A method for self-fitting a hearing aid based on neural network deep learning

Abstract

The application discloses a hearing aid self-fitting method based on neural network deep learning, adopts pure tone and combines audiometry audio with user daily life background noise information to perform audiometry, distributes weights to obtain preset WDRC parameters, adopts KNN algorithm based on the preset WDRC parameters to perform classification to obtain characteristic labels, simultaneously performs evaluation and scoring through audition of specific speech groups, combines the characteristic labels and the evaluation and scoring to adopt a neural network deep learning model to obtain a gain and compression ratio correction vector, and finally obtains the final WDRC parameters. The application can fuse the audiometry link to obtain relatively accurate initial setting parameters containing user characteristic information, a deep learning neural network input vector containing characteristic labels, so that the model has higher accuracy.

Description

Technical Field

[0001] This invention relates to fitting technology for digital hearing aids, and in particular to a self-fitting method for hearing aids based on neural network deep learning. Background Technology

[0002] Traditional hearing aid fitting requires a cumbersome process involving professional fitting experts. With the advancement of wearable smart devices, developing user-initiated fitting methods that eliminate the need for experts and specialized equipment is of great research value and promise. When fitting hearing aids, hearing-impaired patients first need to undergo a hearing test, primarily to assess the degree of hearing loss—specifically, the decibel level at which they can hear the lowest possible sound at different frequencies. Different types of hearing loss vary in degree, nature, and location. Establishing patient data files and fully utilizing user information can significantly improve the efficiency of the self-fitting process.

[0003] Researchers have proposed a user-tuned self-fitting method that uses an initial set of parameters derived from an audiogram, allowing users to independently fine-tune certain parameters. However, users face difficulties in understanding the parameters and the adjustment methods, and the coupling between parameters complicates the adjustment process. This method demands a high level of user expertise, making it difficult to implement on a large scale.

[0004] In recent years, neural network deep learning has achieved significant application results in fields such as speech recognition. One approach employs a self-matching method based on sound scene discrimination, where patient users rate the test speech generated from initial optimal parameters, and this is then adjusted in conjunction with a deep learning algorithm. However, the initial parameters matched based on the similarity of patient user data are random and lack objectivity and reliability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a hearing aid self-fitting method based on neural network deep learning.

[0006] To achieve the above objectives, the present invention adopts the following solution:

[0007] A hearing aid self-fitting method based on neural network deep learning is proposed. The method uses pure tone audio combined with audio recordings containing background noise information from the user's daily life for audio testing. Weights are assigned to obtain preset WDRC parameters. Based on the preset WDRC parameters, the KNN algorithm is used to classify and obtain feature labels. At the same time, specific speech groups are listened to for evaluation and scoring. Combining feature labels and evaluation scores, a neural network deep learning model is used to obtain gain and compression ratio correction vectors, thereby obtaining the final WDRC parameters.

[0008] Step 1: A method for self-fitting hearing aids based on neural network deep learning. Pure-tone audiograms and an initial WDRC1 parameter set are obtained through pure-tone testing. Further, mixed audiograms and mixed WDRCn parameter sets are obtained through listening tests under multiple background noises. By setting the initial WDRC1 parameter weight coefficient α1 and the mixed WDRCn parameter weight coefficient set αn, a preset WDRC parameter set is obtained through linear superposition. Where WDRCI is the set of WDRC parameters under the i-th noise background, and

[0009] Specifically, by using several fixed-frequency test tones, the lowest identification threshold, comfort range, and highest tolerance threshold at different frequencies are obtained based on the feedback from hearing-impaired patients to pure-tone audiometry. This forms the lowest identification threshold audiogram, the highest tolerance threshold audiogram, and the comfort range audiogram, and the initial WDRC1 parameter group is set.

[0010] Specifically, based on the background noise of multiple common scenarios recorded by hearing-impaired patients, frequency domain analysis is used to obtain the spectrum diagram and generate colored background noise corresponding to different fields. These are then superimposed onto pure tones for audiometry to obtain audiogram sequences and WDRC parameter sets corresponding to different background noise conditions.

[0011] Step 2: A method for self-fitting hearing aids based on neural network deep learning, characterized by the following steps:

[0012] A database of WDRC parameter groups corresponding to hearing loss levels, conductive hearing loss, sensorineural hearing loss, and mixed hearing loss was established. The KNN algorithm was used to classify the patients according to the preset WDRC parameter groups. The label vector P = [p1, p2, ..., pn] was set, where n is the number of hearing loss features. The classification was based on the historical patient WDRC database.

[0013] Hearing-impaired patients wear hearing aids with preset WDRC parameter groups, conduct specific audiometry tests on speech groups, and evaluate the speech in six dimensions, namely: volume w1, speech clarity w2, speech dullness w3, speech sharpness w4, amount of background noise w5, and semantic comprehension difficulty w6. The evaluation level of each dimension is scored as good, medium, and poor, with a score range of 0 to 5. An evaluation vector W = [w1, w2, ... w6] is generated.

[0014] For the gain vector B = [b1,b2,...,km] and compression ratio vector K = [k1,k2,...,km] in the WDRC parameters, where m is the number of WDRC channels, the compression ratio and gain in the final WDRC parameters are compared with those in the initial WDRC parameters to obtain the gain correction vector ΔB = [Δb1,Δb2,...,Δkm] and the compression ratio correction vector ΔK = [Δk1,Δk2,...,Δkm].

[0015] A gain correction deep learning network and a compression ratio correction deep learning network are constructed. The augmented vector obtained is based on the label vector P and the evaluation vector W, and the gain correction vector ΔB and the compression ratio correction vector ΔK are used as the outputs. The network is trained to obtain the gain correction network model and the compression ratio correction network model, respectively.

[0016] The gain correction network model is obtained by using the trained gain correction network model. The user's augmented vector Z is used as input to obtain the user's gain correction vector ΔB and compression ratio correction vector ΔK. The final WDRC parameter set of the user is obtained by superimposing the correction values ​​onto the preset WDRC parameter set. Attached Figure Description

[0017] Figure 1 This is a flowchart of a self-fitting method according to the present invention. Detailed Implementation

[0018] The present invention will now be further described in conjunction with the accompanying drawings.

[0019] Specifically, Figure 1 The flowchart of a self-fitting method of the present invention is as follows: listening tests are conducted by using pure tones combined with audio samples containing background noise information from the user's daily life. Weights are assigned to obtain initial WDRC parameters. Based on the initial WDRC parameters, the KNN algorithm is used to classify and obtain feature labels. At the same time, evaluation and scoring are performed by listening to specific speech groups. Combining feature labels and evaluation scores, a neural network deep learning model is used to obtain gain and compression ratio correction vectors, thereby obtaining the final WDRC parameters.

[0020] Pure-tone audiometry: By using several fixed-frequency test tones, the lowest recognition threshold, comfort range, and highest tolerance threshold at different frequencies are obtained based on the feedback from hearing-impaired patients. This forms the lowest recognition threshold audiogram, the highest tolerance threshold audiogram, and the comfort range audiogram, and the initial WDRC1 parameter set is set.

[0021] Audiometry combining background noise information from the user's daily life: Based on the background noise of multiple common scenarios recorded by hearing-impaired patients, frequency domain analysis is used to obtain the spectrum diagram and generate colored background noise corresponding to different fields. These are then superimposed onto pure tones for audiometry to obtain audiogram sequences and WDRC parameter groups corresponding to different background noise conditions.

[0022] Specifically, the method for generating colored background noise in different scenarios is as follows: Based on the time-domain signal of noise in common scenarios, a fast algorithm of Discrete Fourier Transform (FFT) is used to transform the time-domain signal of the speech to the frequency domain, and the spectral function structure G(f) is obtained through discretization. n ), f n Let Δf be the frequency band width, n be the index of the frequency band in ascending order, and the vertical axis be the amplitude of the signal in that frequency band.

[0023] Furthermore, using white noise as the excitation source, colored noise is obtained by filtering it through a filter with the aforementioned preset spectral shape.

[0024] Furthermore, by adjusting the overall amplitude gain k of the colored noise, the overall energy of the colored noise is made to reach a preset energy E, thus forming noise with a more suitable energy. Specifically... Where p is the maximum frequency band index of the spectral structure function after spectral discretization.

[0025] Initial WERC parameter set calculation: Pure tone audiograms and initial WDRC1 parameter sets are obtained through pure tone testing. Further, mixed audiograms and mixed WDRCn parameter sets are obtained through listening tests with multiple background noises. By setting the initial WDRC1 parameter weighting coefficient α1 and the mixed WDRCn parameter weighting coefficient set αn, the preset WDRC parameter set is obtained through linear superposition. Where WDRCI is the set of WDRC parameters under the i-th noise background, and

[0026] Category Tags: A database of WDRC parameter groups was established, with elements such as hearing loss level, conductive hearing loss, sensorineural hearing loss, and mixed hearing loss. Among them, the hearing loss level information is divided into 5 categories, represented by numbers 1-5 in the database; conductive hearing loss is used as judgment information, represented by 0 or 1 to indicate yes or no; sensorineural hearing loss is used as judgment information, represented by 0 or 1 to indicate yes or no; and mixed hearing loss is used as judgment information, represented by 0 or 1 to indicate yes or no.

[0027] Based on the preset WDRC parameter set, the KNN algorithm is used to classify and predict the type and degree of hearing loss. At the same time, a hearing loss label vector P = [p1, p2, ..., pn] is set, where n is the number of hearing loss features.

[0028] Deep learning neural network training includes the following steps:

[0029] Hearing-impaired patients wear hearing aids with preset WDRC parameter groups, conduct specific audiometry tests on speech groups, and evaluate the speech in six dimensions, namely: volume w1, speech clarity w2, speech dullness w3, speech sharpness w4, amount of background noise w5, and semantic comprehension difficulty w6. The evaluation level of each dimension is scored as good, medium, and poor, with a score range of 0 to 5. An evaluation vector W = [w1, w2, ... w6] is generated.

[0030] For the gain vector B = [b1,b2,...,km] and compression ratio vector K = [k1,k2,...,km] in the WDRC parameters, where m is the number of WDRC channels, the compression ratio and gain in the final WDRC parameters are compared with those in the initial WDRC parameters to obtain the gain correction vector ΔB = [Δb1,Δb2,...,Δkm] and the compression ratio correction vector ΔK = [Δk1,Δk2,...,Δkm].

[0031] A gain correction deep learning network and a compression ratio correction deep learning network are constructed. The augmented vector obtained is based on the label vector P and the evaluation vector W, and the gain correction vector ΔB and the compression ratio correction vector ΔK are used as the outputs. The network is trained to obtain the gain correction network model and the compression ratio correction network model, respectively.

[0032] User parameter adjustment: The gain correction network model is obtained by using the trained gain correction network model. The user's augmentation vector Z is used as input to obtain the user's gain correction vector ΔB and compression ratio correction vector ΔK. The final WDRC parameter set of the user is obtained by superimposing the correction values ​​onto the preset WDRC parameter set.

[0033] Any content not described in detail in this specification is prior art known to those skilled in the art.

[0034] Although the present invention has been described above with reference to the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Any modifications and refinements made by those skilled in the art without departing from the concept and scope of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A method of neural network deep learning based hearing aid self-fitting, characterized in that: The pure tone audiogram and the initial WDRC1 parameter set are obtained through the pure tone test, the mixed audiogram set and the mixed WDRCn parameter set are obtained through the hearing test under the superposition of various background sounds, the initial WDRC1 parameter weight coefficient α1 and the mixed WDRCn parameter weight set αn are set, the preset WDRC parameter set is obtained through linear superposition, and the preset WDRC parameter set is used for Wherein, WDRCi is the WDRC parameter set under the i-th noise background, and ; According to the feedback of the pure tone audiometry of the hearing loss patient, the minimum recognition threshold, the comfortable range and the maximum tolerance threshold at different frequencies are obtained to form the minimum recognition threshold audiogram, the maximum tolerance threshold audiogram and the comfortable range audiogram, and the initial WDRC1 parameter set is set; According to the background noise of multiple common scenes recorded by the hearing loss patient, the frequency spectrum is obtained through frequency domain analysis to generate different field corresponding colored background noise, which is respectively superimposed into the pure tone for audiometry to obtain the corresponding hearing level sequence and the WDRC parameter set under different background noise; The self-fitting method comprises the following steps, S1: Establish a WDRC parameter set database corresponding to the elements of hearing loss level, conductive hearing loss, sensorineural hearing loss and mixed hearing loss, classify according to the preset WDRC parameter set using KNN algorithm, set the label vector P=[p1, p2,..., pn], wherein n is the number of hearing loss characteristics; S2: The hearing loss patient wears a hearing aid with a preset WDRC parameter set, performs a specific test on the test speech group, and evaluates the six dimensions of the speech, wherein the evaluation dimensions are: volume w1, speech intelligibility w2, speech dullness w3, speech sharpness w4, noise w5, and speech meaning understanding difficulty w6; each dimension is scored according to good, medium and poor, and the score range is 0-5; an evaluation vector W=[w1, w2,...w6] is generated; S3: For the gain vector B=[b1, b2,..., bm] and the compression ratio vector K=[k1, k2,..., km] in the WDRC parameter, wherein m is the number of WDRC channels; compare the compression ratio and gain in the final WDRC parameter with the initial WDRC parameter to obtain the gain correction vector ΔB=[Δb1, Δb2,..., Δbm] and the compression ratio correction vector ΔK=[Δk1, Δk2,..., Δkm]; S4: Constructing the gain correction deep learning network and the compression ratio correction deep learning network based on the label vector P and the evaluation vector W to obtain an augmented vector As input, respectively taking the gain correction vector ΔB and the compression ratio correction vector ΔK as output, training respectively obtains the gain correction network model and the compression ratio correction network model; S5: The gain correction network model is obtained by using the trained gain correction network model, the user's augmented vector Zuser is input, the user's gain correction vector ΔBuser and compression ratio correction vector ΔKuser are obtained, and the final WDRC parameter set of the user is obtained by adding the correction value to the preset WDRC parameter set.

Images