Method for determining wearer's own voice activity, hearing device and hearing device system

Abstract

The invention relates to a method for determining the own voice activity of a wearer of a hearing device (4) by means of a signal processing device (8) of the hearing device (4), in which a first input signal is generated by a first input converter (6) of the hearing device (4) and a second input signal is generated by a second input converter (6) of the hearing device (4), the two input signals are fed to a determination unit (22) of the signal processing device (8), which determination unit has a neural network (24) and a preceding stage (26) connected upstream of the neural network (24), an information signal is generated by the preceding stage (26) on the basis of the two input signals, the information signal is evaluated by the neural network (24), and the signal processing device (8) is configured to process the information signal. The determination unit (22) outputs a determination result on the basis of an evaluation of the information signal by the neural network (24).

Description

Technical Field

[0001] The invention relates to a method for determining own speech activity of a wearer of a hearing device. The invention also relates to a hearing device and a hearing device system. Background Art

[0002] Conventional hearing aids used to care for people with hearing impairments are often referred to as hearing devices. However, in a broad sense, the term also refers to devices designed to support people with normal hearing. Hearing devices designed to assist people with normal hearing are also called "personal sound amplification products" or "personal sound amplification devices" (abbreviated: "PSAD"). Unlike conventional hearing aids, such hearing devices are not prescribed to compensate for hearing loss, but are specifically designed to support and improve normal hearing in specific hearing situations.

[0003] Regardless of the intended use, a hearing device generally has at least one input converter, a signal processing device and an output converter as essential components. The at least one input converter is generally formed by an acoustic-electric transducer, i.e., for example, by a microphone or by an electromagnetic receiver, such as an induction coil. In many cases, a plurality of input converters are installed, i.e., for example, one or more acoustic-electric transducers and electromagnetic receivers. An electroacoustic transducer, such as a miniature loudspeaker (also referred to as a receiver), or an electromechanical transducer, such as a bone conduction earpiece, is generally used as the output converter. The signal processing device is generally implemented by an electronic circuit implemented on a printed circuit board and generally has a separate amplifier or amplifier unit. In addition, the hearing device is generally equipped with a transmitting and / or receiving unit, which enables wireless communication with other electronic devices, in particular with other hearing devices.

[0004] In some applications, two such hearing devices form a hearing device system, in particular a binaural hearing device system. Typically, one of the hearing devices is designed for the left ear, and the other hearing device is designed for the right ear.

[0005] It is known that, in particular when using corresponding hearing instrument systems, the perception of one's own voice changes and therefore appears unfamiliar and strange. Since this effect is generally undesirable, during the signal processing by the signal processing device, different signal processing processes are preferably used depending on whether the signal based on which the signal processing is carried out also depicts one's own voice.

[0006] In order to enable such a discriminatory signal processing, the corresponding hearing devices are provided for so-called own voice detection (OVD). These hearing devices thus determine when the hearing device wearer's own voice is active and when it is inactive. Such own voice detection is described, for example, in document EP 3 222 057 B1. In particular, if the own voice detection is also personalized in some way so that the recognition of the specific voice of a specific hearing device wearer is possible, it is often also referred to as own voice recognition. Summary of the invention

[0007] On this basis, the technical problem to be solved by the present invention is to provide an advantageous method for determining the own speech activity of a hearing device wearer. The technical problem to be solved by the present invention is also to provide an advantageously designed hearing device and an advantageously designed hearing device system.

[0008] The technical problem is solved according to the invention by a method for determining the own speech activity of a wearer of a hearing device by means of a signal processing device of a hearing device, wherein a first input signal is generated by a first input transducer of the hearing device and a second input signal is generated by a second input transducer of the hearing device, wherein the two input signals are fed to a determination unit of the signal processing device, the determination unit having a neural network and a preamplifier connected upstream of the neural network, wherein an information signal is generated by the preamplifier based on the two input signals, wherein the information signal is evaluated by the neural network, and wherein the determination unit outputs a determination result based on the evaluation of the information signal by the neural network. The technical problem is also solved according to the invention by a hearing device, which is configured to carry out at least one method step of a method according to the aforementioned type in at least one operating mode. The technical problem is also solved according to the invention by a hearing device system, which has two hearing devices according to the aforementioned type. Preferred further developments are contained in the description. The advantages and preferred embodiments mentioned with respect to the method can also be transferred to the device, i.e. the hearing device and the hearing device system, and vice versa.

[0009] The method according to the invention is used to determine the activity of own speech or one's own voice, i.e. the voice of a person wearing a hearing device. The method advantageously comprises at least one method part, i.e. the determination of the activity itself, wherein in particular the method part is carried out by the hearing device, i.e. by means of a signal processing device of the hearing device.

[0010] The hearing device according to the invention is in turn designed and arranged to carry out at least one method step of the method according to the invention in at least one operating mode. In particular, the hearing device according to the invention is designed to carry out at least one of the above-mentioned method parts, i.e. the determination of the activity itself. For this purpose, the hearing device according to the invention has the above-mentioned signal processing device. Therefore, the hearing device according to the invention is arranged to automatically or automatically recognize when the sound of the hearing device, i.e. the wearer of the hearing device according to the invention, is active and when it is inactive in at least one operating mode.

[0011] Finally, the hearing device system according to the invention has two hearing devices according to the invention of the above-described type. Preferably, one of the hearing devices is designed for the left ear and the other hearing device is designed for the right ear. Furthermore, the hearing device system is preferably designed as a so-called binaural hearing device system.

[0012] During the implementation of the method, ie the method according to the invention and in particular at least one method part, a first input signal is generated by a first input transducer of the hearing device, ie the hearing device according to the invention, and a second input signal is generated by a second input transducer of the hearing device.

[0013] The two input signals are then fed to a determination unit of a signal processing device, which has a neural network and a preceding stage connected upstream of the neural network. Alternatively, the two input signals are fed to a filter bank of the signal processing device and decomposed there into partial signals. In this case, the partial signals are fed to the determination unit. In other words, the two input signals are fed to the determination unit undecomposed or decomposed.

[0014] The corresponding filter bank is usually designed in such a way that two or more frequency channels, frequency bands or frequency ranges, i.e. n frequency channels, frequency bands or frequency ranges, are predetermined by the filter bank and are decomposed into partial signals at least approximately in such a way that each partial signal has all frequency components of the input signal that can be assigned to a frequency channel / frequency band / frequency range k of k=1...n in the frequency channels / frequency bands / frequency ranges. A hearing device with such a filter bank is also referred to as a multi-frequency channel hearing device. If the hearing device is designed as a multi-frequency channel hearing device, the hearing device is expediently designed as an n-frequency channel hearing device, where n is, for example, 2, 4, 8, 16, 32 or 48. The frequency spectrum is thus divided into n frequency bands, frequency ranges or frequency channels.

[0015] Furthermore, information signals are generated by the preceding stage based on the two input signals, i.e. the undecomposed or decomposed input signals, and these information signals are subsequently evaluated by the neural network. The determination unit outputs a determination result based on the evaluation of the information signal by the neural network. The determination result is usually a simple binary signal, which indicates whether the own voice, i.e. the voice of the wearer of the hearing device, is currently active.

[0016] Based on the determination result, an output signal is preferably generated for an output transducer of the hearing device. Typically, different signal processing procedures or parameterizations are used to generate the output signal depending on whether the determination result indicates that the wearer's own voice is currently active.

[0017] Furthermore, in the method described herein, the personalization of the determination unit is preferably omitted. That is to say, the own voice determination achieved by the method is not adjusted or adapted to a specific wearer, i.e. a specific person. Therefore, it is further preferred that the manufacturer does not adjust the determination unit after delivery or delivery. In this case, the manufacturer not only presets the hardware of the determination unit, but also the software, programming and / or settings of the determination unit. In particular, no subsequent changes of parameters or parameter values ​​occur, at least no changes in which the wearer is involved in any way.

[0018] The above-mentioned neural network is preferably a so-called artificial neural network or KNN (English: artificial neural network, ANN) for short. Depending on the application, the artificial neural network is designed as a so-called feedforward network, for example. Preferably, the artificial neural network is designed as a so-called recursive network. Independently of this, the artificial neural network preferably has two or more so-called layers. Therefore, the artificial neural network is especially designed as a multilayer neural network (English: deep neuronal network, DNN). Alternatively, only one layer is designed. Suitable topological structures, for example, present artificial neural networks designed in the form of LSTM (long short-term memory), GRU (gated recursive unit) or CNN (convolutional neural network) or neural networks with mixed topological structures.

[0019] In order to implement the method or at least at least one of the method parts described above, i.e. the determination of the activity itself, the front stage also preferably has two filters, i.e. a first filter and a second filter, wherein the first filter corresponds to a first filter type and the second filter corresponds to a second filter type. In at least one operating mode of the hearing device, a first filtered signal is generated by the first filter and a second filtered signal is generated by the second filter.

[0020] The first filter type and the second filter type are usually filters for spatial separation, in particular so-called beamformers. Therefore, these filters are not filters for pure frequency separation, for example filters for realizing the above-mentioned filter banks.

[0021] In particular, if the hearing device is designed as a multi-frequency channel hearing device as described above, it is further advantageous that the pre-stage is designed as a multi-frequency channel pre-stage. That is, either the sub-signals are fed to the pre-stage in the above-described manner, or the pre-stage itself has a filter bank of the above-described type, by which the fed undecomposed input signal is decomposed into sub-signals. Typically, the pre-stage is designed as an n-frequency channel pre-stage, wherein n is, for example, equal to 2, 4, 6, 8, 16, 32 or 48. Thus, the spectrum is divided into n frequency bands, frequency ranges or frequency channels. Accordingly, in the pre-stage, at least after the filter bank, there are n sub-signals for each input signal. The further processing of the sub-signals in the pre-stage is usually carried out in a frequency channel-by-frequency manner. That is, typically in a frequency channel k of k=1...n, only the sub-signals in the first input signal having frequency components from the frequency band k and the sub-signals in the second input signal having frequency components from the frequency band k are further processed.

[0022] In an advantageous embodiment, the multi-frequency channel front stage has two filters for each frequency channel k of k=1 ... n, namely a first frequency channel k filter and a second frequency channel k filter, wherein the first frequency channel k filter corresponds to a first filter type and the second frequency channel k filter corresponds to a second filter type. In at least one operating mode, two filter signals are generated in each frequency channel k, namely a first frequency channel k filter signal is generated by a filter of the first type and a second frequency channel k filter signal is generated by a filter of the second type.

[0023] As mentioned above, filters of the first filter type and filters of the second filter type are usually filters for spatial separation. Preferably, each of the two filter types is designed so that the common directional characteristics of the two microphones of the hearing device are respectively realized by these filter types, that is, the first directional characteristic is realized by the first filter type and the second directional characteristic is realized by the second filter type. Therefore, preferably two different directional characteristics are realized by these two filter types.

[0024] The two directional characteristics differ from one another in particular with regard to a predefined spatial direction, along which the sensitivity of the respective directional characteristic reaches its maximum value. The two respective spatial directions are expediently predefined relative to the frontal direction, i.e. in particular by an angle for each of the two spatial directions relative to an angle of 0° predefined for the frontal direction. If the hearing device is worn by a wearer, the frontal direction usually points in the direction in which the face of the wearer of the hearing device is facing. If the wearer is also looking straight ahead, the frontal direction also points in the viewing direction of the wearer.

[0025] For the first filter type, a spatial direction is preferably predetermined relative to the front direction and in particular pointing toward the front hemispherical space starting from the hearing device wearer and his head orientation. Further preferably, a spatial direction is predetermined in the range between 270° and 90° and in particular in the range between 340° and 20°, for example approximately 0°, relative to the front direction.

[0026] For the second filter type, a spatial direction is preferably provided relative to the front direction and, in particular, from the hearing device wearer and his head orientation, pointing into the rear hemispherical space. Further preferably, a spatial direction is provided in a range between 95° and 265° and in particular in a range between 160° and 200°, for example, approximately 180° relative to the front direction.

[0027] In a further advantageous embodiment variant, the first and second filter types are designed such that the first filter type suppresses the own speech less strongly than the second filter type. Thus, during operation, the second filter type suppresses signal components that reproduce the own speech more strongly than the first filter type.

[0028] In a further advantageous embodiment variant, the first filter type is designed to extract or highlight useful signals from a defined useful signal source, in particular from a defined useful signal source located in front of the hearing device wearer. The conversation partner is usually provided as a defined useful signal source.

[0029] At least in some advantageous embodiments, the second filter is designed to shield or suppress the own speech and thus also the own speech of the hearing device wearer, in particular the current wearer of the hearing device. In this case, the signal component that reproduces the own speech is preferably shielded or suppressed more strongly than all other signal components.

[0030] It is also advantageous that the two filter types are designed as static filters with respectively fixedly preset filter coefficients. In this case, the filter coefficients are preferably preset by the manufacturer and are no longer changed after being preset by the manufacturer.

[0031] According to an alternative variant, only one of the two filter types is designed as a static filter with fixedly preset filter coefficients. In this case, the other filter type is usually designed as an adaptive filter. A variant in which both filter types are designed as adaptive filters is also suitable. In the case of corresponding adaptive filters, the filter coefficients are not fixedly preset. Instead, the filter coefficients are adjusted by the hearing device during operation of the hearing device.

[0032] As described above, according to at least one preferred embodiment, the first filter type is designed as a filter for spatial separation and is designed to extract or highlight useful signals from a defined useful signal source, and the filter has a preset spatial direction in terms of directional characteristics. It should be noted that as soon as the preset spatial direction deviates from the direction in which the useful signal source is located, the filter used for spatial separation will generally also at least partially suppress the useful signal. This is a typical result of commonly used filtering algorithms. How strong this suppression is depends primarily on how much the spatial direction deviates from the direction of the useful signal source.

[0033] Such presetting of the filter coefficients is usually not part of the at least one method part (hereinafter referred to as the second method part) described above, but is part of a further method part (hereinafter referred to as the first method part), which is usually not performed by the hearing device. The further method part, i.e. part of the first method part, is preferably a method step of determining suitable filter coefficients. The determined suitable filter coefficients are then predefined, in particular stored in the hearing device, for the determination of the filter and therefore for the second method part, i.e. the activity itself.

[0034] Suitable filter coefficients are preferably determined by means of an analysis of the recorded acoustic signals. The recording is usually carried out using a hearing device prototype or a hearing device model, and the analysis is further preferably carried out by means of a separate computing unit. The respective hearing device prototype, the hearing device model or the respective computing unit further preferably has adaptive test filters or simulates respective adaptive test filters, by means of which suitable filter coefficients for the static filter of the hearing device are determined.

[0035] In order to determine suitable filter coefficients for the second filter type, recorded acoustic signals are usually used, in which the own speech of different test wearers is recorded. That is, different people act as test wearers, i.e., in particular as test wearers of the aforementioned hearing device prototype or hearing device model, and analyze the acoustic signals with their own speech in order to find suitable filter coefficients for the second filter type and then use these filter coefficients as fixed preset filter coefficients. On the other hand, in order to determine suitable filter coefficients for the first filter type, recorded acoustic signals are usually used, in which different useful signal sources are recorded, which are in particular located in front of the front of the aforementioned test wearer or the simulation head. Usually a conversation partner is used as the useful signal source. In both cases, preferably a recorded acoustic signal without background noise is used.

[0036] In the first method part described above, the neural network is preferably trained, in particular without the subsequent hearing device wearer being involved. This training is usually carried out by the hearing device manufacturer.

[0037] Furthermore, it is further preferred that the neural network is trained with the aid of some recorded acoustic signals. These acoustic signals typically include recorded acoustic signals for determining suitable filter coefficients. Alternatively or additionally, the some recorded acoustic signals used for training the neural network include recorded acoustic signals with different background noises.

[0038] To train the neural network, according to one embodiment variant, for example, a recorded acoustic signal is used in which the own speech of different test wearers is recorded together with different background noises; and / or a recorded acoustic signal is used in which different useful signal sources are recorded together with different background noises; and / or a recorded acoustic signal is used in which the own speech of different test wearers is recorded together with different useful signal sources and with different background noises.

[0039] According to a further embodiment variant, the own speech of different test wearers, different useful signal sources and different background noises are recorded separately and subsequently combined to train the neural network, if necessary using different weights.

[0040] Furthermore, it is advantageous to use the aforementioned hearing device prototype or hearing device model in the training process in order to record the aforementioned acoustic signals. In some cases, the recorded signals are also used to first train a neural network implemented in the hearing device prototype or hearing device model or implemented by a separate computing unit. The training of the neural network in the hearing device is then further preferably carried out by transferring the training state of the neural network in the hearing device prototype, the hearing device model or the separate computing unit to the neural network in the hearing device.

[0041] The preceding stage of the determination unit preferably generates two or more filtered signals, i.e. a first filtered signal and a second filtered signal or a first frequency channel n filtered signal and a second frequency channel n filtered signal in the above-described manner. It is further preferred that each filtered signal is associated with a reference signal in the preceding stage in order to generate at least one information signal, the reference signal being based on at least one of the two input signals.

[0042] In an advantageous embodiment, a real-valued attenuation variable is determined by associating one of the plurality of filter signals with one of the plurality of reference signals, wherein an information signal is advantageously generated based on the attenuation variable. According to an alternative, a complex-valued attenuation variable is determined. Two implementation variants are provided for this alternative. In the case of a first implementation variant, one information signal is generated based on the complex-valued attenuation variable. In the case of a second implementation variant, two information signals are generated based on the complex-valued attenuation variable. In this case, for example, separate information signals are generated for the real part and the imaginary part or for the amplitude and the phase.

[0043] As described above, the information signal generated in the preceding stage is evaluated by the neural network, wherein the neural network expediently outputs an evaluation result based on the evaluation. The evaluation of the information signal and in particular the attenuation variable is further preferably performed in such a way that the neural network outputs a prediction value, i.e. a probability value, in particular a value in the value range of 0 to 1 as the evaluation result.

[0044] It is further preferred that the evaluation result of the neural network is fed to a final stage of a determination unit, which preferably has a comparator unit and is suitably connected downstream of the neural network. In the comparator unit, the evaluation result is preferably compared with a reference value, in particular a threshold value, so that a determination result is generated. For example, if the above-mentioned probability value is higher than the above-mentioned threshold value, it is determined as a determination result that the wearer's own voice is active. Conversely, if the probability value is lower than the threshold value, it is determined as a determination result that the own voice is not active. The determination result is preferably output by the determination unit.

[0045] According to another embodiment variant, although the evaluation result is compared with the reference in the comparator unit, only a preliminary determination result is generated thereby. This preliminary determination result is then preferably further analyzed together with other preliminary determination results in the final stage, for example in an additional logic unit of the final stage, so as to obtain the determination result, i.e. the final determination result. This determination result is usually output by the determination unit.

[0046] The additional preliminary determination result is preferably obtained in another hearing device by another determination unit and transmitted by the other hearing device. The variant of the method is particularly advantageous for a hearing device system having two hearing devices of the above-mentioned type. In such a hearing device system, each of the two hearing devices preferably has the aforementioned signal processing device with a determination unit. In at least one operating mode, a preliminary determination result is preferably generated in each determination unit. Subsequently, the preliminary determination result of one hearing device is usually transmitted wirelessly to another hearing device and analyzed there together with its own preliminary determination result. The final determination result obtained in this process is output by the corresponding determination unit. The final determination result can be used internally. In another hearing device, preferably, almost the same process is carried out simultaneously. In the other hearing device, the two preliminary determination results are also analyzed and the final determination result is obtained.

[0047] Independently of this, each hearing device according to the invention has at least two of the aforementioned input converters, the aforementioned signal processing device and the aforementioned output converter as essential components. Each input converter is usually formed by an acoustic-electric transducer, such as a microphone. Usually, an electroacoustic transducer, such as a microspeaker, or an electromechanical transducer, such as a bone conduction earpiece, is used as the output converter. The signal processing device is usually implemented by an electronic circuit implemented on a printed circuit board and independently of this has a determination unit and usually also an amplifier or an amplifier unit. In addition, each hearing device according to the invention is usually equipped with a transmitting and / or receiving unit, which enables wireless communication with other electronic devices, in particular with other hearing devices according to the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Embodiments of the present invention are described in detail below with the aid of schematic drawings. In the drawings:

[0049] Figure 1 A block diagram showing a binaural hearing device system having a left hearing device and a right hearing device;

[0050] Figure 2 A block diagram showing a signal processing device with a determination unit for a left hearing device;

[0051] Figure 3 A block diagram showing a determination unit having a front stage with four channels;

[0052] Figure 4 A block diagram showing one of four channels having two measurement modules;

[0053] Figure 5 A block diagram showing one of the two measurement modules having a filter; and

[0054] Figure 6 Shows a top view of the test arrangement.

[0055] Parts that correspond to one another are provided with the same reference symbols in all the figures. DETAILED DESCRIPTION

[0056] The following example describes and Figure 1 The hearing device system 2 schematically shown in the figure consists of two hearing devices 4 and is preferably designed as a binaural hearing device system 2. Usually, one of the two hearing devices 4 is designed for the left ear and the other is designed for the right ear, and in the present embodiment, each of the two hearing devices has two input converters 6, a signal processing device 8, an output converter 10 and an antenna unit 12.

[0057] In the present embodiment, each input transducer 6 is formed by a microphone. By way of example, a micro-speaker is used as output transducer 10. The signal processing device 8 is preferably respectively realized by an electronic circuit realized on a printed circuit board. The antenna unit 12 is used for wireless communication and in particular for data exchange between two hearing devices 4.

[0058] In the present embodiment, the signal processing devices 8 of the two hearing instruments 4 are designed substantially identically. A block diagram of one of the signal processing devices 8 is shown in FIG. Figure 2 The signal processing device has a plurality of units, namely a transmitting and receiving unit 14, two analog-to-digital converters 16, a main unit 18, a digital-to-analog converter 20 and a determination unit 22. At least one of these units, namely the determination unit 22, also has a plurality of subunits.

[0059] Each of these units and each of these subunits and in particular each of the subunits mentioned below is formed, depending on the application, by its own electronic circuit, which only forms the corresponding unit or subunit, or by its own program modules / software modules. In particular, a typical design of the signal processing device 8 is that some units or subunits are formed by their own electronic circuits and some units or subunits are formed by their own program modules.

[0060] Independently of this, the determination unit 22 is designed and arranged for so-called own speech determination in at least one operating mode of the hearing device 4. The respective hearing device 4 is therefore arranged to automatically detect when the voice of the wearer of the hearing device 4 is active and when it is inactive in the at least one operating mode. In the at least one operating mode, the determination unit 22 outputs a determination result, which is preferably a simple binary signal, which indicates whether the own speech, i.e. the voice of the wearer of the hearing device 4, is currently active.

[0061] The determination result is based on Figure 2 In the embodiment of the present invention, the output signal of the output transducer 10 of the hearing device 4 is sent to the main unit 18, based on the determination result or according to the determination result, the output signal of the output transducer 10 of the hearing device 4 is then usually generated by the main unit 18. Then, depending on whether the determination result indicates that the wearer's own voice is currently active, different signal processing processes or signal processing programs are used to generate the output signal.

[0062] Independently of the signal processing process being used, the output signal is generated in such a way that the main function of the hearing instrument 4, namely amplifying the sound signal, is fulfilled thereby. This is achieved according to known principles by processing the input signals generated by the input converter 6. These input signals are digitized in the analog-to-digital converter 16 and fed as digital input signals to the main unit 18. Here, the digital input signals are processed by the above-mentioned different signal processing processes, so that digital output signals are generated. These digital output signals are subsequently converted in the digital-to-analog converter and finally fed to the output converter 10.

[0063] In addition, if Figure 2 As shown, the digital input signal is also fed to the determination unit 22 at the same time. Here, the digital input signal is used for the above-mentioned own speech determination. The own speech determination is realized by means of two subunits of the determination unit 22, namely by a neural network 24, i.e. an artificial neural network, and by a pre-stage 26 connected upstream of the neural network 24. The pre-stage 26 is usually designed as a multi-frequency channel pre-stage.

[0064] This means that the two digital input signals are fed to a filter bank (not shown) of the signal processing device 8 and are respectively decomposed into partial signals. The corresponding filter bank is designed in such a way that two or more frequency ranges are predetermined by the filter bank and are decomposed into partial signals at least approximately in such a way that each partial signal has all the frequency components of the digital input signal that can be assigned to exactly one frequency range in the frequency ranges. Figure 3 As shown, the multi-frequency channel front stage is exemplarily designed as a four-frequency channel front stage.

[0065] Corresponding filter banks are known in principle, and hearing devices with such filter banks are generally referred to as multi-frequency channel hearing devices. In the present embodiment, the hearing devices 4 of the hearing device system 2 are designed as four-frequency channel hearing devices and each have a four-frequency channel front stage. Therefore, the spectrum is divided into four frequency ranges, i.e., frequency channels. Therefore, the two digital input signals are sent to the determination unit 22 after being decomposed into partial signals. Subsequently, these partial signals are available to the front stage 26 and are further processed here according to the frequency channels.

[0066] The further processing according to the frequency channels is carried out by means of a subunit of the upstream stage 26, namely a channel module 28, wherein a channel module 28 is implemented for each frequency channel k of k=1 . . . 4, namely for each frequency range k. Such a channel module 28 is Figure 4 It is shown as an example in FIG.

[0067] Each channel module 28 has a first submodule A as a first subunit and a second submodule B as a second subunit. Each submodule A also has a filter a as a main component, which corresponds to the first filter type. Each submodule B has a filter b as a main component, which corresponds to the second filter type. Therefore, the pre-stage 26 has four filters of filter type a and four filters of filter type b in total.

[0068] The filter a of the first filter type and the filter b of the second filter type are filters for spatial separation. Each of the two filter types is designed so that the common directional characteristics of the two microphones of the hearing device 4 are respectively realized by these filter types, that is, the first directional characteristic is realized by the first filter type and the second directional characteristic is realized by the second filter type.

[0069] The two directional characteristics differ from one another in particular with regard to a predetermined spatial direction, along which the sensitivity of the respective directional characteristic reaches its maximum value. Figure 6 As shown, the two corresponding spatial directions are advantageously preset relative to a frontal direction of 0°, towards which the face of the hearing device wearer is usually oriented. In this embodiment, a spatial direction of 0° is preset for the first filter type and a spatial direction of 180° is preset for the second filter type.

[0070] Preferably, the two filter types are also designed as static filters with respectively fixedly predefined filter coefficients. The filter coefficients are preferably selected and predefined by the manufacturer.

[0071] In the thus designed pre-stage 26, two partial signals, namely the kth partial signal of one digital input signal and the kth partial signal of another digital input signal, are respectively fed in the frequency channel k to two submodules A and B of the kth channel module 28. Subsequently, in the submodule A of the kth channel module 28 and the submodule B of the kth channel module 28, respectively, a filtered signal is generated and a real-valued attenuation variable is determined from the filtered signal.

[0072] A total of eight real-valued damping variables are therefore determined, and each damping variable is subsequently transmitted via its own information signal to the neural network 24. In the present exemplary embodiment, eight information signals are therefore fed to the neural network.

[0073] exist Figure 5The k-th submodule A is shown as an example in FIG. However, the subsequent description can be transferred similarly to each other submodule, that is, submodule A or submodule B. Figure 5 The kth sub-signal of a digital input signal and the kth sub-signal of another digital input signal (hereinafter referred to as the first sub-signal and the second sub-signal) are sent to the kth sub-module A.

[0074] The first partial signal is then preferably fed unprocessed to the filter a of the kth submodule A. On the other hand, the second partial signal is preferably fed processed to the filter a of the kth submodule A, wherein the second partial signal is processed for this purpose in the equalizer unit 30. The above-mentioned spatial separation is achieved by the filter a and the equalizer unit 30.

[0075] Then, based on the two partial signals, a filter signal is generated by filter a and fed to the measuring unit 32. In addition, the unprocessed first partial signal is fed to the measuring unit 32, and then one of the above-mentioned attenuation variables is finally determined in the measuring unit 32 based on the filter signal and the first partial signal.

[0076] The above embodiment of the k-th submodule A is only an expedient embodiment. In this embodiment, the function of the equalizer unit 30 is usually to multiply the incoming signal by a frequency-dependent complex-valued scalar weight. In an alternative embodiment, the filter a assumes this function and the equalizer unit 30 is omitted.

[0077] To determine the attenuation variable, for example, an auxiliary variable is first determined in the measuring unit 32, namely a first auxiliary variable based on the first partial signal and a second auxiliary variable based on the filtered signal, and the two auxiliary variables are then compared with one another. Such an auxiliary variable is, for example, power.

[0078] According to an alternative not shown in detail, a complex-valued attenuation variable is determined instead of a real-valued attenuation variable in the measuring unit 32. For this purpose, amplitude and phase information in the signals, ie the first partial signal and the filtered signal, is usually evaluated in the measuring unit 32.

[0079] Two implementation variants are provided for this alternative. In the case of a first implementation variant, there are a total of eight complex-valued damping variables, and each of these eight complex-valued damping variables is transmitted to the neural network 24 via its own information signal.

[0080] In the case of the second embodiment variant, a total of eight variables are present twice, and each of these sixteen variables is transmitted via its own information signal to the neural network 24. In this case, for example, the real part and the imaginary part or the amplitude and the phase are transmitted to the neural network 24 in separate information signals.

[0081] Furthermore, in some applications it may be expedient to feed one or more further information signals, ie information signals which contain further information, to the neural network 24 in addition to the above-mentioned information signals.

[0082] In the exemplary embodiment shown, however, for the sake of simplicity, exactly eight information signals are fed by way of example to the neural network 24, wherein eight real-valued attenuation variables are determined in the manner described above and wherein each attenuation variable is transmitted to the neural network 24 via its own information signal.

[0083] The neural network 24 has some artificial neurons, not shown in detail, which are arranged in one or more layers. The determined attenuation variable is then evaluated by these neurons, thereby obtaining an evaluation result. The evaluation result is finally output by the neural network 24. In the present embodiment, an evaluation signal having a probability value, that is, a value in the value range of 0 to 1, is output as the evaluation result.

[0084] The evaluation signal is then fed to a comparator unit (not shown in detail) of the final stage 34 of the determination unit 22 and compared there with a reference, i.e. a reference value. The reference value is fixedly predefined and is usually in the value range of 0.3 to 0.7 and in particular in the value range of 0.4 to 0.6, for example 0.5. As a result, a determination result is usually generated by the comparator unit, which indicates whether the own speech was determined by the determination unit 22.

[0085] In at least one operating mode of the hearing device 4, the above-described process for natural speech determination usually takes place permanently in the background and preferably applies to each of the two hearing devices 4 of the hearing device system 2. Therefore, the above-described determination results are respectively generated independently of one another in the two hearing devices 4. Based on the respective determination results, a respective output signal is then generated in the corresponding hearing device 4.

[0086] However, a preferred embodiment variant is that the above-mentioned comparator units of the two hearing devices 4 respectively generate only preliminary determination results. Further preferably, each hearing device 4 transmits its preliminary determination result to the corresponding other hearing device 4 by means of the transmitting and receiving unit 14 and the antenna unit 12 connected thereto. In each of the two hearing devices 4, the two preliminary determination results of the two determination units 22 of the two hearing devices 4 are further analyzed together. This usually takes place in a logic unit (not shown in detail) of the final stage 34. Thus, a determination result, i.e. a final determination result, is finally generated.

[0087] As described above, the two filter types a, b of the submodules A, B are preferably designed as static filters with fixedly preset filter coefficients. The filter coefficients are preferably preset by the manufacturer. Further preferably, the determination unit 22 of at least two hearing devices 4 of the hearing device system 2 is not personalized or individualized.

[0088] Furthermore, the corresponding filter coefficients are preferably determined in a coefficient determination method. This determination is preferably carried out without the involvement of a subsequent wearer, ie an end user. It is also expedient to use a hearing device model in the coefficient determination method to determine the filter coefficients. The hearing device model has at least two Figure 6 3. In the coefficient determination method, input signals are preferably generated in different tests by the test input converter 36. These input signals are suitably recorded and analyzed, for example, in a computer not shown in detail.

[0089] In these different tests, different scenarios with different test subjects are preferably simulated. It is further preferred that during the test, different test subjects act as test wearers 38 of the test device, and further preferably different test subjects act as useful signal sources 40 in the sense of conversation partners, wherein the useful signal sources 40 are preferably located in front of the test wearer 38, such as Figure 6 shown.

[0090] Then, preferably in each test and for each simulated scenario, suitable filter coefficients are determined by computer, in particular for the above-mentioned filter a and filter b. In this way, a data set with a large number of suitable values ​​is generated for each filter coefficient.

[0091] For example, the so-called k-means algorithm is applied to the corresponding data set in order to obtain the filter coefficients that are ultimately predefined for the filters a, b of the submodules A, B.

[0092] Reference numerals list

[0093] 2Hearing equipment system

[0094] 4. Hearing equipment

[0095] 6 Input Converter

[0096] 8Signal processing device

[0097] 10 output converter

[0098] 12 antenna units

[0099] 14 transmitting and receiving units

[0100] 16 ADC

[0101] 18 Main Units

[0102] 20 DAC

[0103] 22 Determine the unit

[0104] 24 Neural Networks

[0105] 26 Pre-stage

[0106] 28 channel module

[0107] 30 Equalizer Units

[0108] 32 measuring units

[0109] 34 Final Level

[0110] 36 Test Input Converter

[0111] 38 test wearers

[0112] 40 Useful signal sources

[0113] A submodule

[0114] Submodule B

[0115] aFilter

[0116] bFilter

Claims

1. A method for determining the own speech activity of a wearer of a hearing device (4) by means of a signal processing device (8) of a hearing device (4), wherein: - generating a first input signal by a first input transducer (6) of the hearing device (4), and generating a second input signal by a second input transducer (6) of the hearing device (4), - feeding the two input signals to a determination unit (22) of the signal processing device (8), the determination unit having a neural network (24) and a preceding stage (26) connected upstream of the neural network (24), - generating an information signal based on the two input signals by the preceding stage (26), - evaluating the information signal by means of the neural network (24), and The determination unit (22) outputs a determination result based on the evaluation of the information signal by the neural network (24).

2. The method according to claim 1, wherein: The personalization of the determination unit (22) is omitted.

3. The method according to claim 1 or 2, wherein: The neural network (24) is trained with the aid of a plurality of recorded acoustic signals in which the own speech of different test wearers is recorded.

4. The method according to any one of claims 1 to 3, wherein: The front stage (26) has two filters (a, b), namely a first filter (a) and a second filter (b), wherein the first filter (a) corresponds to a first filter type and the second filter (b) corresponds to a second filter type, and wherein a first filtered signal is generated by the first filter (a) and a second filtered signal is generated by the second filter (b).

5. The method according to claim 4, wherein: The first filter type is designed to extract a useful signal from a useful signal source, and the second filter type is designed to mask out the own speech.

6. The method according to claim 4 or 5, wherein: These two filter types are designed as static filters (a, b) with fixed preset filter coefficients.

7. The method according to claim 6, wherein: The filter coefficients for the second filter type are predetermined on the basis of an analysis of recorded acoustic signals in which the own speech of different test wearers was recorded.

8. The method according to claim 7, wherein: At least one adaptive test filter is used for the analysis.

9. The method according to any one of claims 4 to 8, wherein: Each filtered signal is associated with a reference signal in order to generate an information signal, the reference signal being based on at least one of the two input signals.

10. The method according to claim 9, wherein: By correlating one of the filtered signals with one of a plurality of reference signals, an attenuation variable is determined, which is transmitted to the neural network (24) via a corresponding information signal.

11. The method according to any one of claims 1 to 10, wherein: The information signal is evaluated by the neural network (24), wherein a prediction value is output by the neural network (24) as the evaluation.

12. The method according to claim 11, wherein: The information signal is evaluated by a neural network (24) designed as a recurrent neural network (24), wherein a prediction value is output by the neural network (24) as an evaluation.

13. The method according to any one of claims 1 to 12, wherein: A preliminary determination result is obtained by the determination unit (22), wherein the preliminary determination result is analyzed in the determination unit (22) together with a further preliminary determination result, the further preliminary determination result being obtained in a further hearing device (4) by the further determination unit (22) and transmitted via the further hearing device (4), and wherein the determination result is obtained in the determination unit (22) by analyzing the two preliminary determination results.

14. A hearing instrument (4) which is configured to carry out at least one method step of the method according to any of the preceding claims in at least one operating mode.

15. A hearing device system (2) comprising two hearing devices (4) according to claim 14.

Images