Hearing system and its operation
By combining a deep convolutional neural network speech enhancement model and noise embedding signals into hearing instruments, the problem of insufficient adaptability of hearing instruments in different acoustic environments is solved, achieving more efficient speech enhancement and energy saving, and improving the user experience.
Patent Information
- Application Number
- CN202480091091.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-09-25
- Filing Date
- 2024-10-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing hearing aids struggle to adapt dynamically to various acoustic environments, resulting in a lack of clarity for users in noisy environments and a delayed response to sudden noise changes.
A deep convolutional neural network speech enhancement model (SEM) is used in conjunction with noise-embedded signals. Through the collaborative work of a signal processor and external devices, it dynamically adapts to different acoustic environments, suppresses noise components, and preserves speech components.
It improves the adaptability of hearing instruments in various acoustic environments and speech clarity, reduces computational complexity and energy consumption, and enhances the user experience.
Smart Images

Figure CN123125901A_ABST
Abstract
Description
[0001] This invention relates to a hearing system comprising at least one hearing instrument. The invention also relates to a method of operating the hearing system.
[0002] Generally, a hearing instrument is an electronic device designed to support the hearing of the person wearing it (referred to as the "user" or "wearer" of the hearing instrument). In particular, the present invention relates to hearing instruments specifically configured to at least partially compensate for hearing loss in users with hearing impairment. Such hearing instruments are also referred to as "hearing aids." In addition to the aforementioned hearing aids, there are also hearing instruments designed to support the hearing of users with normal hearing (i.e., people without hearing loss). Such hearing instruments are sometimes referred to as "personal sound amplification products" (PSAPs), for example, to enhance the wearer's hearing in complex acoustic environments or to protect the wearer's hearing from damage or excessive stress.
[0003] Hearing instruments, especially hearing aids, are typically designed to be worn in or over a user's ear, such as behind-the-ear (BTE) or in-the-ear (ITE) devices. Internally, a hearing instrument generally includes at least one (acoustic-electrical) input transducer, a signal processor, and an output transducer. During operation, at least one input transducer captures sound from the environment surrounding the hearing instrument and converts it into an input audio signal (i.e., an electrical signal that transmits sound information). In the signal processor, the input audio signal is processed, particularly amplified according to frequency, for example, to at least partially compensate for the user's hearing loss. The signal processor outputs the processed signal (also called the output audio signal) to the output transducer. In most cases, the output transducer is an electroacoustic transducer (also called a "receiver") that converts the output audio signal into processed airborne sound, which is emitted into the user's ear canal. In BTE devices, the receiver may be located in a housing worn behind the user's ear. In this case, the sound output by the receiver is guided into the ear canal through a hollow sound tube connecting the tip of the housing to an earpiece to be inserted into the ear canal. In another type of BTE device called an in-ear receiver (RIC) device, the receiver is located in the earpiece and therefore outside the housing. Alternatively, the output transducer can be an electromechanical transducer that converts the output audio signal into structured sound (vibration), which is transmitted, for example, to the user's skull.
[0004] In addition to classic hearing aids, there are implantable hearing devices, such as cochlear implants, and hearing devices that use output transducers to directly stimulate the user's auditory nerve.
[0005] The term "hearing system" refers to a device or component of a device and / or other structure that provides the functions required for the operation of a hearing instrument. A hearing system can consist of a single, independent hearing instrument. Alternatively, a hearing system may include a hearing instrument and at least one additional electronic device, which may be, for example, another hearing instrument for the user's other ear, a remote control, and a programming tool for the hearing instrument. Furthermore, modern hearing systems typically include at least one hearing instrument and software applications for controlling and / or programming that at least one hearing instrument. These software applications (hereinafter referred to as "hearing applications") are installed or can be installed on a computer, particularly a mobile communication device such as a mobile phone (smartphone). In the latter case, the computer is typically not part of the hearing system but is used by the hearing system only as a resource for data storage, numerical computation capabilities, and communication services. In most cases, the computer (especially a mobile communication device) on which the hearing application is installed or can be installed will be manufactured and sold independently of the hearing system.
[0006] Generally speaking, hearing aids enhance a user's auditory perception and quality of life, especially in challenging noisy environments. This is particularly important for users with hearing loss, as this impairment not only affects their ability to communicate effectively but also isolates them from their social environment. Among other things, modern hearing aids can distinguish between speech and noise, enhancing the former while suppressing the latter, thus allowing users to participate in conversations even in busy environments such as restaurants or markets.
[0007] Speech enhancement (SE) is one of the core functions of modern hearing instruments. Real-time methods for providing SE, as disclosed in references [1] and [2], have traditionally relied on statistical models. However, the situation has changed significantly in contemporary times, with data-driven, deep learning-based models trained on large, available datasets showing great promise in SE tasks. In particular, convolutional neural networks (CNNs)... n N) Use masking and direct feature mapping techniques to enhance speech. A state-of-the-art solution known as “Deepfilternet” is proposed in reference [3], which is famous for its excellent noise suppression and speech intelligibility.
[0008] It has been recognized that the deeper C n The N-architecture produces improved results. However, their computational complexity can challenge real-time processing capabilities, especially in small devices such as hearing aids. Furthermore, they often struggle to dynamically adapt to diverse real-world acoustic environments. This leads to significant gaps in user satisfaction and usability. For example, users report a lack of clarity in noisy environments and delayed response to sudden changes in noise.
[0009] The technical problem to be solved by the present invention is to provide a solution for speech enhancement in hearing instruments, which has improved adaptability and performance in various acoustic environments.
[0010] According to a first aspect of the invention, the technical problem is solved by a method for operating a hearing system as defined in claim 1. According to a second aspect of the invention, the technical problem is also solved by a hearing system as defined in claim 13. Preferred embodiments of the invention are described in the dependent claims and the following description.
[0011] In general, the present invention relates to a hearing system and a method for operating the hearing system, the hearing system comprising at least one hearing instrument. As described above, the hearing system may consist of a single, independent hearing instrument, or may include at least one additional device or application besides the at least one hearing instrument. In particular, the hearing system may be a binaural hearing system comprising two hearing instruments for both of the user's ears.
[0012] The system includes at least one hearing instrument comprising at least one (first) input transducer configured to capture ambient sound from the environment surrounding the hearing instrument. Optionally, in addition to the at least one (first) input transducer located within the at least one hearing instrument, the hearing system may include or access at least one (second) input transducer from an external device, i.e., at least one (second) input transducer located outside the at least one hearing instrument. The external device may be part of the hearing system (e.g., a dedicated remote control for the hearing instrument). Alternatively, the second input transducer and the external device may not be part of the hearing system. For example, the at least one (second) input transducer may be, for example, a microphone on a smartphone with a hearing application installed.
[0013] At least one hearing instrument in the system includes a signal processor configured to process the captured sound signal to generate a processed sound signal; and an output transducer configured to output the processed sound signal to the user of the hearing system. Preferably, at least one hearing instrument is designed to be worn on or in the user's ear, for example as a BTE (especially RIC) device or as an ITE device or any other design described above. Preferably, the output transducer of at least one hearing instrument is a receiver (i.e., an electroacoustic transducer). Alternatively, within the scope of the invention, the output transducer may be an electromechanical transducer or an output transducer that directly stimulates the user's auditory nerve.
[0014] According to the method for operating a hearing system, in a first sound capture step, ambient sound from at least one (first) input transducer of the hearing instrument is captured by at least one input transducer of the hearing instrument to generate a first input audio signal, wherein the ambient sound includes a speech component and a noise component. Here, the speech component is not necessarily always included in the ambient sound. Instead, the speech component may exist only in certain speech intervals interrupted by non-speech intervals (where the ambient sound does not contain a speech component). Furthermore, the speech component may be defined as belonging to a dominant speech (i.e., a sound with a certain loudness and / or originating from the vicinity of the hearing instrument) or a specific speech (i.e., the speech of a particular individual speaker). In these cases, the noise component may include the speech of one or more other speakers, such as the cacophony of many distant sounds in a party environment. Finally, the ambient sound may consist of a speech component and a noise component (i.e., only one speech component and one noise component), or it may include at least one additional component, a musical component, and / or another speech component belonging to another speech.
[0015] The method also includes a signal processing step, wherein the first input audio signal is modified to generate a processed audio signal. The entire signal processing step can be performed by a signal processor of at least one hearing instrument. Alternatively, only a portion of the signal processing step may be performed by a signal processor, while another portion may be performed externally, for example, by a functional part of the hearing application. In a binaural version of the hearing system, the signal processors of the two hearing instruments may perform the signal processing step together in a collaborative manner.
[0016] Finally, in the output step, the processed signal is output as perceptible sound by the output transducer of at least one hearing instrument to the user of at least one hearing instrument.
[0017] The signal processing steps also include a speech enhancement step (SE step). In the SE step, a speech enhancement model (SEM) is applied to the first input audio signal to produce an enhanced audio signal, wherein noise components of ambient sounds are suppressed while speech components are preserved. Preferably, the SEM comprises a (deep) neural network, particularly a convolutional neural network. More preferably, the SEM is implemented in at least one hearing instrument, particularly as a (hardware and / or software) component of a signal processor. Optionally, in the case of a binaural version of the hearing system, the SEM may be implemented in both hearing instruments of the system. Alternatively, within the scope of the invention, only one of the two hearing instruments of a binaural hearing system may be equipped with a SEM. In a less preferred alternative, but still within the scope of the invention, the SEM may be implemented outside at least one hearing instrument, for example, as part of the system's hearing application.
[0018] SEM can be applied directly to a first input audio signal generated by at least one first input audio transducer. Alternatively, SEM can be applied indirectly to the first input audio signal; that is, it can be applied to an audio signal derived from the first input audio signal through preprocessing.
[0019] In embodiments of the invention, the captured audio signal can be fed to the SEM as a (raw or preprocessed) time-domain signal. In alternative embodiments, features such as time-frequency representations can be extracted from the captured audio signal, and these features are then fed to the SEM as input signals.
[0020] Before or during normal operation of the hearing system, a machine learning algorithm is used to train the SEM to suppress noise components. Preferably, the training is performed offline, i.e., not on the hearing system itself. In other words, preferably, the SEM is provided to the hearing system in a pre-trained state.
[0021] According to the present invention, an embedded signal including information characterizing noise components is provided to the SEM to configure the latter for suppressing noise components.
[0022] The use of such embedded signals (also known as “embedding” or “fingerprint”) is known in itself, for example from reference [4].
[0023] Generally, an "embedding" is an auxiliary input to a model that provides additional information to support the model in performing its task. In this case, the embedding is a "noise embedding" or "noise fingerprint," which defines the noise that the SEM should remove from its input signal. Noise embeddings have been developed for applications such as long-distance voice transmission (e.g., telephone, video conferencing, or radio applications) or audio recording, where embeddings can benefit from the high numerical computational power of fixed computer systems. Embeddings have not yet been used in hearing systems for SE. Known SEMs using embeddings are unsuitable for implementation in hearing systems due to their large numerical requirements, as they typically add additional complexity to already complex SE solutions. The inventors still propose implementing SEMs using noise embeddings in hearing systems because they have found that using noise embeddings allows for a reduction in the complexity of the SEM compared to SEMs without necessarily accepting poorer SE performance. In other words, the inventors have shown that the additional complexity introduced by providing embeddings can be compensated for by reducing the complexity of the SEM without sacrificing SE performance, thus allowing the use of noise embeddings in hearing systems. However, compared to conventional SEMs that do not use embedding, the solution according to the present invention has the advantage of improved adaptability to different acoustic environments.
[0024] Within the scope of this invention, the embedded signal can be provided in different ways and by using different hardware platforms.
[0025] In a first preferred embodiment of the invention, in the embedding selection step, the embedding signal to be provided to the SEM is selected from a plurality of pre-generated embedding signals. This selection is based on acoustic scene detection, i.e., using a sound classifier to analyze the noise components in ambient sound and classifying the noise components into one of several predetermined noise categories based on this analysis. The selection of the embedding signal to be provided to the SEM is based on this classification. Specifically, a pre-generated embedding signal is stored for each predefined noise category, and the pre-generated embedding signal belonging to the sound category to which the noise component is classified is selected and used to configure the SEM. The sound classifier can be applied to a first input audio signal, i.e., ambient sound captured by at least one input transducer of at least one hearing instrument. Alternatively, the sound classifier can be applied to a second input audio signal, wherein the ambient sound is captured by a second input transducer that does not belong to at least one hearing instrument. For example, a hearing application can use the microphone of a smartphone on which the application is installed to capture ambient sound in the user's environment, thereby generating the second input audio signal. The advantage of this embodiment is that it is easy to implement in a hearing system and requires very few numerical resources during the operation of the hearing system. Within the scope of this invention, the pre-generated embedding can be stored in at least one hearing instrument or external to the hearing instrument, such as in the system's hearing application or in a data cloud. Preferably, the embedding selection step is repeated whenever the classification of the acoustic scene changes, i.e., whenever the sound classifier changes the noise category to which the noise component of the ambient sound is classified.
[0026] In a second preferred embodiment of the invention, the embedded signal is generated "online" (i.e., during operation of the hearing instrument) in the embedding generation step. Here, an embedding generation model (EGM) is applied to either the first or second input audio signal (as described above). The EGM analyzes the noise components of the ambient sound to generate the embedded signal. Preferably, similar to SEM, the EGM includes a (deep) neural network, particularly a convolutional neural network. Within the scope of the invention, the EGM can be implemented in at least one hearing instrument, particularly as a (hardware and / or software) component of a signal processor. However, in a preferred implementation of the method, the EGM is implemented as part of a hearing application and is therefore configured to run on an external computer, particularly a user's smartphone.
[0027] Therefore, in a variant of the second preferred embodiment, the embedded signal can be generated on at least one hearing instrument itself. The advantage of this embodiment is that it is suitable for the independent operation of a single hearing instrument. In other words, it does not require the hearing instrument to be connected to another device or structure.
[0028] On the other hand, the advantage of generating the embedding outside at least one hearing instrument is that it is not constrained by the limited numerical capabilities and battery power of the hearing instrument. Instead, the greater numerical capabilities and battery power of an external device (such as a user's smartphone) can be used to generate the embedded signal.
[0029] Preferably, in a further variant of the first and second preferred embodiments, the embedded signal is selected or generated only in non-speech regions where the ambient sound does not contain speech components. To this end, in the speech activity detection step, a speech activity detector (VAD) is applied to either the first or second input audio signal (as described above). The VAD analyzes the ambient sound to detect the non-speech regions. Based on the output of the VAD, the embedding selection step or the embedding generation step is performed only during the detected non-speech regions. In other words, the embedding selection step or the embedding generation step is not performed during speech regions (i.e., when the VAD detects speech components in the ambient sound and as long as it does). In summary, the EGM operates only when the background noise to be suppressed can be directly captured from the environment.
[0030] Since the background noise typically does not change rapidly, the EGM can operate intermittently with large interruptions (pauses) compared to the sampling rate required for normal hearing aid signal processing. Therefore, in a preferred embodiment of the method, the SEM updates the enhanced audio signal at a first sampling rate (i.e., more precisely, the SEM generates a time series of values for the enhanced audio signal, where subsequent values of the enhanced audio signal are generated at the first sampling rate), and the EGM updates the embedded signal at a second sampling rate (i.e., more precisely, the EGM generates a time series of values for the embedded signal, where subsequent values of the embedded signal are generated at the second sampling rate), wherein the first sampling rate is at least 10... 2 10 preferred 3 Especially 10 4 The coefficient exceeds the second sampling rate (embedding generation rate). For example, EGM runs once per second or longer intervals, while SEM produces output values at the sampling rate of the hearing instrument, such as 48 kHz.
[0031] Preferably, the SEM includes an encoder portion that reduces a first input audio signal or an audio signal derived therefrom to a first intermediate signal with lower temporal and / or spectral resolution (compared to the input signal of the SEM); and at least one decoder portion that expands the first intermediate signal to a second intermediate signal with higher temporal and / or spectral resolution (compared to the first intermediate signal). In a preferred implementation of the method, the first input audio signal or the audio signal derived therefrom is reduced to the first intermediate signal by the encoder portion. The first intermediate signal is then expanded to the second intermediate signal by at least one decoder portion.
[0032] Therefore, preferably, the SEM forms a so-called “bottleneck” represented by a low-dimensional first intermediate signal. The bottleneck structure, along with the training of the encoder portion and at least one decoder portion, effectively forces the SEM to discard unwanted noise components of the first input signal. Such speech enhancement models are known in themselves, for example from references [3] and [5], the contents of which are incorporated herein by reference. Preferably, such a SEM, supplemented only by embedding, is used in the method and system according to the invention.
[0033] In an advantageous embodiment of the method, during its speech enhancement step, the embedded signal is applied to the “bottleneck” of the SEM, i.e., to the first intermediate signal, specifically by adding the embedded signal to the first intermediate signal (especially by subtraction, which is considered a special form of addition). Accordingly, in this embodiment, the dimension of the embedded signal is adapted to the dimension of the first intermediate signal. Thus, the dimension of the embedded signal is lower than the dimension of the input signal of the SEM. For example, the input signal of the SEM may be a 48kHz speech signal of several seconds in length, while the embedding is a vector of a fixed length, for example, 128, which is constant for the entire embedded signal (but may of course be different for different embedded signals). In an alternative implementation, the first intermediate signal and the embedded signal are combined using a multi-head attention block (as described in, for example, [6]). Preferably, the multi-head attention block has four attention heads, wherein the first intermediate signal is used as a query and the embedded signal is used as both the key and value of the attention mechanism. The output of the multi-head attention block is then added to the first intermediate signal, and the result is used as the input of at least one decoder.
[0034] In an advantageous embodiment of the method and system according to the invention, the encoder portions of the EGM and SEM are implemented as the same type of neural network, differing only in the number of neurons and / or layers. Here, unlike the conventional SEM implementation using embeddings, the EGM is formed to have higher complexity compared to the encoder portion of the SEM because the EGM has more neurons and / or more layers than the encoder portion of the SEM.
[0035] Generally, a hearing system according to a second aspect of the invention is configured to automatically execute the method according to a first aspect of the invention. For this purpose, the hearing system further includes a sound-embedded imager (SEM) (as described above) and means for providing an embedded signal to the SEM, the embedded signal configuring the latter to suppress noise components in ambient sound.
[0036] Since the hearing system according to the second aspect of the invention is configured to automatically execute the method according to the first aspect of the invention, each embodiment or variation of the method corresponds to an embodiment or variation of the hearing system. Therefore, any disclosure related to the method is also applicable to the hearing system with the necessary modifications, and vice versa.
[0037] Preferably, the signal processor of at least one hearing instrument is designed as a digital electronic device. It can be a single unit or composed of multiple subprocessors. According to the invention, at least one of the signal processors or said subprocessors can be a programmable device (e.g., a microcontroller). In this case, the above-described functions or a portion thereof are implemented as software (especially firmware), which is stored in the memory of the signal processor in an executable form. Alternatively, at least one of the signal processors or said subprocessors can be a non-programmable device (e.g., an ASIC). In this case, the above-described functions or a portion thereof are implemented as hardware circuitry.
[0038] In particular, both the SEM and the means for providing the embedded signal can be implemented by software or hardware (i.e., non-programmable electronic circuits such as ASICs) or a combination of both. Within the scope of the invention, the SEM and the means for providing the embedded signal can be implemented in at least one hearing instrument, outside at least one hearing instrument, or distributed to external parts of at least one hearing instrument and hearing system (such as hearing applications).
[0039] As can be seen from the above, the main effect of this invention is to save the computational power and energy consumption required to run the SEM in a hearing system by providing noise embedding to the SEM. Noise embedding supports the operation of the SEM, thus allowing for a reduction in the complexity of the SEM without sacrificing SE performance, which in turn allows for a reduction in the computational power and energy consumption required to run the SEM. In particular, in embodiments where the embedded signal is generated during the operation of the hearing system, the computational power and energy consumption required to run the SEM can be saved by designing the EGM to have greater complexity compared to the encoder portion of the SEM.
[0040] The numerous features of the various embodiments of the present invention described above further support this main effect by allowing embedded signals to be provided in a manner that uses little or no digital capacity and power supply of at least hearing instruments. This effect is achieved in the following way: The embedded signal is selected from the pre-generated embedded signal (instead of generating the embedded signal in the hearing instrument during its operation). The embedded signal is generated outside the hearing instrument, that is, in the external part of the hearing system, such as by the hearing application (rather than in the hearing instrument, during its operation). The embedded signal is generated at an embedding generation rate much lower than the SEM sampling rate, and / or Embedded signals are generated only in the detected speech regions.
[0041] Unless otherwise stated, the above teachings may be applied individually or in combination.
[0042] Exemplary embodiments of the invention will then be described in more detail with reference to the accompanying drawings, in which: Figure 1 A schematic diagram of a first exemplary embodiment of a hearing system is shown, the hearing system including a hearing instrument to be worn on a user's ear and a hearing application installed on the user's smartphone, wherein a speech enhancement model (SEM) and means for providing embedded signals to the SEM are implemented in the hearing instrument as part of its signal processor; Figure 2 A schematic diagram of a second exemplary embodiment of a hearing system (similar to) is shown. Figure 1 The SEM is implemented in a hearing instrument, wherein the device for providing embedded signals to the SEM is implemented as a functional part of the hearing application; and Figure 3 and Figure 4 It shows Figure 1 or Figure 2 Block diagrams of different embodiments of the functional structure of the hearing system.
[0043] In the accompanying drawings, unless otherwise stated, the same reference numerals always denote the same parts, structures, and elements.
[0044] Figure 1 A hearing system 2 is shown, which includes a hearing instrument 4 configured to be worn in or on the ear of a user 5. Figure 3 , 4 Preferably, the hearing instrument 4 is a hearing aid, that is, a hearing instrument configured to support the hearing of a user with hearing loss. For example... Figure 1 As shown, by way of example, the hearing device 4 can be designed as a behind-the-ear (BTE) hearing device. However, in other embodiments of the invention, the hearing device 4 can be designed as any other configuration described above or known in the art; for example, as a RIC or ITE device. Optionally, the system 2 includes a second hearing device (not shown) to be worn in or on the user's other ear to provide binaural support to the user 5.
[0045] The hearing instrument 4 includes two microphones 8 as (first) input transducers and a receiver 10 as an output transducer within the housing 6. The hearing instrument 4 also includes a battery 12 and a signal processor 14. Preferably, the signal processor 14 includes both programmable subunits (such as microprocessors) and non-programmable subunits (such as ASICs).
[0046] The signal processor 14 is powered by the battery 12, that is, the battery 12 provides the power supply voltage U to the signal processor 14.
[0047] During normal operation of the hearing instrument 4, the microphone 8 captures ambient airborne sound from the environment surrounding the hearing instrument 4. The microphone 8 converts the airborne sound into a (first) input audio signal I (also referred to as the "captured sound signal"), which is an electrical signal containing the captured sound information. The first input audio signal I is fed to the signal processor 14. The signal processor 14 processes the first input audio signal I, for example, to provide directional sound information (beamforming), perform noise reduction and dynamic compression, and individually amplify different spectral portions of the first input audio signal I based on the user 5's audiogram data to compensate for the user's specific hearing loss. The signal processor 14 transmits a processed audio signal O, which is an electrical signal containing processed sound information, to the receiver 10. The receiver 10 converts the processed audio signal O into processed airborne sound, which is transmitted into the user 5's ear canal through the sound channel 16 connecting the receiver 10 to the tip 18 of the housing 6 and a flexible sound tube (not shown) connecting the tip 18 to an earpiece inserted into the user 5's ear canal.
[0048] In addition to at least one hearing instrument 4, the hearing system 2 also includes a software application (hereinafter referred to as "hearing application" 20) installed on the user 5's mobile phone 22. Here, the mobile phone 22 is not part of the hearing system 2. Instead, it is used by the hearing system 2 only as an external resource to provide computing power, data storage (memory), and communication services.
[0049] Hearing instrument 4 and hearing application 20 exchange data via wireless link 24, for example, based on the Bluetooth standard. For this purpose, hearing application 20 accesses the wireless transceiver (not shown) of mobile phone 22, particularly a Bluetooth transceiver, to send data to and receive data from hearing instrument 4. Hearing application 20 also includes functions for remotely controlling, configuring, and updating hearing instrument 4.
[0050] The hearing system 2 also includes a speech enhancement model (SEM26) and a device 28 for providing an embedded signal E to the SEM26 (described in further detail below). Figure 1 In the example, both SEM26 and device 28 are implemented as software components and are installed in the signal processor 14 of hearing instrument 4 in an executable form.
[0051] Figure 2 Alternative arrangements of hearing system 2 are shown, unless otherwise stated, corresponding to Figure 1 The illustrated embodiment. However, in Figure 2 In the embodiments, with Figure 1 Unlike other embodiments, device 28 is implemented as part of hearing application 20. In this case, device 28 provides an embedded signal E to SEM 26 via wireless link 24.
[0052] Figure 3A schematic simplified sketch of the functional structure of one embodiment of the hearing system 2 is shown. As shown, the microphone 8 of the hearing instrument 4 captures ambient sound from the user 5's environment, which at least in certain time speech intervals contains a speech component S and a noise component N. The noise component N may be traffic noise (such as...). Figure 3 (Illustrated in the diagram) or any other sound unrelated to the dominant speech that user 5 assumes to be listening to. In particular, the noise component N may contain noise from multiple distant sounds.
[0053] During signal processing performed by the hearing instrument 4, an analysis filter bank 30 or a similar frequency analysis unit is applied to the first input audio signal I of the microphone 8 of the hearing instrument 4, directly or via one or more preprocessing units 32, which, for example, apply amplification and / or beamforming to the first input audio signal I. In its original or preprocessed form, the first input audio signal I is a variable signal dependent on time t: I = I(t).
[0054] The analysis filter bank 30 converts the (raw or preprocessed) first input audio signal I into a digital time-frequency representation X. The time-frequency representation X of the first input audio signal I can be considered as a discrete mathematical function that depends on a discrete time variable (called a time frame) and a discrete frequency variable (called a frequency band f). In other words, any value of the time-frequency representation X belongs to a specific time frame k and a specific frequency band f: X → X(k,f).
[0055] SEM26 is implemented as a deep neural network. Unless otherwise stated, in a preferred implementation, the structure of SEM26 corresponds to the structure of the speech enhancement algorithm DEEPFILTERNET 2 disclosed in [5], the contents of which are incorporated herein by reference. In other words, DEEPFILTERNET 2 is used as the basic structure of SEM26. Therefore, as suggested in [5], the time-frequency representation X is fed in parallel to analyzers 34 and 36 of SEM26. Here, analyzer 34 extracts the signal X containing ERB (Equivalent Rectangular Bandwidth) features from the time-frequency representation X. erb In other words, it maps the time-frequency representation X to the ERB frequency scale: X erb = X erb (k, f erb ), where f erb The ERB frequency band is represented, thus the speech envelope is recovered. Analyzer 36 extracts the signal X containing complex features from the time-frequency representation X. df Signal X df It consists of time frame k and discrete frequency variable (hereinafter referred to as "bandwidth") f. df Function: X df = X df (k, f df Signal Xdf This indicates the periodicity of the speech component S. In the preferred implementation, the ERB band f erb and frequency band f df It has a different value compared to the original frequency band f of filter bank 30. Preferably, the ERB frequency band f erb and frequency band f df Different combinations (i.e., summation or averaging) or choices from the original frequency band f. In particular, f df Only the lower range of frequency band f is included (without summing), such as frequency band f up to 5 kHz.
[0056] Two signals X erb and X df The input signal is used as the "heart" of SEM26, which consists of three deep convolutional networks: an encoder part (encoder 38) and two decoder parts (decoders 40 and 42).
[0057] Encoder 38 inputs its signal, namely signal X erb and X df Reduced to two signals X erb and X df The combination (i.e., cascade) of the two signals X has a lower dimension than the first intermediate signal Z. More specifically, the first intermediate signal Z is combined with the two signals X. erb and X df Compared to other combinations, this has a lower frequency resolution.
[0058] According to reference [5], decoders 40 and 42 are used to expand the first intermediate signal Z into a second intermediate signal with a higher dimension compared to the first intermediate signal Z. Here, the second intermediate signal is structured into two parts, namely the frequency-dependent real-valued gain G output by decoder 40. erb It has signal X erb Dimensions (X) erb = X erb (k,f erb ()); and the nth-order depth filter coefficients C output by decoder 42 n (Where, preferably, n is chosen between 1 and 7, for example, set to 5). Therefore, the first intermediate signal Z forms a low-dimensional bottleneck between encoder 38 and decoders 40 and 42.
[0059] As described in reference [5], the real-valued gain G erb The time-frequency representation X of the first input audio signal I is fed to the gain applicator 44 of SEM 26, which applies the real-valued gain G. erb (Or more precisely, its interpolation to the original frequency band f at a frequency scale) is applied to the time-frequency representation X to obtain the short-time spectrum Y. G (Y)G = Y G (k,f)).
[0060] Short-time spectrum Y G and depth filter coefficients C n The depth filter 46 is fed into the SEM26. In a preferred implementation, the depth filter 46 corresponds to the filter disclosed in reference [7], which is incorporated herein by reference. It will feed the depth filter coefficients C n Applied to short-time spectrum Y G To generate an enhanced audio signal Y (Y = Y(k,f)), which is the output signal of SEM26, in which the noise component N of ambient sound is suppressed while the speech component S is preserved.
[0061] The enhanced audio signal Y is fed to the synthesis filter bank 48, either directly or via one or more post-processing units 50, which apply frequency-dependent amplification, dynamic compression, etc., to the enhanced audio signal Y. The synthesis filter bank 48 converts the (raw or post-processed) enhanced audio signal Y into a time-dependent processed signal O (O = O(t)), which is then output by the receiver 10.
[0062] Decoder 40 outputs a new real-valued gain G for each frame k. erb Similarly, decoder 42 outputs a new coefficient C for each frame k. n However, the output coefficient C n Used for n-tap filters. This means the spectrum Y G n frames and filter coefficients C n Multiply. These values are summed (actually implemented using matrix multiplication). Therefore, even if each time step requires the spectrum Y... G The decoder takes n frames as input and generates only a single frame of the enhanced audio signal Y at each time step. Therefore, the decoder 42 is designed to operate at a multi-frame scale; that is, it predicts the coefficients C of the n-tap filter. n The filter is applied to the spectrum Y. G To generate a new frame that enhances the audio signal Y.
[0063] Unlike DEEPFILTERNET 2 disclosed in [5], SEM26 uses an embedded signal E as an additional input, where the embedded signal E configures SEM26 to suppress noise component N. In a preferred implementation, the embedded signal E has the same dimension as the first intermediate signal Z and is applied to the latter by adder 52. More precisely, adder 52 subtracts the embedded signal E from the first intermediate signal Z. Thus, unlike DEEPFILTERNET 2, the first intermediate signal Z is indirectly fed to decoders 40 and 42 via adder 52, which modifies the first intermediate signal Z with the embedded signal E.
[0064] exist Figure 3 In the embodiments, the means 28 for providing the embedded signal E is implemented as an embedding generative model (EGM 54). EGM 54 has the same structure as the first part of SEM 26 (including analyzers 34, 36 and encoder 38); that is, EGM 54 includes a first analyzer 56 with the same structure and function as analyzer 34, a second analyzer 58 with the same structure and function as analyzer 36, and a convolutional neural network 60 similar to encoder 38. In fact, neural network 60 differs from encoder 38 only in its complexity. It has higher complexity than encoder part 38, i.e., more neurons and / or more layers. In a preferred implementation, neural network 60 of EGM 54 consists of four convolutional layers with 128 channels followed by a GRU (gated recurrent unit) layer with a hidden size of 64 or larger, while encoder 38 uses four convolutional layers with 16 channels and a GRU with a hidden size of 16, and is therefore much smaller than neural network 60.
[0065] exist Figure 3 In this embodiment, EGM54 receives its input signal from microphone 62, which also captures ambient sound in the user's environment, including a speech component S and a noise component N. Microphone 62 generates a second input audio signal I', which is fed to analysis filter bank 64 corresponding to analysis filter bank 30. Analysis filter bank 64 converts the time-dependent second input audio signal I' (I' = I'(t)) into a time-frequency representation X' (X' = X'(k,f)), similar to time-frequency representation X. The time-frequency representation X' is fed to analyzers 56 and 58. Analyzer 56 extracts signal X' from the time-frequency representation X'. erb (X' erb = X' erb (k,f erb ), similar to signal X erb It contains ERB (Equivalent Rectangular Bandwidth) characteristics. Analyzer 58 extracts signal X'. df (X' df = X' df(k,f df ), similar to signal X df It contains complex features of the time-frequency representation X'. The neural network 60 takes its input signal, namely signal X', as its input signal. erb and X' df This is reduced to a low-dimensional embedded signal E.
[0066] The neural networks of SEM26 and EGM54 (i.e., encoder 38, decoders 40, 42, and neural network 60) are trained together to effectively suppress the noise component N in the first input audio signal I, thereby generating an enhanced audio signal Y. The training process is performed during the development of the hearing system 2 before it is fully operational. Therefore, the hearing system 2 is manufactured with SEM 26 and EGM 54 in a pre-trained state. Preferably, SEM26 and EGM54 are trained using gradient descent or variations thereof such as stochastic gradient descent or Adam. Here, EGM54 is specifically trained to analyze the noise component N in ambient sound.
[0067] To support the analysis of noise component N by the EGM54, preferably, the EGM54 operates only in non-speech intervals, i.e., during time intervals when the ambient sound does not contain speech component S. For this purpose, the hearing system 2 includes a speech activity detector (VAD66) that analyzes the first input audio signal I to detect the presence (and absence) of speech component S. VAD66 controls the EGM54, for example via a control signal C, such that the EGM54 is activated when VAD66 detects the absence of speech, and deactivated when VAD66 detects the presence of speech.
[0068] exist Figure 3 In a variation of the embodiment, the hearing system 2 can be configured such that the VAD 66 receives its input from the microphone 62. In this case, preferably, the time-frequency representation X' generated by the analysis filter bank 64 is fed to the VAD 66.
[0069] Alternatively, as a further measure to reduce the numerical requirements and energy consumption of the hearing system 2, the EGM54 operates intermittently, causing it to generate the embedded signal E at a rate (embedding generation) much lower than the sampling rate of the enhanced audio signal Y updated by the SEM26. For example, the EGM54 runs once per second, while the SEM 26 generates the enhanced audio signal Y at the sampling rate of the hearing instrument 4, for example, at 48 kHz.
[0070] according to Figure 3 The embodiment of hearing system 2 can be Figure 1 Arrangement or Figure 2This is implemented in the arrangement. In the latter case, preferably, the EGM54 uses the microphone of the mobile phone 22 as microphone 62, and the analysis filter bank 64 is implemented as part of the hearing application 20. Preferably, the VAD66 is implemented as part of the hearing application 20 (especially in embodiments where it receives input from the microphone of the mobile phone 22). In another embodiment, the VAD66 is implemented as part of the hearing instrument 4.
[0071] exist Figure 3 In another variation of the embodiment, the hearing system 2 may be implemented such that the EGM 54 receives its input from the microphone 8 of the hearing instrument 4 (instead of from the microphone 62). In this case, preferably, the time-frequency representation X generated by the analysis filter bank 30 is fed to the EGM 54. In this variation, the microphone 62 and the analysis filter bank 64 are not necessary and are preferably not provided to the hearing system 2.
[0072] In a preferred implementation of the hearing system 2, all parts of SEM26 and EGM54, as well as filter banks 30, 48 and 64, preprocessing unit 32, postprocessing unit 50, and VAD66, are implemented as software modules. However, alternatively, any of these units can be implemented as hardware circuitry.
[0073] Figure 4 Another embodiment of hearing system 2 is shown, which corresponds to, unless otherwise stated, Figure 3 The illustrated embodiment. However, in Figure 4 In one embodiment, the means 28 for providing the embedded signal E includes a sound classifier 70 and a selection unit 72, which replaces... Figure 3 The EGM54. Sound classifier 70 indirectly analyzes the second input audio signal I' generated by microphone 62 to identify the current noise component N of the ambient sound as one of several predetermined noise categories K. In practice, it receives a time-frequency representation X' as input.
[0074] The sound classifier 70 transmits an indication of the sound category K corresponding to the current noise component N to the selection unit 72. The selection unit 72 accesses a data memory 74, which stores a plurality of different pre-generated embedding signals E, including one embedding signal E for each predetermined sound category K. The selection unit 72 selects the pre-generated embedding signal E corresponding to the sound category K transmitted from the sound classifier 70 from the data memory 74 and transmits the selected embedding signal E to the SEM 26 for configuration. Whenever the sound classifier 70 changes the noise category K to which the noise component N is classified, the selection unit 72 repeats this embedding selection step, thereby updating the embedding signal E fed to the SEM 26.
[0075] according to Figure 4 The embodiment of hearing system 2 can be Figure 1 Arrangement or Figure 2 This is implemented in the arrangement. In the latter case, preferably, the microphone of the mobile phone 22 is used as microphone 62, and the analysis filter bank 64 is implemented as part of the hearing application 20.
[0076] exist Figure 4 In a variant of the embodiment, the hearing system 2 can be implemented such that the sound classifier 70 receives its input from the microphone 8 of the hearing instrument 4 (instead of from the microphone 62). In this case, preferably, the time-frequency representation X generated by the analysis filter bank 30 is fed to the sound classifier 70. In this variant, the microphone 62 and the analysis filter bank 64 are not necessary and are preferably not provided to the hearing system 2.
[0077] In a preferred implementation of the hearing system 2, the sound classifier 70 and the selection unit 72 are implemented as software modules. However, alternatively, either of these units can be implemented as hardware circuitry. The data memory 74 can be part of the hearing instrument 4. Alternatively, the hearing application 20 can use a portion of the data memory of the mobile phone 22 as the data memory 74.
[0078] Those skilled in the art will understand that various changes and / or modifications can be made to the invention as illustrated in the specific examples without departing from the spirit and scope of the invention as broadly described in the claims. Therefore, these examples should be considered illustrative rather than restrictive in all respects.
[0079] References
[0080] [1] Y. Ephraim, D. Malah, “Speech enhancement using a short-time spectral amplitude estimator with minimum mean square error”, IEEE Trans, on acoustics, speech, and signal processing, vol. 32, no. 6, pp. 1109-1121, 1984; DOI: 10.1109 / TASSP.1984.1164453
[0081] [2] P. Wang and D. Wang, “Enhanced Spectral Features for Distortion-Independent Acoustic Modeling,” in INTERSPEECH, 2019, pp. 476-480; DOI: 10.21437 / lnterspeech.2O19-1493
[0082] [3] H. Schroeter, et al., “DEEPFILTERNET: A Low-Complexity Full-Bandwidth Audio Speech Enhancement Framework Based on Depth Filtering”, in ICASSP 2022-2022 IEEE Int. Conf, on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7407-7411; DOI: 10.1109 / ICASSP43922.2022.9747055
[0083] [4] S. Liu, et al., “N-HANS: A Neural Network-Based Toolkit for Audio Enhancement in the Field”, Multimedia Tools and Appl., vol. 80, no. 18, pp. 28365-28389, 2021; DOI:10.1007 / s11042-021-11080-y
[0084] [5] H. Schroeter, et al., “DEEPFILTERNET2: Real-time Speech Enhancement for Full-Bandwidth Audio in Embedded Devices”, 2022 Int. Workshop on Acoustic Signal Enhancement (IWAENC), 2022, pp. 1-5; DOI: 10.48550 / arXiv.2205.05474
[0085] [6] A. Vaswani, et al., “Attention Is All You Need,” 31st Conf, onNeural Information Proc. Systems (NIPS 2017), 2017; DOI: 10.48550 / arXiv.1706.03762
[0086] [7] H. Schroter, et al., “CLC n et al.: "Deep Learning-Based Hearing Aid Noise Reduction Using Complex Linear Encoding", in IEEE Int. Conf, on Acoustics, Speech and Signal Processing (ICASSP), 2020; DOI: 10.48550 / arXiv.2001.10218.
[0087] List of reference numerals
[0088] 2. Hearing system
[0089] 4. Hearing instruments
[0090] 5 users
[0091] 6. Shell
[0092] 8 microphones
[0093] 10. Receiver
[0094] 12 batteries
[0095] 14 Signal Processors
[0096] 16 channels
[0097] 18 Tips
[0098] 20. Hearing Applications
[0099] 22 Mobile phones
[0100] 24 Wireless Links
[0101] 26 SEM (Speech Enhancement Model)
[0102] 28. Device (for providing embedded signals to SEM)
[0103] 30 (Analysis) Filter Bank
[0104] 32 Preprocessing Units
[0105] 34 Analyzer
[0106] 36 Analyzer
[0107] 38 encoder
[0108] 40 decoders
[0109] 42 Decoders
[0110] 44 Gain Applicator
[0111] 46 Depth Filter
[0112] 48 (Synthetic) Filter Bank
[0113] 50 post-processing units
[0114] 52 Adders
[0115] 54 EGM (Embedded Generative Model)
[0116] 56 Analyzer
[0117] 58 Analyzer
[0118] 60 Neural Networks
[0119] 62 microphones
[0120] 64 (Analysis) Filter Bank
[0121] 66 VAD (Voice Activity Detector)
[0122] 70 Sound Classifier
[0123] 72 Selection Unit
[0124] 74 Data Storage
[0125] C control signal
[0126] C' control signal
[0127] C n Depth filter coefficients
[0128] E Embedded Signal
[0129] G erb Real-value gain
[0130] I (First) Input audio signal
[0131] I' (Second) Input audio signal
[0132] K Noise Category
[0133] N noise components
[0134] 0 Processed audio signal
[0135] U power supply voltage
[0136] S-voice components
[0137] X Time-Frequency Representation
[0138] X' time-frequency representation
[0139] X erb Signal
[0140] X df Signal
[0141] X' erb Signal
[0142] X' df Signal
[0143] Y-enhanced audio signal
[0144] Y G Short-time spectrum
[0145] Z (First) Intermediate Signal
Claims
1. A method for operating a hearing system (2) including at least one hearing instrument (4), the method comprising: - Sound capture step, wherein at least one first input transducer (8) of the hearing instrument (4) captures ambient sound from the environment of the hearing instrument (4) to generate a first input audio signal (I), wherein the ambient sound includes a speech component (S) and a noise component (N). - A signal processing step, wherein the first input audio signal (I) is modified at least in part by the signal processor (14) of the hearing instrument (4) to generate a processed audio signal (O); and - Output step, wherein the processed signal (O) is output as a perceptible sound to the user of the hearing instrument (4) by the output transducer (10); The signal processing steps further include a speech enhancement step, wherein, - The speech enhancement model (26) is applied to the first input audio signal (I) or an audio signal (X) derived from the first input audio signal to generate an enhanced audio signal (Y) in which the noise component (N) of the ambient sound is suppressed while the speech component (S) is preserved; and - An embedded signal (E) including information characterizing the noise component is provided to the speech enhancement model (26) to configure the speech enhancement model (26) to suppress the noise component (N).
2. The method according to claim 1, wherein, The speech enhancement model (26) includes neural networks (38, 40, 42), especially convolutional neural networks.
3. The method according to claim 1 or 2, wherein, The speech enhancement model (26) is implemented in the hearing instrument (4).
4. The method according to any one of claims 1 to 3, further comprising an embedding selection step, wherein, - A sound classifier (70) is applied to either the first input audio signal (I) or the second input audio signal (I'), in which the ambient sound is captured by the second input transducer (62) of an external device (22), wherein the sound classifier (70) analyzes the noise component (N) to classify the noise component (N) into one of several predetermined noise categories (K), and - Based on the classification, one of a plurality of pre-generated and stored embedded signals (E) is selected to be provided to the speech enhancement model (26).
5. The method according to any one of claims 1 to 3, further comprising an embedding generation step, wherein, The embedding generation model (54) is applied to the first input audio signal (I) or to the second input audio signal (I'), in which the ambient sound is captured by the second input transducer (62) of the external device (22), wherein the embedding generation model (54) analyzes the noise component (N) of the ambient sound to generate the embedding signal (E) during the operation of the hearing instrument (4).
6. The method according to claim 4 or 5, further comprising a voice activity detection step, wherein, - Apply a speech activity detector (66) to the first input audio signal (I) or the second input audio signal (I'), wherein the speech activity detector (66) analyzes the ambient sound to detect non-speech regions in which the ambient sound does not contain the speech component (S), and - The embedding selection step or the embedding generation step is performed only during the detected non-speech intervals.
7. The method according to claim 5 or 6, wherein, The embedded generative model (54) includes neural networks (60), especially convolutional neural networks.
8. The method according to any one of claims 5 to 7, wherein, The embedded generative model (54) is implemented as part of a software application (20) installed on an external computer, especially a smartphone (22).
9. The method according to any one of claims 5 to 8, - in, The speech enhancement model (26) updates the enhanced audio signal (Y) at a first sampling rate. - Wherein, the embedding generation model (54) updates the embedding signal (E) at a second sampling rate; and - Wherein, the first sampling rate is at least 10 2 10 preferred 3 Especially 10 4 The coefficient exceeds the second sampling rate.
10. The method according to any one of claims 2 to 9, wherein, In the speech enhancement step, - Through the encoder portion (38) of the speech enhancement model (26), the first input audio signal (I) or the audio signal (X) derived from the first input audio signal is reduced to a first intermediate signal (Z) with lower time and / or spectral resolution, and - The first intermediate signal (Z) is expanded into a second intermediate signal (G) with higher temporal and / or spectral resolution by at least one decoder portion (40, 42) of the speech enhancement model (26). erb C n ).
11. The method according to claim 10, wherein, In the speech enhancement step, the embedded signal (E) is applied to the first intermediate signal (Z) in particular by adding the embedded signal (E) to the first intermediate signal (Z).
12. The method according to claim 5 or 11, wherein, The encoder portions (38) of the embedding generation model (54) and the speech enhancement model (26) are implemented as the same type of neural network that differ only in the number of neurons and / or layers, wherein the embedding generation model (54) has more neurons and / or more layers than the encoder portion (38) of the speech enhancement model (26).
13. A hearing system (2) comprising at least one hearing instrument (4), said at least one hearing instrument (4) comprising: - At least one first input transducer (8) configured to capture ambient sound from the environment of the hearing instrument (4) to generate a first input audio signal (I), wherein the ambient sound includes a speech component (S) and a noise component (N). - A signal processor (14) configured to modify the first input audio signal (I) to generate a processed audio signal (0); and - Output transducer (10) configured to output the processed audio signal (0) to the user of the hearing instrument (4); The hearing system (2) also includes - A speech enhancement model (26) configured to generate an enhanced audio signal (Y) from the first input audio signal (I) or an audio signal (X) derived from the first input audio signal by suppressing the noise component (N) of the ambient sound and preserving the speech component (S); and - A device (28) for providing an embedded signal (E) including information characterizing the noise component (N) to the speech enhancement model (26) to configure the speech enhancement model (26) to suppress the noise component (N).
14. The hearing system (2) according to claim 13, wherein, The speech enhancement model (26) includes neural networks (38, 40, 42), especially convolutional neural networks.
15. The hearing system (2) according to claim 13 or 14, wherein, The speech enhancement model (26) is implemented in the hearing instrument (4), particularly as part of the signal processor (14).
16. The hearing system (2) according to any one of claims 13 to 15, wherein, The means (28) for providing the embedded signal (E) includes - A sound classifier (70) is provided, wherein the first input audio signal (I) or the second input audio signal (I') is fed to the sound classifier, wherein the second input audio signal (I') contains ambient sound captured by the second input transducer (62) of the external device (22), wherein the sound classifier (70) is configured to analyze the noise component (N) in the ambient sound to classify the noise component (N) into one of a plurality of predetermined noise categories (K); and - Selection unit (72), which is configured to select one of a plurality of predetermined embedded signals (E) based on the classification to provide to the speech enhancement model (26).
17. The hearing system (2) according to any one of claims 13 to 15, wherein, The device (28) for providing the embedded signal (E) includes an embedding generation model (54) to which either the first input audio signal (I) or the second input audio signal (I') is fed, wherein the second input audio signal (I') contains the ambient sound captured by the second input transducer (62) of the external device (22), wherein the embedding generation model (54) is configured to analyze the noise component (N) of the ambient sound to generate the embedded signal (E) during operation of the hearing instrument (4).
18. The hearing system (2) according to claim 16 or 17, wherein, The device (28) for providing the embedded signal (E) includes a speech activity detector (66), to which either the first input audio signal (I) or the second input audio signal (I') is fed. The speech activity detector (66) is configured to analyze the ambient sound to detect non-speech intervals in which the ambient sound does not contain the speech component (S). The hearing system (2) is configured to activate only during the detected non-speech intervals. - The sound classifier (70) and the selection unit (72) or - The embedded generative model (54).
19. The hearing system (2) according to claim 17 or 18, wherein, The embedded generative model (54) includes neural networks (60), especially convolutional neural networks.
20. The hearing system (2) according to any one of claims 17 to 19, wherein, The embedded generative model (54) is implemented as part of a software application (20) to be installed on an external computer, especially a smartphone (22).
21. The hearing system (2) according to any one of claims 17 to 20. - in, The speech enhancement model (26) is configured to update the enhanced audio signal (26) at a first sampling rate. - Wherein, the embedding generation model (54) is configured to update the embedding signal (E) at a second sampling rate; and - Wherein, the first sampling rate is at least 10 2 10 preferred 3 Especially 10 4 The coefficient exceeds the second sampling rate.
22. The hearing system (2) according to any one of claims 14 to 21, wherein, The speech enhancement model (26) includes - Encoder section (38), configured to reduce the first input audio signal (I) or the audio signal (X) derived from the first input audio signal to a first intermediate signal (Z) with lower time and / or spectral resolution, and - At least one decoder section (40, 42) configured to expand the first intermediate signal (Z) into a second intermediate signal (G) with higher time and / or spectral resolution. erb C n ).
23. The hearing system (2) according to claim 22, wherein, The speech enhancement model (26) is implemented such that the embedded signal (E) is applied to the first intermediate signal (Z), in particular by adding the embedded signal (E) to the first intermediate signal (Z).
24. The hearing system (2) according to claim 17 or 23, wherein, The encoder portions (38) of the embedding generation model (54) and the speech enhancement model (26) are implemented as the same type of neural network that differ only in the number of neurons and / or layers, wherein the embedding generation model (54) has more neurons and / or more layers than the encoder portion (38) of the speech enhancement model (26).
