "less than", it is not used with "greater than".

[0022] To address the aforementioned technical issues, this application provides a hearing aid device, method, and apparatus with adaptive hearing compensation. Based on a conventional audio playback device architecture, it is not a traditional hearing aid. It does not require the acquisition of ambient sound; instead, it processes only the internal audio signal played by the device itself. Through full-band equalization linear compensation logic, it adjusts the gain proportionally to each frequency point based on the user's hearing loss data, ensuring that the gain amplitude is positively correlated with the user's hearing loss at each frequency. The compensation process does not compress the audio dynamic range or alter the original spectral structure, thus fully preserving the original audio quality. Furthermore, this application ensures compensation accuracy, curve smoothness, and safety through parameterized constraints. It solves the problem of conventional audio playback devices being unsuitable for hearing-impaired users while avoiding the drawbacks of traditional hearing aids, such as sound quality distortion and limited application scenarios. It combines versatility, convenience, and high-fidelity compensation, significantly improving the listening experience for hearing-impaired users.

[0023] The hearing aid device with adaptive hearing compensation described in this application is not a traditional medical hearing aid, but rather an improved audio output device built on the foundation of conventional audio playback devices. The overall hardware uses the core architecture of general audio playback devices such as Bluetooth speakers, TV stereos, and desktop speakers, retaining the original audio decoding, signal amplification, sound output, and wireless communication modules. On this basis, a main control chip with a built-in hearing compensation model is added, forming a composite framework of "general audio playback hardware + dedicated adaptive hearing compensation module." Hearing compensation is achieved through software algorithms and a core computing chip, resulting in minimal hardware modifications and strong compatibility. The device can optionally be equipped with an environmental sound pickup module to collect ambient noise parameters, assisting in dynamic fine-tuning of the compensation gain to further optimize the listening effect under different acoustic environments. This module does not rely on the sound pickup module to achieve the core hearing compensation function; it is only an optional optimization configuration and does not change the core processing logic of the device.

[0024] From a technical perspective, the higher-level concept of this device is an audio playback device, belonging to the category of audio signal processing and output equipment technology. It possesses all the basic functions of a conventional audio playback device and is compatible with normal listening and use by ordinary users. The lower-level concept is a dedicated audio playback device with adaptive hearing compensation function. It is a refinement and improvement of conventional audio playback devices. Its core limitation is that it is an audio playback device designed for hearing-impaired users, capable of achieving full-frequency linear hearing compensation based on personalized hearing loss data, and processing only its own internal audio signals. This distinguishes it from ordinary audio playback devices, traditional hearing aids, hearing testing devices, and other products in the same field.

[0025] This device is applicable to a wide range of scenarios and is not limited by indoor or outdoor environments. It can be used in indoor and in-vehicle scenarios such as home audio-visual, TV audio, car audio, and desktop multimedia playback, as well as outdoor scenarios such as personal audio listening, outdoor leisure playback, and near-field personal listening. Whether indoors or outdoors, the device focuses on personalized gain compensation for the target user. It only performs precise adaptation on the audio signal played by the device itself and does not perform public amplification, public broadcasting, or environmental sound amplification. It can provide the target user with a stable, high-fidelity, and comfortable personalized listening effect under various acoustic conditions such as quiet environments, ordinary indoor environments, and outdoor background noise environments.

[0026] The hearing aids and related methods and devices provided in this application are described in detail below.

[0027] Please see Figure 1 , Figure 1 This is a schematic diagram of the hearing aid device structure provided in the embodiments of this application, such as... Figure 1 As shown, the hearing aid device includes:

[0028] The wireless communication module is used to receive the user's hearing loss data;

[0029] The main control chip has a pre-trained hearing compensation model built in. The main control chip is used to extract hearing features based on the hearing loss data. The hearing loss data and the hearing features are pre-processed to construct an input dataset. The input dataset is input into the pre-trained hearing compensation model, which infers and outputs a personalized frequency response gain curve adapted to the user. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band. The gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies the fitting accuracy constraint, curve smoothness constraint, and clinical hearing safety constraint.

[0030] The audio output module is used to perform full-band frequency-point gain compensation on the original digital audio signal according to the personalized frequency response gain curve before outputting it.

[0031] The wireless communication module is mainly used to enable wireless data interaction between the device and external terminal devices. It can use conventional wireless communication methods such as Bluetooth, Wi-Fi, and near field communication (NFC) to establish a stable wireless connection with terminal devices such as user mobile phones, tablets, and dedicated hearing testing terminals, thereby receiving and acquiring user-specific hearing loss data.

[0032] This hearing loss data is generated after the user completes a hearing test through a professional hearing testing institution, hospital ENT department, or compliant hearing testing equipment. It includes full-band hearing loss parameters at various frequency points, specifically core parameters such as the user's auditory sensitivity threshold at different frequency points and the degree of hearing loss classification. It can accurately reflect the user's hearing loss in each frequency band and provide basic data support for subsequent personalized hearing compensation.

[0033] For example, hearing loss data is measured using the Pure Tone Audiometry (PTA) standard at standard frequency points. The standard frequency point set is f={125,250,500,750,1000,1500,2000,3000,4000,6000,8000}Hz. The hearing loss value corresponding to each frequency point is expressed in decibel hearing levels (dB HL), and the measurement range is uniformly set to 0-120dB HL. This test standard and parameter range conform to clinical hearing testing specifications, which can ensure that the hearing loss data input into the model is accurate and effective, and provide a reliable basis for subsequent model inference and personalized gain compensation.

[0034] After receiving the user's hearing loss data transmitted by the wireless communication module, the main control chip first performs data preprocessing on the hearing loss data. Specifically, it adopts a normalization method to map the values ​​of the hearing loss data to a standard numerical range to ensure numerical stability and convergence speed during model training. The mathematical expression for the normalization process is as follows:

[0035]

[0036] The output value is the normalized value, ranging from [0,1]. This is the raw hearing loss data. This represents the minimum value among the hearing loss data. This represents the maximum value in the hearing loss data.

[0037] After normalizing the hearing loss data, further audiological feature engineering was performed. Combining expertise in audiology, key hearing features were extracted from the normalized raw hearing loss data. The specific feature types and calculation methods are as follows: One is the average hearing loss (PTA4). This indicator is calculated by averaging the hearing loss values ​​at four of the five key frequencies: 500Hz, 1000Hz, 2000Hz, and 4000Hz. It is a commonly used clinical hearing assessment indicator, and its calculation formula is as follows:

[0038]

[0039] Secondly, there is the high-frequency hearing loss index, used to assess the degree of hearing loss in the 3000Hz-8000Hz high-frequency range; thirdly, there is the hearing loss slope, used to characterize the changing trend of hearing loss values ​​in different frequency bands. By extracting the above key features such as average hearing loss, high-frequency hearing loss index, and hearing loss slope, we can more accurately capture the overall pattern and local characteristics of a user's hearing loss, providing a feature input foundation that meets the requirements of clinical and audiological specialties for subsequent model inference.

[0040] The main control chip integrates the hearing loss data after the above preprocessing and feature engineering with the extracted hearing features to construct a standardized input dataset.

[0041] The hearing compensation model adopts a core architecture design that combines a deep feedforward neural network with residual connections. It is specifically designed to generate a suitable frequency response gain curve based on the user's hearing loss data and audiological characteristics. Its overall network structure includes an input layer, an encoder layer, multiple hidden layers, a decoder layer, and an output layer.

[0042] In the model input stage, the raw hearing loss data is combined with the extracted audiological features to construct a fixed-dimensional input vector. This input vector is then passed through the input layer to the encoder layer, where linear transformations and activation function operations complete the mapping from the raw features to a high-dimensional representation space. The mapping process satisfies the expression:

[0043]

[0044] Where x is the input dataset, W1 is the encoder layer weight matrix, b1 is the encoder layer bias vector, and f( ) is the activation function, and h1 is the high-dimensional feature output by the encoder layer.

[0045] The high-dimensional features output from the encoder layer are fed into multiple hidden layers for deep feature extraction. These hidden layers employ a residual connection mechanism, which effectively addresses the vanishing gradient problem during deep network training, improving the training performance of deep networks. The feature extraction process satisfies the expression:

[0046]

[0047] in W is the input feature of the i-th hidden layer. i Let b be the weight matrix of the i-th hidden layer. i h is the bias vector of the i-th hidden layer. i 1 represents the output feature of the (i-1)th hidden layer, realizing residual fusion of cross-layer features, h i+1The output feature vector of the (i+1)th hidden layer is the final feature result after residual fusion.

[0048] The deep features output from the hidden layer are mapped from high-dimensional features to the target feature space by the decoder layer, and then fed into the output layer for final inference calculation. The output layer generates the target frequency response gain curve through linear transformation and activation function operations, and its operation process satisfies the expression:

[0049]

[0050] Among them, h n This represents the feature vector output by the decoder layer, where n is the total number of hidden layers. In other words, this feature vector is the result of the last hidden layer being transformed by the decoder layer. out The weight matrix represents the output layer weights, used to weight the output features h of the decoder layer. n Perform a linear transformation, the dimension of which is adapted to the decoder output features and the target output dimension; b out The bias vector representing the output layer is used to offset the result after the linear transformation; g( ) represents the activation function used in the output layer, and the appropriate function type is selected based on the numerical range characteristics of the gain value; y represents the final result vector generated by the model output layer, which contains the gain values ​​corresponding to 11 frequency points, i.e., the frequency response gain curve.

[0051] This hearing compensation model also integrates several key designs to ensure model performance, including: setting batch normalization operations in each layer of the network to accelerate model convergence and improve the stability of model training; introducing the Dropout regularization mechanism to effectively prevent overfitting by randomly deactivating some neurons; and using a nonlinear activation function to accurately capture the complex nonlinear mapping relationship between the degree of hearing loss and frequency response gain, ensuring that the gain value output by the model can match the actual characteristics of the user's hearing loss.

[0052] In some embodiments, the fitting accuracy constraint is that the difference between the gain amplitude at each frequency point and the clinical target gain amplitude does not exceed a preset accuracy threshold; the curve smoothing constraint is that the fluctuation difference between the gain amplitudes of adjacent frequency points does not exceed a preset smoothing threshold; and the clinical hearing safety constraint is that the gain amplitude at each frequency point does not exceed the preset hearing tolerance threshold corresponding to that frequency point.

[0053] Among them, the fitting accuracy constraint is that the difference between the gain amplitude of each frequency point output by the model and the clinical target gain amplitude of the corresponding frequency point does not exceed the preset accuracy threshold, so as to ensure the accuracy and reliability of hearing compensation; the curve smoothing constraint is that the fluctuation difference of the gain amplitude between adjacent frequency points does not exceed the preset smoothing threshold, so that the gain curve is continuous and smooth, avoiding audio distortion and hearing discomfort caused by sudden gain changes, thereby completely preserving the sound quality and dynamic range of the original audio; the clinical hearing safety constraint is that the gain amplitude output at each frequency point does not exceed the preset hearing tolerance threshold corresponding to that frequency point, ensuring that the compensated sound will not cause damage to the user's hearing and meet the requirements for safe use.

[0054] The preset thresholds are determined comprehensively based on clinical audiological standards, user-specific audiograms, device acoustic performance, and audio fidelity requirements. These thresholds are all personalized and configurable parameters, which are uniformly loaded by the main control chip during model training and online compensation phases, and together constitute the safety and quality constraint system for full-band linear hearing compensation in this application.

[0055] In summary, the complete workflow of a hearing aid device is as follows: After the hearing aid device is powered on, it first completes system initialization, and simultaneously performs self-tests and parameter presets on the wireless communication module, main control chip, and audio output module to ensure that each module is in normal working condition; then, the wireless communication module establishes a stable wireless connection with the user's mobile phone, tablet, or dedicated hearing testing terminal through Bluetooth, Wi-Fi, or near-field communication, and actively acquires the user's full-frequency hearing loss data issued by professional hearing testing institutions, hospitals, or compliant testing equipment;

[0056] After receiving hearing loss data, the main control chip performs preprocessing operations such as normalization calibration and outlier removal. At the same time, it extracts the corresponding frequency domain features and hearing loss grading features to construct a dataset that meets the input standards of the built-in hearing compensation model. After preprocessing, the main control chip inputs the dataset into the pre-trained hearing compensation model. Through model calculation and inference, it generates a full-band continuous personalized frequency response gain curve adapted to the user. The gain amplitude at each frequency point in the curve is positively correlated with the hearing loss at the corresponding frequency point, and the curve meets the triple parameter constraints of preset fitting accuracy, curve smoothness, and clinical hearing safety throughout the process.

[0057] When the hearing aid is connected to the original digital audio signal, the audio output module performs full-band, frequency-point equalization linear gain compensation on the audio signal based on the personalized frequency response gain curve output by the main control chip. The compensation process does not compress the dynamic range of the audio or change the original spectral structure; it only performs personalized loudness adaptation for the target user. Finally, the compensated audio signal is amplified and converted into an acoustic signal for output, providing hearing-impaired users with clear, balanced, and high-fidelity audio listening effects. If the device is equipped with an optional environmental sound pickup module, it can simultaneously collect ambient noise parameters to assist the main control chip in dynamically fine-tuning the compensation gain, further optimizing listening comfort in different scenarios.

[0058] As can be seen from the embodiments of this application, the hearing aid device includes a wireless communication module, a main control chip, and an audio output module. The main control chip has a built-in pre-trained hearing compensation model. The wireless communication module receives the user's hearing loss data, preprocesses it to construct an input dataset, and then inputs it into the hearing compensation model. The model infers and outputs a personalized frequency response gain curve that is continuous across the entire frequency band and whose gain amplitude is positively correlated with the hearing loss. Based on this curve, the audio output module performs full-band, frequency-point equalization linear gain compensation on the original digital audio signal played by the device itself. The entire process does not compress the audio dynamic range or change the original spectral structure, thus preserving high-fidelity sound quality. This invention differs from traditional hearing aids and conventional audio playback devices. It only targets the internal audio processing of the device, without the need for sound pickup and noise reduction. It is suitable for multimedia scenarios such as home theaters and TV audio. It can accurately adapt to the hearing characteristics of hearing-impaired users, achieve clear and balanced listening, and avoid sound quality distortion, taking into account compensation accuracy, user comfort, and scenario versatility.

[0059] Please see Figure 2 , Figure 2 This is a schematic diagram of the training method for the hearing compensation model provided in the embodiments of this application, as shown below. Figure 2 As shown, the training process of the hearing compensation model includes the following steps S201-S206:

[0060] Step S201: Initialize the model parameters of the hearing compensation model and the corresponding adaptive gradient optimizer, and configure the training control parameters.

[0061] First, the model parameters are initialized by unifying all trainable parameters in the hearing compensation model into an initial parameter set θ0. This parameter set specifically includes the weight matrix W1 and bias vector b1 of the encoder layer, and the weight matrix W of the multiple residual hidden layers. i With bias vector b i The weight matrix W of the output layer out With bias vector b outFor the aforementioned trainable parameters, a Gaussian distribution with a mean of 0 and a variance of 0.01 is used to randomly initialize the weight matrix of each layer, and the bias vector of each layer is initialized to 0 at the same time to ensure that the initial parameters of the model have no systematic deviation, thus laying a stable foundation for subsequent iterative optimization.

[0062] Next, the adaptive gradient optimizer is initialized. In this embodiment, the Adam adaptive gradient optimizer is used to implement the iterative update of the model parameters. The specific initialization operations include: initializing the first-order moment estimate variable m0 and the second-order moment estimate variable v0 of the optimizer to all-zero vectors that perfectly match the dimension of the parameter set θ0; and preset the core hyperparameters of the optimizer, including the first-order moment decay coefficient β1=0.9, the second-order moment decay coefficient β2=0.999, and the numerical stability constant. =10 -8 This ensures the adaptive capability of the optimizer's gradient updates and the stability of numerical computation.

[0063] Finally, configure the complete set of training control parameters. Specific preset parameters include: total number of training iterations T, and number of warm-up iterations t. warmup Adjust the maximum step size α max With minimum value α min Reconstruction loss weighting coefficient λ1, smoothness loss weighting coefficient λ2, clinical constraint loss weighting coefficient λ3; L2 regularization coefficient λ L2 The parameters include: Dropout random inactivation probability p; training batch sample size; threshold for determining the number of rounds with no continuous decrease in validation set loss; threshold for loss fluctuation amplitude; and the standard deviation of Gaussian noise used in the data augmentation stage. By configuring the above full training parameters, the model training process is standardized, taking into account model convergence efficiency, fitting accuracy, and clinical safety constraints.

[0064] Step S202: Extract batch sample data, input the sample data into the model, and generate an initial frequency response gain curve.

[0065] The sample data includes hearing loss data, hearing characteristics, full-band clinical target gain curves, and clinical hearing constraint parameters; the gain amplitude at each frequency point in the full-band clinical target gain curve is positively correlated with the hearing loss at the corresponding frequency point.

[0066] During the data loading process, small Gaussian noise is applied to the hearing features for data augmentation to increase the diversity of the training samples and improve the model's generalization ability. The augmented hearing features are then used as input to the hearing compensation model. After forward propagation through the encoder layer, multiple residual hidden layers, and the output layer, the initial frequency response gain curve of the model output is obtained.

[0067] Step S203: Construct positive correlation fitting constraint loss, curve smoothness loss and clinical safety loss respectively. Weight each loss according to preset weights and aggregate them together. Combine with regularization constraints to obtain the final optimized loss.

[0068] After obtaining the initial frequency response gain curve in step S202, since this curve is only the initial inference output of the model, it cannot simultaneously guarantee the fitting accuracy with the clinical target, the smoothness and continuity of the frequency response curve, and the clinical safety of hearing compensation. Therefore, it is necessary to construct a multi-dimensional constraint loss to normalize and constrain the model output, guiding the model to gradually generate a personalized frequency response gain curve that meets actual usage needs during iterative training. Specifically, positive correlation fitting constraint loss, curve smoothness loss, and clinical safety loss are constructed respectively. The positive correlation fitting constraint loss is used to reduce the deviation between the model output and the clinical standard target, ensuring the degree of matching between gain compensation and hearing loss; the curve smoothness loss is used to suppress drastic jumps in gain amplitude at adjacent frequency points, improving auditory comfort and acoustic stability; and the clinical safety loss is used to limit the sound pressure level after compensation to not exceed the user's safe tolerance range, avoiding secondary hearing damage.

[0069] In some embodiments, the positive correlation fitting constraint loss is used to constrain the deviation between the initial frequency response gain curve and the full-band clinical target gain curve, so that the full-band gain curve output by the model maintains a positive correlation between the gain amplitude at each frequency point and the hearing loss at the corresponding frequency point; the curve smoothness loss is determined based on the continuous consistency of the gain amplitude at adjacent frequency points of the initial frequency response gain curve; the clinical safety loss is determined based on the fit relationship between the gain amplitude of the initial frequency response gain curve and the clinical hearing constraint parameters.

[0070] The positive correlation fitting constraint loss is implemented using reconstruction loss. The core purpose of setting this loss is to ensure that the model's output gain curve accurately matches the standard target for clinical hearing fitting, strictly guaranteeing that the gain amplitude at each frequency point is positively correlated with the user's hearing loss level at that frequency. The more severe the hearing loss, the greater the corresponding gain compensation, thus preventing undercompensation or overcompensation. Its calculation formula is:

[0071]

[0072] Among them, L MSE Represents the reconstruction loss; N represents the total number of samples in the current batch; F represents the total number of frequency points contained in a single gain curve; n is the sample number, ranging from 1 to N; i is the frequency point number, ranging from 1 to F; G pred (f i ) represents the nth sample at the i-th frequency point f i The model prediction gain value at point G; target (f i ) represents the nth sample at the i-th frequency point fi Clinical target gain value at the location.

[0073] This loss measures the deviation between the predicted gain and the clinical target gain in the form of mean squared error, so that the gain amplitude output by the model changes synchronously with hearing loss, thus achieving positive correlation constraint.

[0074] The device in this application must fully preserve the original audio quality. The initial linear gain curve is prone to abrupt changes in gain amplitude between adjacent frequency points. These abrupt changes directly lead to audio timbre distortion, harshness, and sound fragmentation, damaging the high-fidelity playback effect. Therefore, this loss is set to constrain the continuous smoothness of the gain curve across the entire frequency band, ensuring audio quality and listening comfort under linear compensation logic, and preventing gain jumps from affecting the original listening experience. The formula for calculating the curve smoothness loss is:

[0075]

[0076] Among them, L smooth Indicates the curve smoothness loss; N represents the total number of samples in the batch; F represents the total number of frequency points; n is the sample number; i is the intermediate frequency point number, ranging from 2 to F-1; G pred (f i 1) G pred (f i ), G pred (f i+1 ) represent the model prediction gain values ​​corresponding to the (i-1), i, and i+1 frequency points, respectively.

[0077] This loss avoids abrupt changes in gain between adjacent frequencies by constraining the second-order difference of the gain curve, ensuring a continuous and smooth curve and improving listening comfort.

[0078] The clinical safety loss is achieved through clinical constraint loss. Although the device in this application is based on a modified conventional audio playback device, it still needs to strictly ensure the safety of hearing-impaired users. The compensation gain must comply with clinical hearing safety standards to avoid excessive gain causing damage to the user's hearing, while not affecting the device's basic function as a general-purpose audio playback device. Therefore, this loss is set to rigidly constrain the safety boundary of linear gain compensation, balancing safety and universality. The calculation formula is:

[0079]

[0080] Among them, L clinical SPL represents clinical constraint loss; N represents the total number of samples in the batch; F represents the total number of frequency points; n is the sample number; i is the frequency point number; SPL input (f i ) represents the input sound pressure level corresponding to the i-th frequency point; G pred (f i) represents the model prediction gain value corresponding to the i-th frequency point; UCL(f i ) represents the user discomfort threshold corresponding to the i-th frequency point; max(0, The value indicated is the larger of 0 and the value within the parentheses. This loss is used to penalize compensation gain that exceeds the safe range, ensuring that the total sound pressure level after compensation does not exceed the user's tolerance limit and meets clinical hearing safety constraints.

[0081] After calculating the three individual losses, the losses are weighted and aggregated according to preset weights to obtain the total training loss:

[0082]

[0083] Among them, L total λ1 represents the total training loss; λ2 and λ3 are the preset weight coefficients corresponding to the reconstruction loss, smoothness loss, and clinical constraint loss, respectively. They are used to adjust the optimization priority of each constraint objective according to the actual application scenario, so that the model can achieve a balance between fitting accuracy, curve smoothness, and hearing safety.

[0084] To further prevent overfitting during training and improve generalization ability and robustness, L2 regularization constraints are superimposed on the total training loss, resulting in the final optimized loss used for parameter updates:

[0085]

[0086] Among them, L reg λ represents the final optimization loss; L2 W is the L2 regularization coefficient; l is the model network layer number; W l Let W be the weight matrix of the l-th layer network; l || F The weight matrix W l The Frobenius norm is used. Weight decay is achieved by applying a penalty to the weight matrix, avoiding the model's over-reliance on local noise in the training data and ensuring stable, compliant, and reliable personalized frequency response gain curves for different users' hearing characteristics.

[0087] This operation applies a penalty to the weights of each layer of the model to achieve weight decay, avoiding excessively large model weight values ​​and overlearning of noisy data in the training samples. It improves the model's adaptability to different user hearing data and different general audio device hardware, and finally outputs an optimized model that combines accurate linear compensation, high-fidelity sound quality, safe use and strong versatility.

[0088] Step S204: Calculate the parameter gradient of the model parameters based on the final optimization loss. The parameter gradient is used to characterize the adjustment trend and relative adjustment magnitude of the model parameters.

[0089] Following step S203, the final optimized loss L is obtained, which balances fitting accuracy, curve smoothness, clinical safety, and generalization constraints. reg Afterwards, the loss optimization objective needs to be transformed into a specific basis for adjusting the model parameters through parameter gradients. This step, based on the final optimized loss, uses the chain rule to accurately calculate the parameter gradients, laying the foundation for subsequent iterative updates of the model parameters.

[0090] In practice, the ultimate optimization loss L is used. reg To optimize the objective, the set of model parameters θ corresponding to the current iteration t is... t Taking the derivative, we obtain the gradient of the parameters g. t The calculation formula is:

[0091]

[0092] Among them, g t This represents the parameter gradient corresponding to this iteration, whose dimension is related to the model parameter θ. t Exact match, θ t It includes all weight matrices and bias vectors of the model encoder layer, residual hidden layer, and output layer; θ L represents the gradient calculation operation with respect to the model parameters θ. reg This represents the final optimized loss obtained from the aforementioned calculations.

[0093] The parameter gradients calculated here directly correspond to the actual adjustment direction of the model parameters: the positive or negative value of the gradient indicates the adjustment trend of the parameter. A positive gradient means that the current value of the parameter will increase the final optimization loss, and it needs to be adjusted in the direction of decreasing the parameter value, thereby reducing the deviation between the model's predicted gain curve and the clinical target, ensuring that the gain curve is smooth and meets clinical safety constraints; a negative gradient means that the current value of the parameter will make the final optimization loss too large, and it needs to be adjusted in the direction of increasing the parameter value, thereby optimizing the model output effect. The absolute value of the gradient corresponds to the relative adjustment magnitude of the parameter. The larger the absolute value, the more significant the impact of the parameter on the final optimization loss, and the greater the impact on the fitting accuracy, smoothness, or clinical safety of the full-band gain curve, the greater the adjustment magnitude required in subsequent iterations; the smaller the absolute value, the less significant the impact of the parameter on the loss and gain curve, and only small adjustments are needed in subsequent iterations to avoid over-adjustment that could damage the model stability.

[0094] The calculation process employs the chain rule of differentiation, which propagates backward from the final optimized loss layer by layer, sequentially solving the gradients of the parameters in the output layer, multiple residual hidden layers, and encoder layer. This ensures that each trainable parameter can obtain a corresponding gradient value, and that the gradient corresponds one-to-one with the parameter without dimensional bias. This guarantees that subsequent parameter updates can accurately align with the direction of loss optimization, continuously optimizing the model's adaptive hearing compensation effect.

[0095] Step S205: The model parameters are iteratively updated using a dynamic step size scheduling strategy in conjunction with an adaptive gradient optimizer.

[0096] After the parameter gradient calculation is completed in step S204, if the parameters are updated directly using a fixed step size, it is easy to encounter problems such as slow convergence in the early stage of training and difficulty in convergence in the later stage of oscillation. Moreover, a single gradient update method cannot adapt to the distribution characteristics of different parameter gradients. Therefore, this step combines dynamic step size scheduling and adaptive gradient optimization to balance training efficiency and parameter update stability. This ensures that the model can quickly fit the clinical target gain curve while avoiding excessive parameter update amplitude that could damage the curve smoothness and clinical safety constraints.

[0097] In some embodiments, in the process of using a dynamic step-size scheduling strategy in conjunction with an adaptive gradient optimizer to complete the iterative update of the model, the main control chip is specifically used to: determine the adjustment step size of the model parameters in the current iteration round through the dynamic step-size scheduling strategy, wherein the adjustment step size is the correction magnitude of the model parameters in a single iteration; adaptively correct the parameter gradient through the adaptive gradient optimizer; and adjust the model parameters according to the adjustment step size and the corrected parameter gradient to complete the iterative update of the model.

[0098] In the specific implementation process, the adjustment step size for the current iteration round is first determined through a dynamic step size scheduling strategy:

[0099] In some embodiments, regarding determining the adjustment step size of the model parameters for the current iteration round through the dynamic step size scheduling strategy, the main control chip is specifically used to: divide the dynamic step size adjustment process into a preheating stage and an annealing stage, using a preset number of preheating rounds as the stage division boundary; during the preheating stage, if the number of iteration rounds does not exceed the preset number of preheating rounds, control the adjustment step size to increase linearly with the iteration process until it reaches a preset maximum step size; during the annealing adjustment stage, if the number of iteration rounds exceeds the preset number of preheating rounds, control the adjustment step size to decrease smoothly with the iteration process until it drops to a preset minimum step size.

[0100] This application uses the current iteration round number t and the preset preheating round number t warmup Based on the size relationship, the calculation and control of the adjustment step size are carried out in two stages. The specific implementation method is as follows:

[0101] The first stage is the warm-up stage, which satisfies the iteration round condition t≤t. warmup This means that the current iteration number has not exceeded the preset warm-up number of iterations. This stage is the initial transition period for model training. Its core function is to stabilize the parameter gradient direction through small step-size iterations, adapt to the distribution of hearing sample data, and avoid the parameters deviating from the optimal direction due to an excessively large initial step size, thus breaking the positive correlation between gain amplitude and hearing loss. This stage is performed according to the formula... Calculate the current iteration adjustment step size α t , where α max The preset maximum adjustment step size is t, where t is the current iteration number. warmup The preset number of preheating cycles is used; through this linear operation, the adjustment step size starts from 0 and increases linearly and uniformly with the number of iterations until the preset maximum adjustment step size α is reached. max Gradually transition to the normal training rhythm to ensure a smooth and orderly initial parameter update.

[0102] The second stage is the cosine annealing stage, which satisfies the iteration round condition t > t. warmup This means that the current iteration rounds exceed the preset warm-up rounds. At this stage, the model has completed initial fitting, and its core function is to gradually reduce the parameter update amplitude, avoid parameter oscillations, accurately optimize the smoothness of the gain curve and clinical safety constraints, allowing the model to gradually converge to the optimal parameter state. This stage follows the formula... Calculate the current iteration adjustment step size α t , where α min The preset minimum adjustment step size is T, and the preset total number of training iterations is T. The introduction of the cosine function enables a smooth, non-linear decay of the adjustment step size, avoiding training stagnation caused by a sudden drop in step size, and eventually gradually reducing it to the preset minimum adjustment step size α. min This enables refined convergence of the model.

[0103] As can be seen, in this embodiment, the dynamic step size scheduling strategy adjusts the step size through phased differentiated regulation. On the one hand, the step size decreases non-linearly and smoothly with the iteration process, rather than dropping abruptly, effectively avoiding the problems of parameter oscillation and repeated fluctuations in loss in the later stages of model training, ensuring that the model continues to converge toward the optimal state. On the other hand, by gradually reducing the magnitude of single parameter correction, the gain curve can be finely optimized, focusing on refining the curve smoothness and clinical safety threshold constraints. This approach does not compress the dynamic range of the audio or destroy the original spectral structure, fully preserving the original sound quality, while ensuring that the compensation gain accurately meets clinical safety requirements. At the same time, it improves the model's generalization ability, adapts to different user hearing characteristics and the hardware characteristics of different general audio playback devices, and achieves stable and high-fidelity adaptive hearing compensation.

[0104] In some embodiments, in adaptively correcting the parameter gradient using the adaptive gradient optimizer, the main control chip is specifically configured to: iteratively update the first-order moment estimate and the second-order moment estimate of the adaptive gradient optimizer based on the parameter gradient, and perform bias correction on the updated first-order moment estimate and second-order moment estimate; the first-order moment estimate is used to characterize the iterative mean trend of the parameter gradient, and the second-order moment estimate is used to characterize the iterative variance trend of the parameter gradient; and perform normalization smoothing processing on the parameter gradient based on the bias-corrected first-order moment estimate and second-order moment estimate to obtain the corrected parameter gradient.

[0105] This application uses the Adam adaptive gradient optimizer to complete the above-mentioned adaptive gradient correction and subsequent parameter updates, fully meeting the model's hearing compensation optimization requirements. The specific implementation steps are as follows:

[0106] First, a moment estimation update is performed, based on the gradient g of the current round parameter calculated in step S204. t The first and second moment estimates of the optimizer are updated iteratively, respectively, according to the following formula:

[0107] m t =β1m t-1 +(1-β1)g t

[0108] v t =β2v t-1 +(1-β2)g t 2

[0109] Where, m t This is the first-order moment estimate for the current round, used to characterize the iterative mean trend of the parameter gradient. It integrates historical gradient directions, weakens the random fluctuations of single gradients, ensures that the parameter update direction always aligns with the loss minimization objective, and avoids gain curve distortion caused by gradient jitter; v t β1 represents the estimated second moment of the current iteration, used to characterize the iterative variance trend of the parameter gradient, record the gradient fluctuation amplitude, suppress large gradients, and amplify small gradients, ensuring that model parameters with different sensitivities are updated appropriately; β1 and β2 are preset hyperparameters of the Adam optimizer, with common values ​​of 0.9 and 0.999 respectively. t-1 v t-1 This is the moment estimate from the previous round, enabling iterative inheritance of gradient information.

[0110] Next, bias correction is performed to eliminate the calculation bias caused by the small value of the moment estimate in the early stage of training, and to ensure the accuracy of gradient correction. The corresponding formula is:

[0111]

[0112]

[0113] In the formula, To correct the estimate of the second first moment, To correct the second-order moment estimate, where t is the current iteration round, the bias correction addresses the numerical offset caused by insufficient iteration of the moment estimate in the initial iteration stage, thus adapting to the gradient correction requirements throughout the entire model training cycle.

[0114] Finally, the parameter update is completed by dynamically adjusting the step size. Based on the two types of moment estimates after bias correction, the original parameter gradient is normalized and smoothed to obtain the corrected stable gradient, which is then substituted into the adjustment step size α obtained by dynamic scheduling. t The model parameters are iteratively updated using the following formula:

[0115]

[0116] Where, θ t+1 For the updated model parameters, θ t For the current round parameters, Take 10 -8 , which is a numerical stability constant, to avoid calculation abnormalities caused by a denominator of zero; through this correction and update logic, gradient fluctuations can be smoothed, and excessive parameter update amplitude can be prevented from damaging the smoothness of the gain curve and clinical safety. It can also accelerate the model convergence speed and ensure that the full-band gain curve output by the model always meets the core requirements that the gain amplitude is positively correlated with hearing loss, does not change the original audio quality, and is in compliance with safety and compliance.

[0117] As can be seen from the embodiments of this application, the adaptive gradient optimizer effectively suppresses gradient oscillations and eliminates initial iteration bias by iteratively updating and correcting the first and second moments of the parameter gradient and implementing normalization smoothing. This makes the parameter updates more stable and smooth, ensuring rapid model convergence while avoiding excessive parameter update amplitude that could lead to gain curve distortion or compensation abnormalities. Consequently, the output full-band linear gain curve better fits the positive correlation constraint of hearing loss, has higher smoothness, and stronger clinical safety. At the same time, it improves the model's generalization ability and operational stability on general audio playback devices.

[0118] Step S206: When the preset termination condition is reached, the optimal model parameters after training are obtained. The hearing compensation model supported by these optimal model parameters can be directly deployed to the device's main control chip to achieve high-fidelity, personalized linear compensation of internal audio signals.

[0119] In some embodiments, the preset termination condition is any of the following: the number of iteration rounds reaches a preset maximum number of iteration rounds threshold; or, the fluctuation range of the final optimization loss within a consecutive preset number of rounds is less than a preset fluctuation range.

[0120] Specifically, the main control chip checks whether the termination conditions are met after each iteration. The two types of termination conditions control the training endpoint from the dimensions of "training duration" and "optimization effect," respectively, to adapt to the needs of different training scenarios.

[0121] A pre-set threshold for the maximum number of iterations during model training (e.g., 1000 or 2000 iterations) is established. Training terminates directly when the actual number of iterations reaches this threshold. This is a hard termination rule. The core technical effect is to avoid excessive training time and hardware resource consumption caused by unlimited model iteration. It adapts to the "lightweight training and rapid deployment" characteristics of general audio devices such as Bluetooth speakers and TV stereos, while preventing overfitting caused by excessive iteration and ensuring model compatibility in different use cases.

[0122] Real-time monitoring of the final optimization loss L in each iteration reg The system calculates the difference between the maximum and minimum values ​​within a preset number of rounds (e.g., 50 or 100 rounds). If this difference is less than a preset fluctuation range (e.g., 0.001), the model is considered to have converged to a stable state, and training is terminated. This condition is an effect-oriented termination rule. Its core technical effect is to accurately capture the model's convergence point: when the loss fluctuation is extremely small, it indicates that there is no significant room for optimization of the model parameters, and the accuracy, smoothness, and safety of the output gain curve have reached their optimal levels. Terminating training at this point avoids invalid iterations and ensures the stability of the model's compensation effect, ensuring that the device's compensation of the internal audio is always high-fidelity and compliant.

[0123] When any termination condition is met, parameter iteration stops and the current model parameters are fixed as the optimal parameters. After the parameters are deployed, the device can quickly generate a personalized gain curve that conforms to the full-band linear compensation logic based on the user's hearing loss data. It only performs adaptation processing on the internal audio signal it plays, which not only solves the pain point that conventional audio devices cannot adapt to hearing-impaired users, but also avoids the drawbacks of traditional hearing aids such as sound quality distortion and limited scenarios, taking into account universality, convenience and compensation effect.

[0124] Based on the above detailed description of the hardware structure, module functions, and interaction implementation of the hearing aid device, in order to further clarify the technical solution of this application and make its protection scope more complete and its technical logic clearer, this application also provides an adaptive hearing compensation method. This method is applied to the aforementioned hearing aid device and can realize full-band, high-fidelity, safe and compliant personalized hearing compensation based on the device hardware architecture. It is based on the same inventive concept, technical principle and constraints as the device embodiment. The specific implementation method is as follows.

[0125] Please see Figure 3 , Figure 3This is a flowchart illustrating an adaptive hearing compensation method provided in an embodiment of this application, as shown below. Figure 3 As shown, the method includes the following steps S301-S305:

[0126] Step S301: Obtain the user's hearing loss data;

[0127] Step S302: Extract hearing features based on the hearing loss data;

[0128] Step S303: Preprocess the hearing loss data and the hearing features to construct an input dataset;

[0129] Step S304: Input the input dataset into the pre-trained hearing compensation model. The pre-trained hearing compensation model infers and outputs a personalized frequency response gain curve that fits the user. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band. The gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies the fitting accuracy constraint, curve smoothness constraint, and clinical hearing safety constraint.

[0130] Step S305: Perform full-band frequency-by-frequency gain compensation on the original digital audio signal according to the personalized frequency response gain curve and then output it.

[0131] In some embodiments, the training process of the hearing compensation model includes: initializing the model parameters and corresponding adaptive gradient optimizer of the hearing compensation model, and configuring training control parameters; extracting batch sample data, inputting the sample data into the model, and generating an initial frequency response gain curve; the sample data includes the hearing loss data, hearing features, full-band clinical target gain curve, and clinical hearing constraint parameters; the gain amplitude of each frequency point in the full-band clinical target gain curve is positively correlated with the hearing loss at the corresponding frequency point; constructing positive correlation fitting constraint loss, curve smoothness loss, and clinical safety loss respectively, weighting and aggregating each loss according to preset weights, and combining regularization constraints to obtain the final optimized loss; calculating the parameter gradient of the model parameters according to the final optimized loss, the parameter gradient being used to characterize the adjustment trend and relative adjustment magnitude of the model parameters; using a dynamic step size scheduling strategy in conjunction with the adaptive gradient optimizer to complete the iterative update of the model parameters; and obtaining the optimal model parameters after training when iterating to a preset termination condition.

[0132] In some embodiments, the positive correlation fitting constraint loss is used to constrain the deviation between the initial frequency response gain curve and the full-band clinical target gain curve, so that the full-band gain curve output by the model maintains a positive correlation between the gain amplitude at each frequency point and the hearing loss at the corresponding frequency point; the curve smoothness loss is determined based on the continuous consistency of the gain amplitude at adjacent frequency points of the initial frequency response gain curve; the clinical safety loss is determined based on the fit relationship between the gain amplitude of the initial frequency response gain curve and the clinical hearing constraint parameters.

[0133] In some embodiments, a dynamic step size scheduling strategy combined with an adaptive gradient optimizer is used to complete the iterative update of the model, including: determining the adjustment step size of the model parameters in the current iteration round through the dynamic step size scheduling strategy, wherein the adjustment step size is the correction magnitude of the model parameters in a single iteration; adaptively correcting the parameter gradient through the adaptive gradient optimizer; and adjusting the model parameters according to the adjustment step size and the corrected parameter gradient to complete the iterative update of the model.

[0134] In some embodiments, determining the adjustment step size of the model parameters for the current iteration round through the dynamic step size scheduling strategy includes: dividing the dynamic step size adjustment process into a preheating stage and an annealing stage, with a preset number of preheating rounds as the stage division boundary; during the preheating stage, if the number of iteration rounds does not exceed the preset number of preheating rounds, controlling the adjustment step size to increase linearly with the iteration process until it reaches a preset maximum step size; during the annealing adjustment stage, if the number of iteration rounds exceeds the preset number of preheating rounds, controlling the adjustment step size to decrease smoothly with the iteration process until it drops to a preset minimum step size.

[0135] In some embodiments, adaptively correcting the parameter gradient using the adaptive gradient optimizer includes: iteratively updating the first-order moment estimate and the second-order moment estimate of the adaptive gradient optimizer based on the parameter gradient, and performing bias correction on the updated first-order moment estimate and second-order moment estimate; the first-order moment estimate is used to characterize the iterative mean trend of the parameter gradient, and the second-order moment estimate is used to characterize the iterative variance trend of the parameter gradient; and normalizing and smoothing the parameter gradient based on the bias-corrected first-order moment estimate and second-order moment estimate to obtain the corrected parameter gradient.

[0136] In some embodiments, the preset termination condition is any of the following: the number of iteration rounds reaches a preset maximum number of iteration rounds threshold; or, the fluctuation range of the final optimization loss within a consecutive preset number of rounds is less than a preset fluctuation range.

[0137] In some embodiments, the fitting accuracy constraint is that the difference between the gain amplitude at each frequency point and the clinical target gain amplitude does not exceed a preset accuracy threshold; the curve smoothing constraint is that the fluctuation difference between the gain amplitudes of adjacent frequency points does not exceed a preset smoothing threshold; and the clinical hearing safety constraint is that the gain amplitude at each frequency point does not exceed the preset hearing tolerance threshold corresponding to that frequency point.

[0138] Since the technical solutions of the method embodiments of this application are completely corresponding to and consistent with the above system embodiments, in order to avoid repeated descriptions, the specific content of the method embodiments can be found in the steps and processing flow executed by each module, unit, platform and intelligent agent in the above device embodiments, and will not be repeated here.

[0139] The above primarily describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the server includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0140] This application embodiment can divide the server into functional units according to the above method example. For example, each function can be divided into different functional units, or two or more functions can be integrated into one processing module. The integrated unit can be implemented in hardware or as a software program module. It should be noted that the unit division in this application embodiment is illustrative and only represents a logical functional division, while other division methods may be used in actual implementation.

[0141] In the case of using integrated units, please refer to Figure 4 , Figure 4 A functional unit structural block diagram of an adaptive hearing compensation device provided in this application embodiment, the device comprising:

[0142] Acquisition unit 401 is used to acquire the user's hearing loss data;

[0143] The processing unit 402 is used to extract hearing features based on the hearing loss data; preprocess the hearing loss data and the hearing features to construct an input dataset; input the input dataset into a pre-trained hearing compensation model; and output a personalized frequency response gain curve adapted to the user after the pre-trained hearing compensation model infers and outputs the personalized frequency response gain curve. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band, and the gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies the fitting accuracy constraint, curve smoothness constraint, and clinical hearing safety constraint.

[0144] The output unit 403 is used to perform full-band frequency-point gain compensation on the original digital audio signal according to the personalized frequency response gain curve and then output it.

[0145] As can be seen, in this embodiment, the main control chip has a built-in pre-trained hearing compensation model. Through model inference, it outputs a personalized frequency response gain curve that is continuous across the entire frequency band and whose gain amplitude is positively correlated with hearing loss. Based on this curve, the audio output module performs full-band, frequency-point-by-frequency equalization linear gain compensation on the original digital audio signal played by the device itself. This process does not compress the audio dynamic range or change the original spectral structure, preserving high-fidelity sound quality. Unlike traditional hearing aids and conventional audio playback devices, this method only processes the internal audio of the device, eliminating the need for sound pickup and noise reduction. It accurately adapts to the hearing characteristics of hearing-impaired users, achieving clear and balanced listening, while avoiding sound quality distortion, thus balancing compensation accuracy, user comfort, and versatility.

[0146] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a main control chip provided in an embodiment of this application, as shown below. Figure 5 As shown, the main control chip 50 includes a processor 501, a memory 503, a communication interface 502, and a computer program 5031. The computer program 5031 is stored in the memory 503 and configured to be executed by the processor 501. The program includes methods and apparatuses for executing a hearing aid device, method, and apparatus with adaptive hearing compensation function as described in the above embodiments.

[0147] This application provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implement the steps of any possible embodiment of the method.

[0148] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0149] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0150] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0151] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0152] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0153] If the aforementioned integrated units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0154] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0155] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A hearing aid device having an adaptive hearing compensation function, characterized by include: The wireless communication module is used to receive the user's hearing loss data; The main control chip has a pre-trained hearing compensation model built in; the main control chip is used to extract hearing features based on the hearing loss data. The hearing loss data and hearing features are preprocessed to construct an input dataset. The input dataset is then input into a pre-trained hearing compensation model. The hearing compensation model infers and outputs a personalized frequency response gain curve that is adapted to the user. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band. The gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies the fitting accuracy constraint, curve smoothness constraint, and clinical hearing safety constraint. The audio output module is used to perform full-band frequency-point gain compensation on the original digital audio signal according to the personalized frequency response gain curve before outputting it. The training process of the hearing compensation model includes: The model parameters and corresponding adaptive gradient optimizer of the hearing compensation model are initialized, and the training control parameters are configured. Batch sample data is extracted and input into the model to generate an initial frequency response gain curve. The sample data includes hearing loss data, hearing characteristics, full-band clinical target gain curve, and clinical hearing constraint parameters. The gain amplitude at each frequency point in the full-band clinical target gain curve is positively correlated with the hearing loss at the corresponding frequency point. The hearing loss data includes average hearing loss, high-frequency hearing loss index, and hearing loss slope. We construct positive correlation fitting constraint loss, curve smoothness loss, and clinical safety loss respectively. We then aggregate each loss according to preset weights and combine them with regularization constraints to obtain the final optimized loss. The parameter gradient of the model parameters is calculated based on the final optimization loss. The parameter gradient is used to characterize the adjustment trend and relative adjustment magnitude of the model parameters. A dynamic step-size scheduling strategy combined with an adaptive gradient optimizer is used to complete the iterative update of the model parameters; When the preset termination condition is reached, the optimal model parameters after training are obtained.

2. The hearing aid device according to claim 1, characterized in that, The positive correlation fitting constraint loss is used to constrain the deviation between the initial frequency response gain curve and the full-band clinical target gain curve, so that the full-band gain curve output by the model maintains a positive correlation between the gain amplitude at each frequency point and the hearing loss at the corresponding frequency point; The curve smoothness loss is determined based on the continuous consistency of the gain amplitude at adjacent frequency points of the initial frequency response gain curve; The clinical safety loss is determined based on the fit between the gain amplitude of the initial frequency response gain curve and the clinical hearing constraint parameters.

3. The hearing assistance device of claim 1, wherein, In employing a dynamic step-size scheduling strategy in conjunction with an adaptive gradient optimizer to complete the iterative update of the model, the main control chip is specifically used for: The adjustment step size of the model parameters in the current iteration is determined by the dynamic step size scheduling strategy, and the adjustment step size is the correction magnitude of the model parameters in a single iteration. The parameter gradient is adaptively corrected using the adaptive gradient optimizer. The model parameters are adjusted according to the adjustment step size and the corrected parameter gradient to complete the iterative update of the model.

4. The hearing assistance device of claim 3, wherein, In determining the adjustment step size of the model parameters for the current iteration round using the dynamic step-size scheduling strategy, the main control chip is specifically used for: The process of dynamically adjusting the step size is divided into a preheating stage and an annealing stage, with the number of preset preheating cycles as the dividing line. During the preheating phase, if the number of iterations does not exceed the preset number of preheating cycles, the adjustment step size is controlled to increase linearly with the iteration process until it reaches the preset maximum step size. During the annealing adjustment phase, if the number of iterations exceeds the preset number of preheating cycles, the adjustment step size is controlled to decrease smoothly with the iteration process until it drops to the preset minimum step size.

5. The hearing assistance device of claim 3, wherein, In terms of adaptively correcting the parameter gradient using the adaptive gradient optimizer, the main control chip is specifically used for: Based on the parameter gradient, the first-order moment estimate and the second-order moment estimate of the adaptive gradient optimizer are iteratively updated, and the updated first-order moment estimate and the second-order moment estimate are subjected to bias correction. The first-order moment estimate is used to characterize the iterative mean trend of the parameter gradient, and the second-order moment estimate is used to characterize the iterative variance trend of the parameter gradient. The parameter gradient is normalized and smoothed based on the first-order moment estimate and the second-order moment estimate after bias correction to obtain the corrected parameter gradient.

6. The hearing assistance device of claim 1, wherein, The preset termination condition is any of the following conditions: The number of iteration rounds reaches a preset maximum iteration round threshold; or, The final optimized loss has a fluctuation range within a preset number of consecutive rounds that is less than a preset fluctuation range.

7. The hearing assistance device of claim 1, wherein, The fitting accuracy constraint is that the difference between the gain amplitude at each frequency point and the clinical target gain amplitude does not exceed a preset accuracy threshold. The curve smoothing constraint is that the fluctuation difference of the gain amplitude between adjacent frequency points does not exceed a preset smoothing threshold. The clinical hearing safety constraint is that the gain amplitude at each frequency point does not exceed the preset hearing tolerance threshold corresponding to that frequency point.

8. A method of adaptive hearing compensation, characterized in that, Applied to the hearing aid device as described in any one of claims 1-7, the method comprises: Obtain user hearing loss data; Hearing features are extracted based on the hearing loss data; The hearing loss data and the hearing features are preprocessed to construct the input dataset; The input dataset is input into a pre-trained hearing compensation model. The pre-trained hearing compensation model infers and outputs a personalized frequency response gain curve that fits the user. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band. The gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies the fitting accuracy constraint, curve smoothness constraint, and clinical hearing safety constraint. The original digital audio signal is output after full-band frequency-point gain compensation based on the personalized frequency response gain curve.

9. An adaptive hearing compensation device, characterized in that, The device includes: The acquisition unit is used to acquire the user's hearing loss data; The processing unit is used to extract hearing features based on the hearing loss data; preprocess the hearing loss data and the hearing features to construct an input dataset; input the input dataset into a pre-trained hearing compensation model; and output a personalized frequency response gain curve adapted to the user after the pre-trained hearing compensation model infers and outputs the personalized frequency response gain curve. The personalized frequency response gain curve is a continuous gain curve across the entire frequency band, and the gain amplitude corresponding to each frequency point is positively correlated with the user's hearing loss at that frequency point, and satisfies the fitting accuracy constraint, curve smoothness constraint, and clinical hearing safety constraint. The output unit is used to perform full-band frequency-point gain compensation on the original digital audio signal according to the personalized frequency response gain curve and then output it. The training process of the hearing compensation model includes: The model parameters and corresponding adaptive gradient optimizer of the hearing compensation model are initialized, and the training control parameters are configured. Batch sample data is extracted and input into the model to generate an initial frequency response gain curve. The sample data includes hearing loss data, hearing characteristics, full-band clinical target gain curve, and clinical hearing constraint parameters. The gain amplitude at each frequency point in the full-band clinical target gain curve is positively correlated with the hearing loss at the corresponding frequency point. The hearing loss data includes average hearing loss, high-frequency hearing loss index, and hearing loss slope. We construct positive correlation fitting constraint loss, curve smoothness loss, and clinical safety loss respectively. We then aggregate each loss according to preset weights and combine them with regularization constraints to obtain the final optimized loss. The parameter gradient of the model parameters is calculated based on the final optimization loss. The parameter gradient is used to characterize the adjustment trend and relative adjustment magnitude of the model parameters. A dynamic step-size scheduling strategy combined with an adaptive gradient optimizer is used to complete the iterative update of the model parameters; When the preset termination condition is reached, the optimal model parameters after training are obtained.

Images