Method for operating a hearing aid

Abstract

The invention relates to a method (28) for operating a hearing aid (2) comprising a neural network (18) with several neurons (22), each of which is assigned a weighting vector (40) with binary weights (44). Each neuron (22) is supplied with an input vector (34) with binary values ​​(36) and processed with the weighting vector (40) to obtain a transfer function (42). The transfer function (42) is processed with an activation function (48) such that a binary result (26) is provided. The invention further relates to a method (52) for training a neural network (18) and a hearing aid (2).

Description

[0001] The invention relates to a method for operating a hearing aid with a neural network. Furthermore, the invention relates to a method for training a neural network and a hearing aid.

[0002] People with hearing loss typically use a hearing aid. This usually involves an electromechanical transducer that captures ambient sound. The resulting electrical (audio) signals are amplified by an amplifier circuit and then delivered to the ear canal via another electromechanical transducer, often in the form of a receiver. The captured audio signals are usually also processed, typically by a signal processor within the amplifier circuit. The amplification is adjusted to the specific hearing loss of the hearing aid user, who is also referred to as the user or wearer.

[0003] Depending on the current situation, it may be necessary to adjust the processing to improve intelligibility for the user. Especially with speech, it is desirable to choose different processing settings for different syllables or sounds. For example, reducing certain frequencies improves intelligibility for some syllables, while reducing intelligibility for others.

[0004] One challenge in this processing lies in distinguishing between speech and background noise in the signal processed by the hearing aid. Artificial neural networks (hereafter referred to simply as neural networks) are particularly well-suited to solving this problem. Neural networks with feedforward components, where only the current input values ​​influence the state, and recurrent components, where the state is also influenced by the results of past processing steps, are especially suitable for processing audio signals. However, for neural networks to be effective in distinguishing between speech and background noise in the signal, very large networks with many layers are necessary; that is, the networks consist of a large number of neurons. This also applies to other typical applications of neural networks for audio signal processing.Due to their size, these neural networks have a large hardware and energy requirement, which is not desirable for a hearing aid, as these should be designed to be as small and energy-efficient as possible.

[0005] The invention is based on the objective of providing a particularly suitable method for operating a hearing aid, a particularly suitable method for training a neural network, and a particularly suitable hearing aid, wherein in particular user comfort is increased and hardware resources and / or energy requirements are advantageously reduced.

[0006] With regard to the method for operating a hearing aid, this problem is solved according to the invention by the features of claim 1, with regard to the method for training a neural network by the features of claim 1, and with regard to the hearing aid by the features of claim 5 and with regard to the hearing aid by the features of claim 8. Advantageous further developments and embodiments are the subject of the respective dependent claims.

[0007] The process is used to operate a hearing aid. For example, the hearing aid is a headphone or includes a headphone, and the hearing aid is, for example, a headset. However, the hearing aid is most commonly referred to as a hearing aid device. The hearing aid device serves to support a person suffering from a reduction in hearing ability. In other words, the hearing aid device is a medical device by means of which, for example, partial hearing loss is compensated. The hearing aid device is, for example, a receiver-in-the-canal (RIC) hearing aid, an in-the-ear (ITE) hearing aid, such as an in-the-ear (ITC) hearing aid, a complete-in-the-canal (CIC) hearing aid, hearing glasses, or a pocket hearing aid. Alternatively, the hearing aid is a behind-the-ear (BTE) hearing aid, which is worn behind the ear.

[0008] The hearing aid is designed and configured to be worn on the human body. In other words, the hearing aid preferably includes a retention device that allows it to be attached to the human body. If the hearing aid is a hearing assistance device, it is designed and configured to be placed, for example, behind the ear or within the ear canal. In particular, the hearing aid is wireless and designed and configured to be inserted, at least partially, into the ear canal.

[0009] The hearing aid preferably includes a microphone for capturing sound. In particular, when in operation, the microphone captures ambient sound, i.e., sound waves, or at least a portion thereof. The microphone is advantageously located at least partially within the housing of the hearing aid and is thus at least partially protected. The microphone is suitably an electromechanical transducer. The microphone may, for example, have only a single microphone unit or several microphone units that interact with each other. Each of the microphone units advantageously has a diaphragm that is set into vibration by sound waves, and the vibrations are converted into an electrical signal by means of a suitable recording device, such as a magnet moved within a coil.Alternatively, the microphone units can be designed capacitively, utilizing the fact that an applied electrical voltage changes when the distance between the diaphragm and a static surface of the microphone unit changes. In this case, the electrical voltage is applied specifically between the diaphragm and the static surface. The microphone units are preferably designed to be omnidirectional. In this or another way, it is at least possible to generate or at least provide an audio signal using the microphone, based on the sound incident on the microphone, namely, in particular, ambient sound.

[0010] Advantageously, the hearing aid includes a receiver for outputting a signal. The output signal is, in particular, an electrical signal, and may be, for example, digital or, more appropriately, analog. The receiver is preferably an electromechanical transducer, such as a loudspeaker. Depending on the design of the hearing aid, in its intended state, the receiver is at least partially positioned within the ear canal of a user of the hearing aid, i.e., a person also referred to as the wearer, user, or hearing aid wearer, or at least acoustically connected to it. The hearing aid primarily serves to output the signal via the receiver, thereby generating a corresponding sound. In other words, the main function of the hearing aid is preferably to output the signal.

[0011] The hearing aid suitably includes a signal processing unit by means of which the microphone and receiver are connected. Advantageously, the hearing aid has a signal processor that, for example, forms the signal processing unit or is at least a component thereof. The signal processor is, for example, a digital signal processor (DSP) or implemented using analog components. The signal processor, or at least the signal processing unit, is used in particular to adapt the audio signal generated by the microphone, preferably to create the output signal. At a minimum, the signal processing unit is suitable for this purpose, and in particular is designed and configured for this purpose. Advantageously, an analog-to-digital converter (ADC) is arranged between the microphone and the signal processing unit, for example, the signal processor, provided the signal processor is designed as a digital signal processor.Preferably, the hearing aid also includes an amplifier, or the amplifier is at least partially formed by the signal processing unit. For example, the amplifier is connected upstream or downstream of the signal processor in terms of signal processing.

[0012] The hearing aid also includes a neural network. The neural network is an artificial neural network. The neural network consists of multiple neurons.

[0013] The neurons are divided into different layers arranged sequentially. Each neuron is assigned a weight vector comprising various weights. The number of weights is equal to or greater than the number of neurons in the preceding layer. Specifically, the weight vectors assigned to neurons in different layers have a different number of weights. The weights are binary. Thus, there are only two different values ​​for each weight, for example, 0 (zero) and 1, or preferably -1 and 1. The two possible values ​​of the weight therefore differ only in their sign. Consequently, it is possible to represent each weight using only a single bit. For example, the neural network is used independently.Alternatively, the neural network is a component of a larger neural network where, for example, the weights are non-binary. Thus, the larger neural network contains the neural network with the binary weights.

[0014] In this method, each neuron is fed an input vector with various values. The number of values ​​corresponds to the number of weights in the respective weighting vector. The values ​​are binary and therefore also have only two different forms. For example, these are 0 (zero) and 1, or preferably -1 and 1. Thus, the different forms of the values ​​differ only in their sign. Here, too, it is possible to represent each value using only a single bit. For example, the values ​​are initially binaryized, particularly using another method, or in a preliminary step. This involves, for example, a comparison with a threshold value to realize the two different forms of each value.

[0015] The input vector and the weighting vector are processed together to create a transfer function. In other words, the input vector and the weighting vector are processed to obtain the transfer function, also known as the propagation function. Specifically, vector multiplication is performed to obtain this function. Conveniently, the dot product of the weighting vector and the input vector is calculated.

[0016] The transfer function is processed by an activation function, and a result is provided. This result is also binary, and the activation function is designed accordingly. Thus, the result has only one of two values, which are expediently -1 and 1. In particular, the binary result is subsequently used as one of the values ​​of another input vector, which is fed to another neuron, preferably a downstream layer. For example, only the values ​​in one direction are exchanged between the layers. In other words, the neural network is expediently a forward-facing neural network. Preferably, however, it is also possible for the result to be fed to the same layer or a preceding layer.This distinguishes it from a purely forward-directed neural network, and the neural network in question is preferably a recurrent neural network. Ideally, the neural network has both recurrent and forward-directed components.

[0017] Preferably, this is done for each neuron of the neural network, and in particular, a complete result, especially a result vector or a specific result value, is provided by means of the last position of the neurons.

[0018] Due to this design, each weight, value, and result can be represented by a single bit, resulting in comparatively low memory requirements. The required calculations are also relatively simple and few in number, thus accelerating the delivery of results and reducing hardware resources. Furthermore, the energy required is reduced. Normalization is unnecessary and, logically, omitted since only binary values / weights / results are used. This further reduces the number of required calculations, thus minimizing hardware resources and energy consumption. Consequently, the hearing aid can be designed to be relatively small and lightweight, without excessively increasing manufacturing costs. At the very least, battery life is extended.Consequently, user acceptance and therefore comfort are increased. This enables the processing of certain hearing aid functions or tasks using the neural network, thus increasing accuracy and / or user comfort.

[0019] For example, the transfer function is the dot product of the weight vector and the input vector, for which each weight is multiplied by its corresponding value, and the resulting products are summed. However, it is particularly preferred to use the XNOR operation, specifically an equivalence function. Therefore, an XNOR gate is expediently used. Thus, instead of creating the product, it is simply checked whether the respective weight equals the respective value. If so, one possible value is taken; otherwise, another possible value is taken. Specifically, the possible values ​​for the weights and the values ​​are -1 and 1, respectively, and the result of the XNOR operation is therefore also either -1 or 1.Such a calculation can be performed relatively quickly and efficiently using software or hardware, and no multiplication is required. The result of the XNOR operation is the same as the result of the product. This further reduces hardware resources and energy consumption while still achieving the same result.

[0020] For example, a comparison function is used as the activation function, and the transfer function is compared to a predefined limit to generate the result. If the transfer function is greater than this limit, one possible value is used as the result, and otherwise, the other possible value is used. Here, the limit can be arbitrary or, preferably, equal to 0 (zero). Particularly preferably, only the sign of the transfer function is used as the activation function. If the transfer function is positive, the sign is 1, and otherwise -1. This is essentially the same as comparing it to a limit of 0, but simplifies processing. When using the sign, it is only necessary to read the bit representing the sign and use it as the result.An explicit comparison with the limit is not required. This further reduces the number of required calculations, and consequently, the necessary hardware resources and energy consumption are also reduced. Furthermore, no subsequent normalization is necessary. It is expedient to use either 1 or -1 for the binary result.

[0021] For example, the input vector is determined based on measurement data from a sensor in the hearing aid. A motion sensor, for instance, is used as the sensor. However, it is particularly preferred that the input vector be created based on captured audio signals, specifically a time series. In this case, the input vector corresponds directly to the audio signal, or at least to the partially processed audio signal, at least in the first layer of the neural network. In subsequent layers, the input vector is advantageously generated using the results of the preceding layer. These are also based on the audio signal, so the respective input vectors are likewise created based on the captured audio signal. Thus, it is possible to determine the hearing aid's current environment with relative accuracy using the neural network.

[0022] Alternatively, or preferably in combination with this, a prediction for future audio signals is generated based on the results—that is, at least a portion of the results provided by the neurons, particularly the results of the neurons in the last layer of the neural network. This constitutes the "output" of the neural network. In other words, the prediction is generated based on the result vector or the result value, or these correspond to the prediction. In summary, an assumption is made about how the acoustic environment of the hearing aid will change, specifically how a current sentence or word will continue. For this purpose, the input vectors are generated based on the captured audio signals. Thus, an assumption is made about what the following syllable will be.Advantageously, the signal processing unit is adjusted based on the forecast, preferably by modifying the processing of the captured audio signal. If the subsequent audio signal then corresponds to the forecast, intelligibility for the user is improved. Since, for example, a recurrent neural network is used, which can analyze time series relatively effectively, the accuracy of the forecast is particularly improved.

[0023] Preferably, the precise design of the prediction / output depends on the application of the neural network, but is preferably further processed by other components of the hearing aid. The prediction / output is preferably used to influence the sound output at the listener, for example, by having the network generate an estimate of the speech and background noise components in the audio signal, which is then used to adjust frequency-dependent gain factors. In this way, intelligibility is improved for the user. Since, for example, it is a recurrent neural network, which can analyze time series relatively effectively, the accuracy of distinguishing between speech and background noise is particularly improved.

[0024] The other method is used to train a neural network for a hearing aid. This network consists of several neurons and is, for example, a recurrent neural network, or at least has recurrent components. Each neuron is assigned a weight vector with binary weights, at least once the neural network is fully trained. In the trained state, i.e., when the neural network is in use, each neuron receives an input vector with binary values. The weight vectors and input vectors are processed together to obtain a respective transfer function. The transfer function is then processed with an activation function to produce a binary result.

[0025] The method serves primarily to determine the weights. This process is carried out, for example, using the hearing aid itself or by the hearing aid manufacturer. In this case, the hearing aid is delivered with the neural network already trained. The method involves several training steps, with the weights being adjusted appropriately in each step. Preferably, the training steps are performed sequentially.

[0026] At each training step, it is advisable to first change the weights. This can be done, for example, arbitrarily, while adhering to certain guidelines, or in a predetermined manner. According to the procedure, the weights are first binarized. In other words, each weight is assigned one of only two possible values. In other words, a digitization, or binary mapping, of an existing weight to one of the two values ​​takes place. Ideally, one of the values ​​is -1 and the other is 1.

[0027] The binarized weights are normalized. This ensures that all weight vectors in this training step and / or subsequent training steps always have the same norm, specifically the same length. For normalization, each weight is multiplied by a corresponding normalization constant. The normalization constant is the same for all weights assigned to the same weight vector. It is possible for the normalization constants assigned to different weight vectors to differ. In summary, each normalization constant applies only to the weights assigned to the same weight vector. The normalization constant is inversely proportional to the Euclidean norm of the respective weight vector.In some cases, each weight is divided by the Euclidean norm, or the normalization constant includes further components. For example, it might also involve multiplication by a specific factor. The Euclidean norm corresponds specifically to the length of the weighting vector. However, since the weights are binary before normalization, and preferably equal to -1 or 1, the normalization constant is thus equal to the number of weights in the respective weighting vector. The normalization constant is therefore always the same and does not depend on statistics or similar factors, which is why only minimal hardware resources are required to perform the training.

[0028] The transfer function for each neuron is then generated based on the respective weight vector, which has now been nominated. For this, the weight vectors are processed with an input vector, where, due to normalization, the weights do not simply have the values ​​-1 and 1. Specifically, the transfer function corresponds to the dot product of the weight vector and the input vector. The transfer function is then processed with the activation function to produce the binary result.

[0029] Preferably, this is performed for all neurons of the neural network, with all neurons being used once per training step. Advantageously, the neural network outputs the resulting vector or value. Preferably, a further value is determined based on this and a cost function. During the execution of the cost function, a comparison is made with a desired result vector / value, which, preferably like at least some of the input vectors, is provided by training data.

[0030] For example, there are specific requirements for choosing weights at the beginning of each training step. However, it is preferable to have no such requirements, and the weights can be chosen essentially arbitrarily at the beginning, especially at the start of the training session, or at each training step. This allows for the use of an existing algorithm, thus reducing the effort required. Furthermore, the process can be implemented relatively efficiently on existing hardware or software.

[0031] After completing the training steps, and especially when the provided result vector / value corresponds sufficiently closely to the training data, the weights are binarized and multiplied by the sign of the respective normalization constant. Thus, each weighting vector then contains only the binary weights. This allows the weighting vectors to be used directly for operating the hearing aid.

[0032] A hearing aid can be, for example, a headset or, more commonly, a hearing aid device. Examples include receiver-in-the-canal (RIC) hearing aids, in-the-ear (ITC) hearing aids, complete-in-canal (CIC) hearing aids, hearing glasses, or pocket hearing aids. Alternatively, a hearing aid can be a behind-the-ear (BTE) hearing aid, which is worn behind the ear.

[0033] The hearing aid preferably includes a microphone. This microphone is, for example, omnidirectional, or its directional characteristics can be adjusted. For this purpose, the microphone preferably has two or more microphone units. The microphone is designed and configured to detect ambient sound. Advantageously, an audio signal is generated by the microphone when ambient sound is detected. The hearing aid preferably includes a signal processing unit, which is preferably connected to the microphone. In particular, the audio signal is fed to the signal processing unit during operation. Preferably, the hearing aid includes a receiver through which the processed audio signal is output, and which is advantageously connected to the signal processing unit.

[0034] The hearing aid further comprises a neural network with multiple neurons. The neurons are suitably distributed across different layers. For example, the neural network is implemented solely by software, or the hearing aid incorporates dedicated hardware for this purpose, in particular a specially adapted chip. Preferably, the neural network is assigned to the signal processing unit. For example, the neural network is either a recurrent neural network or a forward neural network. Preferably, however, the neural network comprises both recurrent and forward components.

[0035] Each neuron is assigned a weighting vector with binary weights, at least during normal operation. The hearing aid operates according to a method in which each neuron is fed an input vector with binary values ​​and processed with the weighting vector to obtain a transfer function. The transfer function is then processed with an activation function to provide a binary result. Alternatively, or preferably in combination with this, a method for training the neural network is performed using the hearing aid, for example, only once, or also during use, expediently at specific intervals. In this process, the weights are binaryized and normalized by multiplying them by a respective normalization constant that is inversely proportional to the Euclidean norm of the respective weighting vector.For each neuron, the respective transfer function is created based on the respective weighting vector, and the transfer function is processed with the activation function in such a way that a binary result is provided.

[0036] The signal processing unit is expediently suitable, in particular designed and equipped, to carry out at least part of one or both procedures.

[0037] The further training and advantages explained in connection with the two procedures can also be applied analogously to the hearing aid and to each other, and vice versa.

[0038] An embodiment of the invention is explained in more detail below with reference to a drawing. The drawing shows: Fig. 1 schematically simplified a hearing aid with a neural network, Fig. 2 schematically the neural network comprising several neurons, Fig. 3 a method for operating the hearing aid, Fig. 4 schematically a neuron, and Fig. 5 a method for training the neural network.

[0039] Corresponding parts are marked with the same reference symbols in all figures.

[0040] In Figure 1A simplified schematic representation of a hearing aid 2 is shown. The hearing aid 2 has a housing 4, inside which a microphone 6 is arranged. The microphone 6 has several microphone units (not shown in detail), each designed as an electromechanical transducer or a capacitive transducer. A signal processing unit 8 is connected downstream of the microphone 6. A receiver 10 is connected downstream of the signal processing unit 8, by means of which, when used as intended, sound can be emitted into the ear canal of the user (not shown in detail).

[0041] The signal processing unit 8 includes a processing unit 12, which processes an audio signal 14 provided by the microphone 6 during operation, thus generating an output signal 16. This involves frequency-selective amplification and / or attenuation, suppressing, for example, noise or other interfering sounds. The processing unit 12 includes, for example, a digital sound processor for this purpose. Compression is also performed, for example, so that the frequency spectrum of the output signal 16 is reduced compared to the audio signal 14. The output signal 16 is routed to the earpiece 10, so that the sound emitted by the earpiece 10 corresponds to the output signal 16.

[0042] The signal processing unit 8 also includes a neural network 18, which is likewise fed the audio signal 14 and has recurrent components. Using the neural network 18, a prediction 20 for future audio signals is generated and fed to the processing unit 12. Depending on the prediction 20, the processing unit 12 is adjusted, specifically by selecting gain factors for different frequencies. The processing unit 12 is then adjusted to the audio signal 14, which is likely to follow in time and is provided by the microphone 6, resulting in improved processing. This improves the intelligibility of the sound output through the headphones 12 for the user. In summary, the neural network 18, in particular, uses the audio signal 14 and, for example, other sensor data, to generate a signal 14.Acceleration values ​​and intermediate values ​​are calculated, representing the prediction 20, which are useful for processing the audio signal 14. For example, the presence or absence of speech is estimated ("voice activity detection").

[0043] In Figure 2The neural network 18, also referred to simply as the neural network, is shown schematically in a simplified form. The neural network 18 has several neurons 22 assigned to different layers 24, with three layers 24 shown in the example. Each neuron 22 provides an output 26, which is assigned to the neurons 22 of the next layer 24. However, it is also possible for one or more of the outputs 26 to be sent to a neuron 22 of the same layer 24 or to one of the preceding layers 24. For clarity, only some of the outputs 26 are shown in Figure 2. The audio signal 24 is assigned to the first layer 24, and the outputs 26 of the neurons 22 of the last layer 24 form the prediction 20.

[0044] In Figure 3A procedure 28 for operating the hearing aid 2 is shown. In a first step 30, the audio signal 14 is provided by means of the microphone 6. This is directed to the neural network 18. In a subsequent second step 32, an input vector 34 is created based on the acquired audio signal 18, which is fed to the neurons 22 of the first layer 24, one of which is in Figure 4 This is shown schematically. The input vector 34 has several binary values ​​36, which are also simply referred to as values. The values ​​36 can take on two different forms, namely -1 and 1.

[0045] In a subsequent third step 38, the input vector 34 is processed with a weight vector 40 assigned to the respective neuron 22, resulting in a transfer function 42. In other words, the input vector 34 is processed with the weight vector 40 to obtain the transfer function 42. The weight vector 40 has several binary weights 44, which are also simply referred to as weights. Each binary weight 44 can only take on two values: -1 and 1. The number of binary weights 44 is equal to the number of binary values ​​36 of the input vector 34, and each weight 44 is assigned one of the values ​​36.

[0046] To create the transfer function 42, an XNOR operation is performed. This checks whether the respective binary weight 44 equals the corresponding binary value 36. If they match, the result of the XNOR operation is 1; otherwise, it is -1. The sum of the results corresponds to the transfer function 42. In summary, the transfer function 42 is the sum of the XNOR operation on all binary values ​​36 and their respective corresponding binary weights 44.

[0047] In a subsequent fourth step 46, the transfer function 42 is processed with an activation function 48 to provide the respective result 26. The sign of the transfer function 42 is used as the activation function 48. The result 26 is therefore also binary, corresponding to 1 if the transfer function is positive and -1 otherwise.

[0048] The results 26 obtained in this way for one of the layers 24 are combined to form the input vector 34 for the subsequent layer 24. If necessary, other results 26 are also added to the respective input vector 34.

[0049] Steps 2 to 4, 32-46, are performed for all neurons 22, in particular layer by layer, or possibly in a different sequence. Since all results 26 are generated, at least indirectly, from the recorded audio signals 18, all input vectors 34 are consequently generated from the recorded audio signals 18. The results 26 of the last layer 24 are then summarized to form the prediction 20.

[0050] After each neuron 22 has generated the respective result 26 at least once, a fifth processing step 50 is carried out, and the prediction 20 is output, based on which the processing unit 12 is then set. Using this unit, the audio signal 24 is then processed according to its creation, so that the output signal 16 can be generated.

[0051] To generate the prediction 20, only 1 bit of storage space is required for each of the weights 44 and for each of the values ​​36, compared to 8 bits when using non-binary values / weights. This reduces the required storage to approximately one-eighth. When performing the XNOR operation, a Boolean operation stored in a respective chip can be used, eliminating the need for multiplication with multiple individual arithmetic operations. Furthermore, no normalization is performed, and to execute the activation function 48, only the bit corresponding to the sign of the transfer function 42 is read. Therefore, no arithmetic operations are required here either. In summary, determining the prediction 20 is comparatively resource-efficient, resulting in a relatively low energy requirement for carrying out the procedure 28 to operate the hearing aid 2.

[0052] In Figure 5 A method 52 for training the neural network 18 is shown. This method is performed, for example, by a manufacturer of the hearing aid 2, particularly before the hearing aid 2 is made available to the user. Alternatively, the user performs the method 52 for training the neural network 18, for example, when starting to use it, or after the user has already used it for a certain period of time, such as a week or a month. In another alternative, the method 52 for training the neural network 18 is performed several times at specific intervals. The method 52 for training the neural network 18 is performed, in particular, using the hearing aid 2 itself.

[0053] The procedure 52 for training the neural network 18 is started in a sixth step 54. In this step, all weights 44 are initially chosen arbitrarily. In other words, there is no restriction regarding the values, and they can be selected from a continuous range of values.

[0054] In a subsequent seventh step 56, the weights 44 are binarized. In other words, each weight 44 is mapped to one of only two distinct values, in this example either -1 or 1. Thus, after the seventh step 56, all weights 44 have one of two distinct values. If the neural network 14 is a component of a larger neural network that is fully trained, it is possible that the larger neural network contains non-binary weights that are not part of neural network 14.

[0055] In a subsequent eighth step 58, the weights 44 are normalized. For this purpose, they are multiplied by a normalization constant 60. The normalization constant 60 is the reciprocal of the Euclidean norm of the respective weighting vector 40. Since the weights 44 are either 1 or -1, the Euclidean norm corresponds to the number of weights 44 of the respective weighting vector 40. Thus, each of the weights 44 is divided by the number of weights 44 of the associated weighting vector 40. Depending on how many results 26 are fed to the respective neuron 22, i.e., depending on the connection of the neural network 18, the normalization constants 60 of different weighting vectors 40 differ.

[0056] In a subsequent ninth step 62, the transfer function 42 is created for each neuron 22 using its respective assigned weight vector 40, which has the normalized weights 44. For this purpose, each neuron 22 is supplied with the corresponding input vector 34 with the binary values ​​36, and the dot product with the weight vector 40 is calculated. The input vector 34, which is used for the first layer 24, is provided using training data and has only binary values.

[0057] Furthermore, the activation function 48 is processed based on the respective generated transfer function 42, resulting in the binary result 26. If the transfer function 42 is positive, the value 1 is used as the result 26; otherwise, it is -1. This is performed for all neurons 22, resulting in the prediction 20. This prediction is then compared with an expected prediction provided by the training data. Depending on this comparison, the weights 44 are adjusted, and steps seven through nine (56-62) are repeated. This process continues until the generated prediction 20 matches the expected prediction with sufficient accuracy.

[0058] If this is the case, a tenth step 64 is performed. In this step, the weights 44, which may have different values ​​due to the normalization last performed in the eighth step 58, are binarized, i.e., set to either -1 or 1. If the respective weight 44 is positive, 1 is used; otherwise, -1 is used. These are then multiplied by the sign of the respective normalization constant 60. Since this constant is always positive, the multiplication is by 1. Subsequently, each of the weight vectors 40 now only has the binary weights 44, and the neural network 18 is trained and can be used to perform the procedure 28 for operating the hearing aid 2.

[0059] In one variant, not shown in detail, the binarization of the weights 44 is performed only on the last repetition of the seventh step 56 or only with the tenth step 64. In this case, the range of values ​​for the weights 44 is preferably specified in the seventh step 56, such that they can only lie between -1 and 1, although several values ​​are possible. With each repetition of the seventh step 64, the possible values ​​then approach either -1 or 1 more and more closely.

[0060] The invention is not limited to the embodiment described above. Rather, other variants of the invention can also be derived by a person skilled in the art without departing from the subject matter of the invention. In particular, all individual features described in connection with the embodiment can also be combined with one another in other ways without departing from the subject matter of the invention. Reference symbol list

[0061] 2 Hearing aid 4 Housing 6 Microphone 8 Signal processing unit 10 Receiver 12 Processing unit 14 Audio signal 16 Output signal 18 Neural network 20 Prediction 22 Neuron 24 Position 26 Result 28 Procedure for operating the hearing aid 30 First step 32 Second step 34 Input vector 36 Binary value 38 Third step 40 Weighting vector 42 Transfer function 44 Binary weight 46 Fourth step 48 Activation function 50 Fifth step 52 Procedure for training the neural network 54 Sixth step 56 Seventh step 58 Eighth step 60 Normalization constant 62 Ninth step 64 Tenth step

Claims

1. Method (28) for operating a hearing aid (2) comprising a neural network (18) with several neurons (22), each of which is assigned a weighting vector (40) with binary weights (44), wherein each neuron (22) is supplied with an input vector (34) with binary values ​​(36) and processed with the weighting vector (40) to obtain a transfer function (42), and wherein the transfer function (42) is processed with an activation function (48) such that a binary result (26) is provided.

2. Method (28) according to claim 1, characterized by that The transfer function (42) is the sum of the XNOR operation on all values ​​(36) with the respective assigned weight (44).

3. Method (28) according to claim 2, characterized by that The sign of the transfer function (42) is used as the activation function (46).

4. Method (28) according to any one of claims 1 to 3, characterized by that the input vector (34) is created from captured audio signals (18), and / or a prediction (20) for future audio signals is created from the results (26).

5. Method (52) for training a neural network (18) according to any one of claims 1 to 4, wherein in each training step - the weights (44) are binarized, - the weights (44) are normalized, for which they are multiplied by a respective normalization constant (60) which is inversely proportional to the Euclidean norm of the respective weighting vector (44), - for each neuron (22) the respective transfer function (42) is created on the basis of the respective weighting vector (44), and - the transfer function (48) is processed with the activation function (48) such that the binary result (26) is provided.

6. Method (52) according to claim 5, characterized by that At the beginning, the weights (44) can be chosen arbitrarily.

7. Method (52) according to claim 5 or 6, characterized by that After performing the training steps, the weights (44) are binarized and multiplied by the sign of the respective normalization constant (60).

8. Hearing aid (2) comprising a neural network (18) with multiple neurons (22) and operated according to a method (28, 52) according to any one of claims 1 to 7.

Images