Determining sound quality of earphones using machine learning algorithms

Abstract

A method for determining a sound quality (26) of an earphone (14) to be inserted into an ear of a user, comprising: receiving user data (22) comprising at least an audiogram of the user; and inputting the user data (22) into a machine learning algorithm (24), and determining the sound quality (26) by means of the machine learning algorithm (24).

Description

Technical Field

[0001] The present invention relates to a method, a computer program, and a computer-readable medium for determining the sound quality of earphones to be inserted into a user's ear, and a method, a computer program, and a computer-readable medium for training a machine learning algorithm for determining sound quality. Background Art

[0002] Hearing devices are generally small and complex devices. A hearing device may include a processor, a microphone, a speaker, a memory, a housing, and other electronic and mechanical components. Some hearing devices have a so-called dome that is inserted into the user's ear. The dome can be open, vented (i.e., can have one or more vent channels), or fully enclosed. A fully enclosed dome may also be referred to as a power dome. The dome is used to input sound from the hearing device into the ear canal and to filter ambient sound. Generally, the filtering of ambient sound is achieved by means of the external-to-internal acoustic coupling of the ear canal performed by the dome.

[0003] Optimal acoustic coupling is generally determined based on the user's hearing loss, experience, desired amplification, self-speech perception, etc. Acoustic coupling can be formally given by the sound quality or cut-off frequency provided by the dome. Too open or too blocked coupling configurations may lead to user rejection, for example due to feedback, occlusion effects, performance degradation of coupling-related features (such as beamforming), and a higher risk of direct sound attenuation.

[0004] Generally, the proposal of the optimal dome is based on rule-based algorithms, which find a balance between rejection metrics for each dome in dealing with occlusion effects, feedback risks, etc. Due to the high complexity of such algorithms, it is challenging to find a suitable parameterization for the optimal proposal acceptable to users and hearing care professionals. In addition, current algorithms generally only consider a subset of relevant end-consumer characteristics, such as mono sound, which may lead to a higher degree of non-acceptance in cases of highly asymmetric hearing loss.

[0005] In WO2021138603 A1, patient information is obtained, including the patient's auditory diagnostic data and patient-specific data. The auditory diagnostic data and patient-specific data are concatenated into an input vector. Using the input vector and multiple feature vectors as inputs to a machine learning training model, the correlation between the input vector and each of the multiple feature vectors is determined, and the multiple feature vectors correspond to multiple hearing device models. Based on the corresponding correlations with the input vector, the multiple hearing device models are ranked, and information corresponding to the highest-ranked hearing device model is output. Summary of the Invention

[0006] The object of the present invention is to simplify the selection and / or manufacture of earphones for a specific user. Another object is to improve the accuracy of determining the suitable sound quality of an earphone for a specific user.

[0007] These objects are achieved by the subject matter of the independent claims. Further exemplary embodiments are apparent from the dependent claims and the following description.

[0008] A first aspect of the present invention relates to a method for determining the sound quality of an earphone to be inserted into a user's ear and / or ear canal. Generally, the sound quality can be defined by the coupling of the earphone between the outside and inside of the ear relative to the transfer function of the earphone. The transfer function can define the phase shift and / or attenuation of a sound signal at a set of frequencies.

[0009] The earphone can be a dome of a hearing device or a hearing aid. The dome can be connected to the hearing device or hearing aid via a hose. The earphone can also be a housing of a hearing device or a part of the housing. The hearing device is adapted to receive a sound signal, process the sound signal and output the sound to the ear of the user (i.e., the user wearing the hearing device). A hearing aid is a hearing device adapted to process a sound signal such that the hearing loss of the user is compensated.

[0010] According to one embodiment, the method includes: receiving user data including at least an audiogram of the user; and inputting the user data into a machine learning algorithm and determining the sound quality by the machine learning algorithm.

[0011] The audiogram can define the auditory ability of the user at a plurality of frequencies. For example, the audiogram includes the hearing level of the user at a plurality of frequencies. Generally, the audiogram is an indicator of the hearing defect and hearing ability of the user intended to wear the earphone. The audiogram and optionally further data are input into a machine learning algorithm that has been trained to output the sound quality that is optimally provided by the earphone. As described in more detail below, the machine learning algorithm has been trained using the audiogram, optionally further user data, and the sound quality of earphones that have been successfully used by the corresponding users.

[0012] By using such a machine learning algorithm, complex rule-based algorithms can be replaced. When new user data is available, the machine learning algorithm can be easily improved. In addition, the output of the machine learning algorithm (i.e., the sound quality) can facilitate the fitting of hearing devices, the selection of suitable domes, the manufacture of customized earphones, the manufacture of individual hearing protectors, etc. For hearing protection, it is also possible for the machine learning algorithm to output an optimal attenuation curve. The optimal attenuation curve can be compared with the curve of a filter in a hearing protection combination, and the most suitable filter can be selected, i.e., the filter having the curve that differs least from the optimal attenuation curve output by the machine learning algorithm.

[0013] According to one embodiment, an audiogram includes at least one of the following: an air conduction audiogram; a bone conduction audiogram; an ipsilateral audiogram of the ear into which the earphone is inserted; a contralateral audiogram of the opposite ear; an uncomfortable loudness level (e.g., at multiple frequencies). The audiograms for air conduction, bone conduction, and / or uncomfortable loudness level may have different levels at different frequencies (typically between 125 Hz and 10,000 Hz). Generally, an audiogram includes measurements of the user's hearing ability at different frequencies. Such measurements may include the user's lowest hearing level, the user's speech hearing level, the uncomfortable loudness level, etc.

[0014] These measurements may have been recorded for air conduction (i.e., sound received by the user via air) and / or bone conduction (i.e., sound received via the skull). Distinguishing between air conduction and bone conduction can help in selecting the sound quality of the earphone.

[0015] The audiogram may include data for the ipsilateral ear (i.e., the ear into which the earphone is inserted), and optionally data for the opposite ear. The data for the opposite ear can indicate how much of a hearing deficit difference there is between the ears, which can also help in selecting the sound quality of the earphone.

[0016] According to one embodiment, the earphone is an earphone of a hearing aid. This can be a dome to be connected to a behind-the-ear hearing aid or an in-the-ear receiver hearing aid. The earphone can also be the housing of an in-the-ear hearing aid.

[0017] According to one embodiment, the user data additionally includes the user's experience level indicating the user's experience with the hearing aid. The experience level can include the number of years the user has worn hearing devices. The experience level can distinguish between novice users and long-term users. For new users, acoustic coupling related to a more natural sound setting may be beneficial, i.e., the sound quality can be determined to be lower.

[0018] According to one embodiment, the user data additionally includes a fitting formula for the hearing aid. Such a fitting formula can be a formula that converts an audiogram into fitting parameters for the hearing device. Such parameters can include frequency-specific and / or input level-specific amplification / gain settings. There are standardized and / or widely used fitting formulas, such as NAL-NL2 (National Acoustic Laboratories Nonlinear 2), DSL (Desired Sensation Level), Phonak Digital (Adaptive Phonak Digital, APD), etc. Utilizing knowledge of the fitting formula can improve the accuracy of determining the sound quality.

[0019] According to one embodiment, the acoustic mass is the acoustic vent mass, which is proportional to the length of the vent and inversely proportional to the cross-sectional area of the vent. As already described, generally, the acoustic mass may include an indicator for the acoustic coupling behavior of the earphone. The acoustic mass may be encoded as the geometry of the vent (i.e., the passage between the outside and the inside of the earphone).

[0020] According to one embodiment, the machine learning algorithm is an artificial neural network, a Gaussian process, polynomial regression, and / or a regression tree.

[0021] In particular, artificial neural networks, especially those with multiple layers, are well-suited for modeling complex functions (e.g., acoustic mass depending on the audiogram and optionally further user data).

[0022] The artificial neural network can be a multi-layer perceptron (MLP). The MLP is suitable for regression prediction problems, where a real-valued quantity is predicted given a set of inputs. The MLP is suitable for classification prediction problems, where the inputs are assigned classes or labels. In addition, the MLP classifier can handle various types of data. They can also accommodate multiple output classes, making them suitable for multi-class classification problems. The MLP is a multi-layer feedforward neural network with an input layer, one or more hidden layers, and an output layer. Each layer except the input layer processes the information from the previous layer and sends the result to the next layer. Generally, a three-layer neural network can approximate a non-linear mapping with arbitrary precision.

[0023] The MLP can be implemented as a regression prediction problem. The MLP can predict continuous numerical values. The classification of the dome type can be done based on the continuous numerical values, for example, by selecting the dome based on comparing the continuous numerical value thresholds. The predicted continuous numerical values can be the acoustic mass of various dome types.

[0024] According to one embodiment, the earphone is the dome of a hearing aid. In particular, when adapting to a new hearing aid, the selection of the dome is a difficult task because the dome significantly affects the user's sound perception. An unfavorably selected dome may significantly reduce the user's acceptance and satisfaction.

[0025] According to one embodiment, the earphone is a hearing protector. The determination of the acoustic mass can also be performed to select other types of earphones, such as hearing protectors. A hearing protector can be a device that passively or actively reduces the level of the sound signal to be provided to the user's ear to protect the user's ear. In this case, the machine learning algorithm can also output an optimal attenuation curve, which is compared with the attenuation curves of different types of hearing protectors for selecting the most suitable hearing protector. For the most suitable hearing protector, the attenuation curve may differ the least from the optimal attenuation curve output by the machine learning algorithm.

[0026] According to one embodiment, the method includes: selecting a type of earphone that provides the determined sound quality. There may be several types of prefabricated earphones. For example, a set of open domes, vented domes, and closed domes. From this set, an earphone can be selected in which the difference in sound quality between the earphone and the sound quality determined by this method is minimized. In particular, the difference in logarithmically compressed sound quality can be used. The selected type can be output, for example, on an adaptation device by this method.

[0027] According to one embodiment, the method includes: determining the geometric dimensions (such as, for example, diameter and length) of the vent of the earphone. This may already have been done on the manufacturer side of the earphone. These geometric dimensions can be used to manufacture the earphone.

[0028] According to one embodiment, the method includes: generating manufacturing data for the earphone. The sound quality can be converted into geometric features of the earphone, such as the geometric shape of the vent of the earphone. Based on the manufacturing data, the earphone can be manufactured.

[0029] Another aspect of the present invention relates to a method for selecting a dome of a hearing aid to be inserted into a user's ear.

[0030] According to one embodiment, the method includes: receiving user data including at least the audiogram of the user; inputting the user data into a machine learning algorithm, and using the machine learning algorithm to determine the dome type for the dome. For example, the machine learning algorithm can determine the sound quality of the dome, and thus the dome type can be determined. The dome type can be selected as the dome type that provides the sound quality closest to the sound quality determined by the machine learning algorithm. There may be a predefined vent quality associated with each dome type. The machine learning algorithm can be used to determine the optimal acoustic vent quality, and based on the predefined acoustic vent quality of the dome or dome type, the best-matching predefined acoustic vent quality can be selected.

[0031] Another aspect of the present invention relates to a training method for training a machine learning algorithm for determining the sound quality of an earphone to be inserted into a user's ear. For an artificial neural network, this training method includes determining the weights of the neurons of the artificial neural network.

[0032] According to one embodiment, the training method includes: receiving a data set of records with user data, each record including at least the audiogram of the user and the type of earphone used by the user; for each record, determining at least one user score for the earphone type according to the user data; generating a filtered data set by excluding records from the data set, where if at least one user score is below a threshold, the record is excluded from the data set; determining the sound quality of each record according to the earphone type; and using the filtered data set to train the machine learning algorithm.

[0033] Generally, the input data for training a machine learning algorithm includes: user data to be input into the sound quality determination method, optional further user data, and the sound quality corresponding to the user data.

[0034] The sound quality is determined based on the headphone type provided in the user data. For example, the corresponding sound quality can be determined according to the shape of the vent or the measurement results for each headphone type.

[0035] In this training method, a user score is determined based on the user data, and the user score indicates the degree of selection of the sound quality for the scenario encoded in the user data. The user score can indicate the degree of satisfaction of the user with the selection of the sound quality and / or the headphones, and / or the improvement in the compensation of the user's hearing defect by the sound quality.

[0036] The user score is determined based on the user data. As an example, the user data can include a questionnaire filled out by the user. The results of the questionnaire can be evaluated to determine the user score.

[0037] However, indirect indicators can also be used to determine the user score, such as the wearing time of the hearing device or the active use of the sound program by the user.

[0038] According to one embodiment, the user score is or includes a customer satisfaction score. In this case, the user's satisfaction with the headphones is used as an indicator of whether the correct sound quality has been used.

[0039] According to one embodiment, each record of the data set additionally includes the wearing time of the headphones, and the user score is a customer satisfaction score determined based on the wearing time. The indicator of user satisfaction is the wearing time. The longer the wearing time, the higher the user's satisfaction.

[0040] According to one embodiment, the user score is or includes an intelligibility score. In this case, the improvement in the intelligibility of the speech or other sounds of the hearing device combined with the headphones is used as an indicator of whether the correct sound quality has been used.

[0041] Such an intelligibility score can be determined based on a specific type of user data without asking the user questions. Generally, the intelligibility of speech and / or sounds can be deduced based on the adaptation of the hearing device in combination with the sound quality of the headphones.

[0042] According to one embodiment, each record of the data set further includes a fitting formula for a hearing aid to be used with the earphone; wherein, for at least one frequency, a desired target gain is determined according to the fitting formula; wherein, for at least one frequency, a target gain limited by a feedback threshold is determined according to the earphone type; and wherein, the intelligibility score is determined according to the target gain limited by the feedback threshold at at least one frequency.

[0043] The target gain of the hearing device at one or more frequencies can be determined according to the audiogram and fitting formula stored in the user data. On the other hand, the target gain limited by the feedback threshold can be determined according to the data of the earphone type in the user data, that is, the maximum possible target gain of a specific sound quality at these one or more frequencies. Using the target gain limited by the feedback threshold at one or more frequencies, an indicator for intelligibility can be derived for the combination of the adapted hearing device and the earphone. Such an indicator can be used as an intelligibility score.

[0044] According to one embodiment, the training method further includes: determining an enhancement score for each record according to at least one user score; when the enhancement score of the record is higher than a threshold, generating an enhanced data set by at least copying the record. On the one hand, one or more user scores can be used to filter out records with low user scores, that is, wherein the sound quality is adapted to the scenario in the user data. On the other hand, one or more user scores can be used to multiply records with high user scores. For this purpose, when one or more user scores are high, a high enhancement score can be determined, and when one or more user scores are low, a low enhancement score can be determined.

[0045] It is possible that the sound quality of the user data and / or the copied records varies, thus generating a higher variance in the data set.

[0046] Other aspects of the present invention relate to a computer program for determining the sound quality of an earphone to be inserted into a user's ear and / or for training a machine learning algorithm using a machine learning algorithm, and to a computer-readable medium in which such a computer program is stored, the computer program being adapted to perform the steps of the methods described above and below when executed by a processor.

[0047] For example, the computer program for determining the sound quality can be executed in an adaptation device. The computer program for determining the sound quality and / or for training a machine learning algorithm can be executed in a computing device for manufacturing hearing devices. The computer-readable medium can be the memory of the corresponding device.

[0048] Generally, a computer-readable medium can be a hard disk, a USB (Universal Serial Bus) storage device, RAM (Random Access Memory), ROM (Read-Only Memory), EPROM (Erasable Programmable Read-Only Memory), or FLASH memory. A computer-readable medium can also be a data communication network that permits downloading of program code, such as the Internet. A computer-readable medium can be a non-transitory or transitory medium.

[0049] It should be understood that the features of the methods as described above and below can be the features of the computer programs and computer-readable media as described above and below, and vice versa.

[0050] With reference to the embodiments described below, these and other aspects of the present invention will be clear and elucidated. Brief Description of the Drawings

[0051] Hereinafter, embodiments of the present invention will be described in more detail with reference to the drawings.

[0052] Figure 1 A hearing device with earphones is shown.

[0053] Figure 2 Different types of domes are shown.

[0054] Figure 3 A method for determining sound quality according to an embodiment of the present invention is shown.

[0055] Figure 4 A method for training a machine learning algorithm according to an embodiment of the present invention is shown.

[0056] Figure 5 A graph with sound quality output by a machine learning algorithm is shown.

[0057] Figure 6 A graph with a median wearing time for determining a customer satisfaction score is shown.

[0058] Figure 7 A graph with a weighting function for determining a customer satisfaction score is shown.

[0059] Figure 8 A graph with a weighting function for determining an intelligibility score is shown.

[0060] The reference numerals used in the drawings and their meanings are listed in the list of reference numerals in an overview form. In principle, the same components are provided with the same reference numerals in the drawings. Detailed Description of the Embodiments

[0061] Figure 1There is shown a hearing device 10, in particular a hearing aid, which comprises a component 12 to be worn behind the ear and an earpiece 14 to be inserted into the ear canal of a user, who is also referred to herein as the user. The component 12 comprises a microphone, a sound processor and other components for processing the sound signal from the microphone and compensating for the user's hearing deficiency. The component 12 is connected to the earpiece 14 via a cable 16, and the earpiece 14 comprises a loudspeaker and a dome 18 to be inserted into the ear canal. More generally, the earpiece 14 can be the housing of the hearing device or a hearing protector.

[0062] Figure 2 There are shown an open dome 18a, a vented dome 18b and a closed dome 18c. The dome 18a has a rather large opening, which allows sound to pass through the dome 18a with little change. Except for the vent 20, the vented dome 18b blocks the ear canal, and the vent 20 is a small passage interconnecting the interior and the exterior with respect to the dome 18b. The closed and / or powered dome 18c does not have any opening connecting the interior to the exterior and completely blocks the ear canal.

[0063] The more open dome 18a has the following advantages: a smaller occlusion effect, better wearing comfort and a more natural user voice. On the other hand, the more closed domes 18b, 18c have the following advantages: better intelligibility, better streaming quality, a lower risk of feedback and better sound cleaning. The selection and / or design of the dome 18 or the more general earpiece 14 is a challenging task, which is significantly simplified by the method proposed herein.

[0064] Figure 3 There is shown a diagram depicting a method for determining the sound quality 26 of an earpiece 14 to be inserted into a user's ear and / or the user.

[0065] The method starts with receiving user data 22. The user data 22 includes at least the audiogram of the user. Such an audiogram can be recorded at the office of a hearing care professional. Generally, the audiogram can include the levels of hearing loss at multiple frequencies. It is also possible that the audiogram includes the levels of uncomfortable loudness at multiple frequencies.

[0066] More specifically, the audiogram includes at least one of the following: an air conduction audiogram, a bone conduction audiogram, an ipsilateral audiogram of the ear into which the earpiece is inserted, and / or a contralateral audiogram of the opposite ear. As described above, all this data can improve the accuracy of the method.

[0067] In addition to the audiogram, the user data 22 can additionally include other data related to the user, such as the user's experience level indicating the user's experience with the hearing aid and / or the fitting formula for the hearing aid 10.

[0068] The sound quality 26 is determined by a machine learning algorithm 24, which is an artificial neural network in the form of a multi-layer perceptron (MLP). The machine learning algorithm 24 has been trained as described herein. User data 22 is input into the machine learning algorithm 24, and then the machine learning algorithm 24 determines the sound quality 26.

[0069] The sound quality 26 can be a value, but can also be a set of values that define a curve, such as a transfer function of the coupling of the earphone 14 between the outside and inside of the ear relative to the earphone. For example, the set of values can include the phase shift and / or attenuation of the sound signal at a set of frequencies. An example of the sound quality 26 as a single value is the acoustic vent quality, which is proportional to the length of the vent 20 and inversely proportional to the cross-sectional area of the vent 20.

[0070] Once the sound quality 26 has been determined, it can be used to select the type 28 of the earphone 14 that provides the determined sound quality 26.

[0071] There can be a list 28 of earphone types stored in the device that executes the method. Each earphone type 28 can be stored together with its sound quality 26, which can be compared with the sound quality 26 determined by the machine learning algorithm 24. The method can output the best earphone type and / or the best dome type that should be used by the user.

[0072] Alternatively or additionally, the geometric dimensions 30 of the vent 20 of the earphone 14 are determined for the sound quality 26, and the geometric dimensions 30 can be used to manufacture the earphone 14.

[0073] Specifically, the dimensions of the vent 20 (such as diameter and length) can be related to a scalar sound quality value (such as the acoustic vent quality). This allows the derivation of the best vent dimensions given the earphone 14 and the above input parameters.

[0074] Alternatively or additionally, manufacturing data 30 for the earphone 14 can be generated. For example, this can be done when manufacturing a custom housing for a hearing device.

[0075] Figure 4 A diagram depicting the training of the machine learning algorithm 24 is shown. The left box 31 shows data preparation, and the right box 33 shows training using the prepared data.

[0076] The starting point is the original data set 32, which can include tens of thousands of fitted data. In particular, each record of the data set 32 relates to a specific adaptation of the earphone 14 to the user. Each record includes one or more audiograms for the adaptation. Such audiograms can be as described above, i.e., can include an air conduction audiogram, a bone conduction audiogram, an ipsilateral audiogram of the ear into which the earphone is inserted, and / or a contralateral audiogram of the opposite ear.

[0077] Each record also includes data about the earphone 14, such as the earphone type, the geometric dimensions of the vent of the earphone, and / or the sound quality of the earphone 14.

[0078] Each record can also include additional user data generated during and / or related to the adaptation, such as questionnaires, wearing time, the age of the user, the user's experience with hearing devices, etc.

[0079] When performing fitting related to the hearing device 10 or hearing aid, the record can also include hearing device data, such as fitting formulas, fitting parameters, etc.

[0080] The data set 32 is received at the start of data preparation. The record includes at least the user's audiogram and the earphone type 28 used by the user.

[0081] After that, for each record, at least one user score 34, 36 of the earphone type is determined according to the data in the record. As shown in the figure, such user scores can be a customer satisfaction score 34 and / or an intelligibility score 36.

[0082] With the aid of the user scores 34, 36, a filtered data set 38 is generated. The filtered data set 38 is generated by excluding records from the data set 32, where, if at least one user score 34, 36 is lower than a threshold, the record is excluded from the data set 32.

[0083] For example, if either of the user scores 34, 36 is lower than a specific threshold of the user score, the record is excluded from the data set 32. For example, the threshold can be 0.5 for the customer satisfaction score 34 and / or 0.9 for the intelligibility score 36. In this case, the user satisfaction and / or the expected hearing performance are too poor for the machine learning algorithm 24 to train on the corresponding record.

[0084] It is also possible to determine an enhancement score 40 for each record of the filtered data set 38 according to at least one user score 34, 36. The enhancement score 40 is used to generate an enhanced data set 42. In the enhanced data set 42, at least the records with an enhancement score 40 higher than the threshold are replicated.

[0085] In the case of only one user score 34, 36, the enhanced score 40 can be the user score, and the threshold for repetition is higher than the threshold for filtering.

[0086] In the case of at least two user scores 34, 36, the enhanced score 40 can be a weighted average of the at least two scores 34, 36, such as a 50% / 50% average. The threshold for the enhanced score 40 can be determined based on the filtering threshold of the user scores.

[0087] For example, at the enhanced score 40 of 0.7 = 0.5*0.5 + 0.9*0.5, a record can simply be kept. For example, at a value of 1, this record can be used 10 times. The enhanced score 40 is scaled (how many times the record is replicated) to ensure that the machine learning algorithm 24 is tuned for high-quality examples.

[0088] The data of the repeated records can vary. The replicated records do not have to be exactly the same as the original records, which can be achieved, for example, by adding a small amount of noise to the audiometry data.

[0089] For training, for each record, the user data 22 for training is extracted from each record of the filtered data set 38 or the optional enhanced data set 42. Additionally, based on data related to the earphone associated with the record (such as the earphone type 28), the sound quality 26 for each record is determined. There can be a list of earphone types 28, where the specific sound quality 26 has been determined for each earphone type 28 and has been stored in the list.

[0090] Then, by inputting the user data 22 and the corresponding sound quality 26 for each record into the training algorithm, the machine learning algorithm 24 is trained using the filtered data set 38 and / or the enhanced data set 42.

[0091] Figure 5 A graph showing the possible output (i.e., the sound quality 26 as a numerical value) of the machine learning algorithm 24 depending on the hearing loss is shown. By setting a threshold, the selection of different dome types can be performed, such as the open dome 18a, the vented dome 18b, and the closed dome 18c. For example, in the case of having a hearing loss along the fourth vertical line, the predicted best sound quality 26 will fall into the category of the closed dome 18c.

[0092] The benefit of predicting the sound quality 26 instead of the dome type lies in the ability to be forward compatible. If, for example, in a future combination, the vented dome 18b is replaced by two new vented domes, one with a sound quality 26 closer to the open dome 18a and the other with a sound quality 26 closer to the closed dome 18c, only the range has to be changed.

[0093] Another novelty of the proposed method compared to previously proposed algorithms is that it is not only based on unilateral hearing loss, but can also consider contralateral hearing loss. Additionally, the user experience is considered to capture the preferences of novice users to more openly conform to the user experience. Two parameters are found to improve the prediction accuracy of the machine learning algorithm 24 during testing.

[0094] Return Figure 4 , the customer satisfaction score 34 can be determined based on the wearing time of the earphone 14. To this end, each record of the data set 32 needs to include data from which the wearing time of the earphone 14 can be determined.

[0095] The customer satisfaction score 34 can be derived based on the average daily wearing time for each record separately. Normalization can be done based on the hearing loss at different frequencies (such as 500, 1000, 2000, 4000 Hz). Such data can be derived based on the user's audiogram.

[0096] Such normalization is as Figure 6 shown. First, the median wearing time 44 is derived.

[0097] This means that users with milder hearing loss wear the earphone 14 for a shorter average period during the day compared to users with higher hearing loss, which is beneficial to consider when deriving the customer satisfaction score 34.

[0098] The median wearing time curve 44 is used to derive the weighting function 46 using the sigmoid function for each record. Figure 7 Shows the weighting function 46 for different hearing losses. The slope of the function is based on the underlying distribution of the data, i.e., the percentile.

[0099] Then, by scaling the daily usage time (as Figure 6 shown), and multiplying by the corresponding weight based on the hearing loss (as Figure 7 shown), the customer satisfaction score 34 for the record is determined.

[0100] Here, the derivation of the customer satisfaction score 34 is described in a rule-based manner. Another approach is to learn the data likelihood given the hearing loss curve and the wearing time. Such an extension would allow capturing more complex dependencies of the wearing time on the hearing loss, which might be lost here by only calculating the average at four frequencies.

[0101] Return again Figure 4, the intelligibility score 36 can be determined based on the parameters of the user's hearing loss and on the fitting formula of the hearing device 10 used with the earphone 14. In this case, each record of the data set 32 additionally includes the fitting formula of the hearing aid 10 to be used with the earphone 14.

[0102] For each record, the intelligibility score 36 is derived through the following steps:

[0103] Step 1: Calculate the expected target gain based on the fitting formula for 50 dB and 65 dB speech. Generally, for at least one frequency, the expected target gain is determined according to the fitting formula.

[0104] Step 2: Calculate the estimated feedback threshold of the selected earphone 14, and derive the feedback threshold-limited target gain achievable using the earphone type provided in the record.

[0105] Step 3: Calculate the expected intelligibility value with the established intelligibility metric (such as the derivation of the speech intelligibility index or the percentage of correct IEEE sentences) based on the feedback threshold-limited target gain. This can also be done for at least one frequency, particularly for two speech input levels of 50 dB and 65 dB.

[0106] Step 4: Derive the intelligibility score 36 based on a normalization method. In particular, the intelligibility value in Step 3 above is normalized using a weight function. Thus, the intelligibility score 36 is determined based on the feedback threshold-limited target gain at at least one frequency.

[0107] Figure 8 The weight function 48 for the clarity score 36 is shown. The weight function 48 is determined based on the intelligibility values of all records determined in Step 3. The normalization is based on bucketizing by the hearing loss at different frequencies (the hearing loss is shown on the right side of the figure).

[0108] The clarity values depending on the hearing loss are Figure 8 shown as a point cloud. The weight function 48 is the median of these intelligibility values.

[0109] The derivation of the intelligibility score 36 is outlined here in a rule-based manner. Another approach is to learn the data likelihood given complete audiometry data. This would allow capturing more complex dependencies on intelligibility that might be lost here by only calculating the average over four AC frequencies.

[0110] Although the present invention has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive; the invention is not limited to the disclosed embodiments. By studying the drawings, the disclosure, and the appended claims, those skilled in the art can understand and implement other variations of the disclosed embodiments and practice the claimed invention. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or controller or other unit may implement the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope. List of Reference Signs 10 Hearing device 12 Behind-the-ear component 14 Earphone 16 Cable 18 Dome 18a Open dome 18b Ventilated dome 18c Closed dome 20 Vent 22 User data 24 Machine learning algorithm 26 Sound quality 28 List of earphone types 30 Geometric dimensions, manufacturing data 32 Set of raw data 34 User scores, customer satisfaction scores 36 User scores, intelligibility scores 38 Filtered data set 40 Enhancement scores 42 Enhanced data set 44 Median wearing time 46 Weighting function 48 Weighting function

Claims

1. A method for determining the acoustic quality (26) of an earphone (14) to be inserted into an ear of a user, the method comprising: receiving user data (22) comprising at least an audiogram of the user; The user data (22) are input into a machine learning algorithm (24) and the sound quality (26) is determined by the machine learning algorithm (24).

2. The method according to claim 1, in, The audiogram includes at least one of the following: air conduction audiogram; Bone conduction audiogram; an audiogram of the same side of the ear into which the earphone is inserted; contralateral audiogram of the opposite ear; Uncomfortable loudness levels at multiple frequencies.

3. The method according to claim 1 or 2, in, The earphone (14) is an earphone of a hearing aid (10); The user data (22) further includes at least one of the following: an experience level of the user indicating the user's experience with the hearing aid; A fitting formula for the hearing aid (10).

4. The method according to one of the preceding claims, in, The acoustic mass (26) is an acoustic vent mass, which is proportional to the length of the vent (20) and inversely proportional to the cross-sectional area of ​​the vent (20).

5. The method according to one of the preceding claims, in, The machine learning algorithm (24) is an artificial neural network, a Gaussian process, a polynomial regression and / or a regression tree.

6. The method according to one of the preceding claims, in, The earpiece (14) is a dome (18) of a hearing aid (10); and / or Wherein, the earphone (14) is a hearing protector.

7. The method according to one of the preceding claims, further comprising: A type of headphone (28) is selected that provides the determined sound quality (26).

8. The method according to one of the preceding claims, further comprising: The geometric dimensions of the vent (20) of the earphone (14) are determined.

9. The method according to one of the preceding claims, further comprising: Manufacturing data of the earphone (14) is generated.

10. A training method for training a machine learning algorithm (24) for determining the sound quality (26) of an earphone (14) to be inserted into an ear of a user, the training method comprising: receiving a data set (32) having records of user data (22), each record comprising at least an audiogram of a user and a type of headphones (28) used by said user; for each record, determining at least one user score (34, 36) for the headphone type based on the user data (22); generating a filtered data set (38) by excluding records from the data set (32), wherein records are excluded from the data set (32) if the at least one user score (34, 36) is below a threshold; determining a sound quality (26) for each recording based on the headphone type (28); The machine learning algorithm (24) is trained using the filtered data set (38).

11. The training method according to claim 10, in, Each record of the data set (32) further includes the wearing time of the earphone (14); The user score is a customer satisfaction score (34) determined based on the wearing time.

12. The training method according to claim 10 or 11, in, The user score is an intelligibility score (36); wherein each record of the data set (32) further comprises a fitting formula for a hearing aid (10) to be used with the earphone (14); wherein, for at least one frequency, a desired target gain is determined according to the fitting formula; wherein, for the at least one frequency, the target gain limited by the feedback threshold is determined according to the headphone type; wherein the intelligibility score (36) is determined based on the target gain at the at least one frequency, which is limited by a feedback threshold.

13. The training method according to one of claims 10 to 12, further comprising: Determining an enhancement score (40) for each record based on the at least one user score (34, 36); When the enhancement score (40) of the record is above a threshold, an enhanced data set (42) is generated by at least copying the record.

14. A computer program for determining the sound quality (26) of an earphone (14) to be inserted into an ear of a user using a machine learning algorithm (24) and / or for training the machine learning algorithm (24), the computer program being adapted to perform the steps of the method of one of the preceding claims when being executed by a processor.

15. A computer readable medium having stored therein a computer program according to claim 14.

Images