Method for classifying an acoustic environment for a hearing instrument

Abstract

The invention relates to a method for signal processing for a hearing instrument (1), wherein a first input signal (E1) is generated from an ambient sound (2) by a first input transducer (M1) of the hearing instrument (1), wherein, based on the first input signal (E1), a first classifier (DNN1) determines a first probability value (w1a-d) for a first plurality of first classes (C1a-d) of ambient noise, which indicates a probability for the presence of the respective first class (C1a-d) of ambient noise in the ambient sound (2), in particular independently of the respective other first probability values ​​(w1a-d), wherein, based on the first input signal (E1) and based on at least one of the first probability values ​​(w1a-d), a second classifier (DNN2) determines a first probability value (w1a-d) for a second plurality of second classes (C2a-c) of ambient noise, which are different from the ambient noise of the first plurality of first classes (C1a-d),a second probability value (w2a-c) is determined in each case, and wherein the at least one of the first probability values ​​(w1 ad) is used as an input variable for the second classifier (DNN2).

Description

[0001] The invention relates to a method for signal processing and in particular for classifying an acoustic ambient situation for a hearing instrument, wherein, based on a first input signal of the hearing instrument, which represents an ambient sound, a classifier determines a probability value for a plurality of classes, which indicates a probability for the presence of the respective class of ambient noise.

[0002] The term "hearing instrument" usually refers to devices that are used to output sound signals to the ear or, more generally, to the hearing center of a user of the device. Hearing aids fall under this term in particular. Hearing aids are used by people with hearing impairment to at least partially compensate for the hearing loss resulting from this hearing impairment. Hearing aids usually have at least one electroacoustic input transducer, usually in the form of a microphone, for detecting an acoustic (ambient) noise and converting it into an electrical input signal. Furthermore, such hearing aids usually have a signal processing device that is designed to filter the input signal(s) for interference components (e.g.noise, ambient noise and the like), to filter and / or attenuate these interference components and to amplify the remaining signal components as a useful signal (such as speech and / or music).

[0003] To output the processed input signal to the ear, hearing aids usually include an electroacoustic output transducer, e.g., in the form of a loudspeaker (also called a receiver or receiver), which converts the processed input signal into an output sound signal and outputs it to the ear of the hearing aid wearer. Alternatively, hearing aids have a cochlear or bone conduction receiver for outputting an output signal in electrical or mechanical form to the ear.

[0004] However, the term “hearing instrument” also includes other devices for sound output, such as headsets (“headphones”), wireless headphones with and without active noise cancellation, so-called “hearables” and the like.

[0005] Hearing aid signal processing units, in particular, usually contain device-specific, relatively complex algorithms for processing the input signals. To tailor the hearing aid to the individual wearer, these algorithms are based on variable parameters, the limits of which are specified for each individual wearer when adjusting the hearing aid to the wearer's hearing loss.

[0006] Signal processing can vary dramatically depending on the general acoustic environment. Signal processing is often simplified by classifying the acoustic environment, i.e., determining or estimating which class the current environmental situation belongs to from a predefined, manageable set of "typical" environmental situations (e.g., but not limited to, so-called "listening situations"), which are defined and differentiated from one another by a number of acoustic characteristics. Subsequently, a signal processing program specific to the determined class (often referred to as a "listening program" or as an automatic program for adaptively changing the signal processing settings) is applied to the input signal from the hearing instrument.This achieves a standardization of the complex dependencies between the acoustic environment and the associated signal processing considered optimal in each case by grouping comparable or similar acoustic environment situations.

[0007] Individual subprocesses of signal processing in a hearing instrument, such as the detection of the user's own speech activity or the direction of a useful signal source, can vary in complexity depending on the acoustic environment and therefore require different resources to determine a result with a sufficiently high degree of certainty. The quality of the applied signal processing depends significantly on the "quality" of the classification—that is, on the one hand, on the possibility of differentiation provided by the selected grouping of various possible acoustic environment situations, and on the other hand, on the precision of its detection.

[0008] It is therefore an object of the invention to allow the most precise possible classification of an acoustic environmental situation for signal processing of a hearing instrument.

[0009] The stated object is achieved according to the invention by a method for signal processing, in particular for classifying an acoustic ambient situation, for a hearing instrument, wherein a first input signal is generated from an ambient sound by a first input transducer of the hearing instrument, and wherein, based on the first input signal, a first classifier determines a first probability value for a first plurality of first classes of ambient noise, which indicates a probability for the presence of the respective first class of ambient noise, in particular independently of the respective other first probability values.

[0010] According to the method, it is provided that, based on the first input signal and based on at least one of the first probability values, a second classifier determines a second probability value for a second plurality of second classes of ambient noise, which are different from the ambient noise of the first plurality of first classes, which second probability value indicates, in particular, a probability for the presence of the respective second class of ambient noise, wherein the at least one of the first probability values ​​is used as an input variable for the second classifier. Advantageous and, in part, inventive embodiments are the subject of the dependent claims and the following description.

[0011] A hearing instrument generally includes any device that is configured to generate an electrical input signal from ambient sound by means of at least one, in particular acousto-electrical, input transducer, to process said input signal into an output signal by means of, in particular, frequency-band-specific amplification and / or compression, and to generate a sound signal from the output signal and feed it to the ear of a wearer of this device, in particular by means of an electro-acoustic output transducer (e.g., a loudspeaker, a so-called balanced metal case receiver, but also a bone conduction receiver). A hearing instrument therefore includes, in particular, headphones (e.g., as an "earbud"), a headset, data glasses with a loudspeaker, etc., which are equipped with a corresponding input transducer.However, a hearing aid in the narrower sense is also included as a hearing instrument, i.e. a device for treating a hearing impairment of the wearer, in which, when the input signal is processed to form the output signal, the former is amplified and / or compressed, in particular depending on the frequency band, in order to at least partially compensate for the hearing impairment of the wearer in a user-specific manner by means of an output sound signal generated from the output signal.

[0012] In particular, the hearing instrument can also be configured as a binaural hearing system with a first local device and a second local device. In this case, the hearing instrument can preferably also have (at least) one additional input transducer, with the first input transducer being arranged in the first local device and the additional input transducer being arranged in the second local device. The applicability of the described method is generally independent of this.

[0013] An input transducer includes, in particular, any device configured to generate a corresponding electrical signal from a sound signal. In particular, preprocessing can also occur during the generation of the first or second input signal by the respective input transducer, e.g., in the form of linear preamplification and / or A / D conversion. The correspondingly generated input signal is, in particular, provided by an electrical signal whose current and / or voltage fluctuations essentially represent the sound pressure fluctuations in the air.

[0014] A classifier is, in particular, a device or program element configured to determine the respective probability of presence for the predetermined plurality of (first or second) classes of ambient noise based on a number (preferably a plurality) of input variables determined using the first input signal and, if applicable, further input signals from additional input transducers. The classifier can be implemented as a standalone device, e.g., via an ASIC or other hard-wired circuit, or run as a program element on a signal processor configured for this purpose. The respective classifier is preferably implemented as an artificial neural network (deep neural net, DNN).

[0015] The first classes of ambient noise are preferably directly given by concrete ambient noises, for example “speech”, “noise / noise”, “quiet”, “music” etc. The second plurality of the second classes of ambient noise is different from the first plurality of the first classes of ambient noise (i.e., not congruent) and preferably completely different (i.e., any class is either a first or a second class), whereby the second classes can preferably be designed as “sub-classes” of a first class (e.g. with regard to the type of “noise” or the type of “music”, or which “speech” is involved).

[0016] A probability value for the presence of the respective first or second class of ambient noise is particularly related to the probability with which the ambient noise underlying the respective class is contained in the ambient sound “in and of itself”. In particular, this means that the first probability values ​​are independent of one another in the sense that if only one specific ambient noise changes (e.g. a decrease in “speech”), the probability values ​​of other classes (e.g. “noise” or “music”) do not necessarily have to change. In particular, this means that the probability values ​​of the first (or second) classes do not necessarily have to add up to one, since a probability value in this case does not specify or estimate which ambient noise orwhich class is dominant in the ambient sound (e.g., "the most probable of all classes"), but rather whether and to what extent the ambient noise in question is present in the ambient sound. The first and second probability values ​​preferably take values ​​from zero to one.

[0017] The second classifier uses at least one first probability value of a first class as an input variable in addition to the input variables derived from the first input signal (and possibly from other input signals). In particular, this means that the second classifier changes or adapts its functionality depending on said first probability value and, for example, for a high first probability value with given other input variables, outputs different second probability values ​​for the second classes than for a low first probability value (and otherwise identical input variables). Preferably, the first probability value used as the input variable of the second classifier is greater than zero and less than one, and in particular is not used merely as a binary trigger.

[0018] Preferably, the input variables for the first and second classifiers are updated at predefined, in particular regular, time intervals, so that corresponding time windows are formed, and the signal components of the first input signal (and possibly further input signals) in the individual time windows are incorporated into the respective input variables of the classifiers. In particular, input variables based on the same time window are used to calculate the first and second probability values ​​by the first and second classifiers, respectively. This means, in particular, that the first and second classifiers preferably operate in parallel in terms of time (in terms of the time windows), but that the second classifier is logically hierarchically subordinate to the first classifier (due to the use of the first probability value).

[0019] By using the first probability value as a "real" input variable (i.e. not just as a trigger / for activation) for the second classifier, a more precise classification can be achieved, since the second classifier not only simply classifies "subcases" of the first classes of the first classifier, but can also directly take into account how an increasing or decreasing presence of certain ambient noises (represented by the corresponding first classes) can change the correlations between the input variables of the second classifier (i.e. the values ​​of its input variables) and the resulting second probability values ​​by incorporating the corresponding first probability value into the second classifier.

[0020] On the other hand, the presented structure of the classification of ambient noise allows an advantageous implementation of the two classifiers by corresponding DNNs, since the tasks of the respective classifier can be clearly defined, thus enabling both a compact topology for the respective DNN and a manageable training effort.

[0021] Preferably, a parameter of a signal processing of the hearing instrument for the first input signal is set based on at least one of the determined second probability values. While the classification of the first and in particular the second class can also be carried out in particular to adapt parameters of the hearing instrument which only relate to a further analysis of the first input signal (and possibly further input signals and / or an intermediate signal derived from the first input signal), such as a detection of a direction of a sound source (Direction of Arrival, DoA), the probability values ​​obtained from the classifications can also be used directly for an adaptation of the signal processing by setting a corresponding parameter of the signal processing of the first input signal (and possibly further input signals and / or an intermediate signal derived from the first input signal), such asa gain value and / or a compression ratio or the like is varied depending on the determined second probability values. In particular, the parameter (and / or at least one further signal processing parameter) of said signal processing can also be adjusted in direct dependence on at least one of the determined first probability values ​​(i.e., in particular, said first probability value itself is included in the signal processing as a control variable, and not only indirectly via the second classifier and its thus determined second probability values).

[0022] Preferably, a second input signal is generated from the ambient sound by a second input transducer of the hearing instrument, with the first classifier and / or the second classifier determining the first probability values ​​or the second probability values ​​based on the second input signal. Hearing instruments often have more than one input transducer. By using the corresponding second input signal to determine the first or second probability values, spatial acoustic information can also be taken into account for the classification.

[0023] Advantageously, a first set of acoustic features is determined based on the first input signal (and possibly also based on the second input signal), wherein the first classifier determines the first probability values ​​of the first classes based on the first set of acoustic features as input variables, and / or a second set of acoustic features is determined based on the first input signal (and possibly also based on the second input signal), wherein the second classifier determines the second probability values ​​of the second classes based on the second set of acoustic features as input variables. In particular, the first and second sets of acoustic features do not need to be identical, but can (in some cases considerably) overlap.Whether the at least one first probability value used as input to the second classifier is considered an acoustic feature of the second set is ultimately a question of definition.

[0024] The advantage of using acoustic features determined from the first input signal (and possibly further input signals and / or from an intermediate signal derived from the first input signal) instead of directly using the signal components of the first input signal as input to the classifiers is that the acoustic features already provide a pre-filtering of the acoustically relevant information for classification. This means that a relatively manageable set of features can be sufficient for classification, whereas direct signal components would require a significantly more complex model for classification. In particular, this further facilitates the implementation of the classifiers using appropriate DNNs.

[0025] In this case, the acoustic characteristics of the first set and / or the second set are preferably selected from the following characteristics: centroid frequency of an overall signal and / or a noise background and / or a band-limited signal, variance of the centroid frequency, modulation depth at at least one given modulation frequency (preferably at a modulation frequency from an interval of 4 Hz ± 1 Hz), modulation phase difference of at least two different frequency bands, stationarity, signal level, noise level for at least one given frequency range (preferably for a high-frequency and / or a medium-frequency frequency range of the audible spectrum), autocorrelation value for at least one given time delay (e.g. for a given time delay from an interval of 3 ms ± 1 ms and / or a given time delay from an interval of 8 ms ± 2 ms), beat orAccent detection, onset detection, coherence, and / or correlation of the two input signals. Beat or accent detection can be specifically designed to detect a rhythm ("beat") in the musical sense. These acoustic features provide a particularly good trade-off between simple detection and high information content regarding the ambient noise to be classified.

[0026] It is expedient for at least the classes “noise” (or “noise”) and “speech” (or equivalent classes such as “noise” instead of “noise”) to be included as first classes, with the associated first probability values ​​preferably indicating the probability of the presence of noise or interference or of speech in the ambient sound.

[0027] Advantageously, the first probability value of the "speech" class is used as the input variable for the second classifier. In particular, the presence of speech can lead to acoustic features used as input variables for the second classifier having different value ranges with regard to the second classes than in the (largely) absence of speech. In other words, the same value of an input variable can lead to a different assessment with regard to the second classes in the presence of speech, and thus to a different second probability for a specific class (or its ambient noise) than in its absence. The transitions here can be continuous, and accordingly, intermediate levels should also be considered preferentially based on the first probability of the "speech" class.In particular, if (at least) the second classifier is implemented by a corresponding DNN, this can be taken into account within a training of the DNN.

[0028] Advantageously, the second classifier is activated based on the first probability value of the “noise” class, and / or a feature value of an acoustic feature of the second set is reset. Activating the second classifier based on the first probability value of the “noise” class, for example via a comparison with a lower limit, is particularly advantageous if the second classes include at least the “babble” and “traffic” classes, since these second classes themselves contain a correspondingly high proportion of noise. Resetting a feature value of an acoustic feature of the second set based on the first probability value of the “noise” class, for example via a comparison with a lower limit, is particularly advantageous for acoustic features that can be influenced and, in particular, distorted by noise onsets or transients.By resetting, the influences of such noise transients can be filtered out from the acoustic feature. An example of such an acoustic feature to which the above-mentioned influence can apply is "stationarity."

[0029] Conveniently, the second classes include at least the classes "babble" and "traffic noise" (or equivalent classes such as "traffic noise" instead of "traffic sounds"). Classifying these classes of ambient noise is particularly advantageous, especially with regard to the resulting signal processing of the first input signal.

[0030] In a particularly advantageous embodiment, for the class “babble of voices”, at least for a range of the associated second probability value, an output signal of the hearing instrument, which is reproduced by an output transducer, has a component of a directional signal that increases monotonically with the second probability value, and / or for the class “traffic noise”, at least for a range of the associated second probability value, the output signal of the hearing instrument, which is reproduced by the output transducer of the hearing instrument, has a component of an omnidirectional signal that increases monotonically with the second probability value.

[0031] This means, in particular: For the "buzz" class, there is at least one range within the value range of the associated second probability value in which the described output signal has a directional signal component that increases monotonically with the second probability value over this range of the value range. Preferably, the range is an interval of at least 30 percentage points, particularly preferably at least 50 percentage points (for the second probability value). A similar situation applies to the "traffic noise" class with regard to the omnidirectional signal.

[0032] It has also proven advantageous if temporal smoothing is applied to at least one second probability value, and at least one time constant of said smoothing is set as a function of the same second probability value at a previous point in time. This can mean, in particular, that such temporal smoothing of the second probability is "guided" in a certain direction (via a comparison with corresponding limit values ​​for the application of the smoothing), i.e. a type of "bias" is applied depending on the previous values. For example, the longer the second probability value lies above a given limit value (e.g. 0.5 or 0.6 or 0.7), the shorter the time constant for an "attack" and the longer the time constant for a "release" can be selected.If the second probability value jumps upwards again, the smoothing reacts quickly to this increase; if the "unsmoothed" second probability value drops briefly, this is initially of little importance for the smoothed value.

[0033] In a further advantageous embodiment, a binaural hearing system with two local hearing aids is used as the hearing instrument, wherein a preliminary second probability value is determined for each of the second classes in each of the two local hearing aids, and wherein for at least some of the second classes, the second probability value is determined based on the two associated preliminary second probability values ​​of the two local hearing aids. In other words, each of the two local hearing aids of the binaural hearing system initially performs the above-described method independently of the other local hearing aid based solely on the local input signal (or local input signals), wherein the preliminary second probability values ​​are determined on each side.These are then transmitted to the other local hearing aid, so that the second probability value can be determined on both sides, preferably according to the same procedure, for example by averaging or by determining a maximum or minimum. In particular, the described synchronization can also be applied to at least one first probability value (i.e., corresponding preliminary first probability values ​​are generated locally on both sides, based on which the final first probability value is then determined after corresponding transmission to the other side).

[0034] The invention further provides a hearing instrument comprising: a first input transducer for generating a first input signal from ambient sound, and signal processing means configured to implement at least a first classifier and a second classifier, wherein the hearing instrument is configured to perform the method described above. The hearing instrument is in particular equipped with corresponding signal processors and / or ASICs to implement the two classifiers.

[0035] The hearing instrument according to the invention shares the advantages of the method according to the invention. The advantages stated for the method and its further developments can be applied analogously to the hearing instrument.

[0036] An embodiment of the invention is explained in more detail below with reference to the accompanying drawings, each of which shows schematically: Fig. 1 in a block diagram a hearing instrument, Fig. 2 in a block diagram a signal flow for classifying an ambient sound in the hearing instrument according to Fig. 1, and Fig. 3 in a block diagram a binaural hearing system as a Fig. 1 alternative listening instrument.

[0037] Corresponding parts and sizes are provided with the same reference numerals in all figures.

[0038] In Fig. Figure 1 schematically shows a block diagram of a hearing instrument 1 having a first input transducer M1 and a second input transducer M2. The first input transducer M1 and the second input transducer M2 are each provided by corresponding microphones. The first input transducer M1 is configured to generate a first input signal E1 from ambient sound 2 during operation of the hearing instrument 1. Accordingly, the second input transducer M2 is configured to generate a second input signal E2 from ambient sound 2 during operation of the hearing instrument 1.The first input signal E1 and the second input signal E2 are fed to a signal processing unit 4, in which both input signals E1, E2 are processed to form an output signal A1, and in particular are amplified and / or compressed in a frequency band-specific manner. The said signal processing of the two input signals E1, E2 to form the output signal A1 is particularly direction-dependent, i.e., in the ambient sound, contributions from individual sound sources from different spatial directions can be amplified to different degrees. Furthermore, the said signal processing can be carried out in particular according to the audiological requirements of a wearer of the hearing instrument 1.

[0039] The hearing instrument 1 further comprises an output transducer L1, which is configured to generate an output sound signal 6 from the output signal A1. The schematic representation of the hearing instrument 1 in Fig. 1 shows a so-called behind-the-ear hearing aid (BTE) with an earpiece 8 in which the output transducer L1 is arranged, but the hearing instrument 1 is also conceivable as a design, in particular as an in-the-ear hearing aid (ITE), an in-the-canal hearing aid (ITC), a completely in-the-canal hearing aid (CIC), a receiver in-the-canal hearing aid (RIC), or in particular also as an earphone not exclusively or not primarily intended for the treatment of a hearing impairment.

[0040] During operation of the hearing instrument 1, the ambient sound 2 is classified into various possible classes of ambient noise based on the first input signal E1, and in particular also based on the second input signal E2. For this purpose, the signal processing device 4 has two artificial neural networks, which implement a first classifier DNN1 and a second classifier DNN2, respectively. The artificial neural networks can run via an ASIC or other hard-wired circuit in the signal processing device 4, or as a program element on a signal processor configured for this purpose in the signal processing device 4.

[0041] In Fig. 2 is a block diagram of a signal flow in the hearing instrument 1 according to Fig. 1 for the said classification of the ambient sound 2 with regard to various possible classes of ambient noise. In the present case, acoustic features Fa-n are initially determined at regular intervals (i.e. with regular updates) on the basis of the first input signal E1 and the second input signal E2. The acoustic features Fa-n preferably comprise a centroid frequency of an overall signal and / or a noise background and / or a band-limited signal, a variance of the centroid frequency, a modulation depth at a given modulation frequency, a modulation phase difference of at least two different frequency bands, a stationarity, a signal level, a noise level for at least one given frequency range, an autocorrelation value for at least one given time delay, a beat orAccent detection, onset detection, and / or coherence and / or correlation of the two input signals. The acoustic features Fa-n are used for the aforementioned classification in a manner to be described below.

[0042] The individual acoustic features Fa-n are each assigned to a first set F1 aj and / or a second set F2a-m, i.e. each of the acoustic features Fa-n can be contained either in the first set F1 aj or in the second set F2a-m or in both sets F1a-j, F2a-m. The first set F1 aj of acoustic features is then fed to the first classifier DNN1 as respective input variables. The first classifier DNN1 and also the second classifier DNN2 each have an internal structure (not shown in more detail), which in particular comprises an input layer, one or preferably several hidden layers and an output layer, wherein the hidden layers in turn have at least one so-called dense layer or one fully connected layer as well as a so-called gated recurrent unit layer.

[0043] The first classifier DNN1 now determines, on the basis of the first set F1a-j of acoustic features, for a first plurality of first classes C1a-d of ambient noise, a first probability value w1 ad for the presence of the corresponding ambient noise in the ambient sound 2. The first classes C1a-d are given here by the first classes “noise” (C1a), “speech” (C1b), “music” (C1c) and “quiet” (C1d), but could also include further classes of background noise without major changes to the said first classes C1a-d.

[0044] The first probability values ​​w1 ad each quantify the probability with which the associated ambient noise of the respective first class C1a-d is contained in the ambient sound 2, i.e., a first probability value w1x = 1 (x = a...d) means that the ambient noise of the associated first class C1x is definitely contained in the ambient sound 2, whereas a first probability value w1 x = 0 means that the ambient noise of the associated first class C1x is definitely not contained in the ambient sound 2. The individual first probability values ​​w1 ad therefore do not have to be normalized to 1 in their sum across all first classes C1a-d (they generally are not). In particular, in the event of a sudden occurrence of an ambient noise of a certain class C1x (i.e. from w1x = 0 to w1x = 0), the first probability values ​​w1 y (y ≠ x) of the other first classes C1y do not necessarily have to change.

[0045] A check is now carried out to determine whether the first probability value w1 a of the first class “noise” (C1a) exceeds a lower limit th1 (y / n).

[0046] If this is the case (y), the ambient sound 2 is further classified using the second classifier DNN2. For this purpose, the second set F2a-m of acoustic features is fed to the second classifier DNN2 as respective input variables. In addition to the second set F2a-m of acoustic features, the second classifier DNN2 also receives the first probability value w1b (0 ≤ w1b ≤ 1) of the first class "speech" (C1b) as a further input variable.

[0047] Based on these input variables, a second probability value w2a-c is now determined for a second plurality of second classes C2a-c of ambient noise for the presence of the corresponding ambient noise in the ambient sound 2. The second classes C2a-c are given here by the second classes “babble of voices” (C2a), “traffic noise” (C2b), and “other noise” (C2c), but could also include further classes of noise without major changes to the said second classes C2a-c (with the exception of the second class “other noise” / C2c).

[0048] Post-processing 9 can also be applied to the second probability values ​​w2a-c determined as described, which can in particular include temporal smoothing. In particular, at least one time constant of the temporal smoothing can depend on the respective second probability value w2a-c itself, for example by selecting a shorter time constant for an "attack" and a longer time constant for a "release" if the second probability value w2a-c lies above a predetermined limit value (e.g., 0.5 or 0.6) for a longer period.

[0049] In particular, the second probability values ​​w2a for “babble” (C2a) and w2b for “traffic noise” (C2b) can now be used to generate the output signal A1 by directional processing 7 of the first input signal E1 and the second input signal E2, if necessary also via intermediate signals derived therefrom and in particular direction-dependent (e.g. a cardioid and an anticardioid signal; not shown), the output signal A1 runs in monotonic dependence on the said second probability values ​​w2a, w2b at least for an interval of the respective second probability values ​​w2a, w2b (e.g. an interval with a lower limit of 0.3 to 0.5 and an upper limit of 0.7 to 1).This includes, in particular, that in such an interval with an increasing second probability value w2a for “babble” (C2a) (with otherwise constant second probability values ​​w2b, w2c of the other second classes C2b, C2c), the output signal A1 has a higher directionality, for example in that a higher proportion of a directional signal generated on the basis of the first and second input signals E1, E2 is included in the output signal A1. Furthermore, this includes, in particular, that in such an interval with an increasing second probability value w2b for “traffic noise” (C2b) (with otherwise constant second probability values ​​w2a, w2c of the other second classes C2a, C2c), the output signal A1 has a lower directionality, for example in that a higher proportion of an omnidirectional signal generated on the basis of the first and / or second input signals E1, E2 is included in the output signal A1.In particular, a parameter for generating the output signal A1 can also be set in direct dependence on at least one of the first probability values ​​(w1 ad) (not shown).

[0050] In Fig. 3 is one for Fig. 1 shows an alternative embodiment of the hearing instrument 1. The hearing instrument 1 is provided as a binaural hearing system 10 with a first local hearing aid HG1 and a second local hearing aid HG2, which are to be worn by a user on his or her left or right ear, respectively. Each of the two local hearing aids HG1, HG2 can, in particular, like the hearing system 1 according to Fig. 1, in particular comprising two input transducers each, so that in each local hearing aid HG1, HG2 two input signals are generated from the ambient sound 2. Each of the local hearing aids HG1, HG2 is also capable of carrying out the Fig. 2, for which purpose, in each of the two local hearing aids HG1, HG2, preferably only acoustic features are used which have been determined based on the respective local input signals of the respective local hearing aid HG1, HG2. In order to achieve a uniform classification on both sides, however, the output of the respective local second classifier (not shown) is not used directly as a second probability value for the associated second class, but is first transmitted as a preliminary second probability value w', w'' via appropriately configured communication means (not shown) to the other local hearing aid HG1, HG2 (dashed lines), so that in each of the two local hearing aids HG1, HG2, both completely locally generated preliminary second probability values ​​w', w'' are present.On each side, the respective second probability value w2y (y= a...c) for each of the second classes C2y can then be determined from the preliminary second probability values ​​w', w'' using the same rule, for example by averaging or as a maximum or minimum. In particular, this can be done using . Fig. 3 described procedure of generating on the basis of preliminary second probability values ​​can, mutatis mutandis, also be applied to at least one of the first probability values ​​(not shown).

[0051] Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited to the disclosed examples and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention. List of reference symbols 1 hearing instrument 2 Ambient noise 4 Signal processing unit 6 Output sound signal 7 directional processing 8 Earpiece 9 Post-processing 10 binaural hearing system A1 output signal C1a-d first classes C2a-c second classes DNN1 / 2 first / second classifier E1 / 2 first / second input signal Fa-n acoustic characteristics F1 aj acoustic features of the first set F2a-m acoustic features of the second set HG1 / 2 first / second local hearing aid (of the hearing instrument) L1 output converter M1 / 2 first / second input converter th1 lower limit w1a-d first probability values w2a-c, w2y second probability values w', w'' preliminary second probability value

Claims

[1] Method for signal processing for a hearing instrument (1), - wherein a first input signal (E1) is generated from an ambient sound (2) by a first input transducer (M1) of the hearing instrument (1), - wherein, based on the first input signal (E1), a first classifier (DNN1) determines a first probability value (w1a-d) for a first plurality of first classes (C1a-d) of ambient noise, which indicates a probability for the presence of the respective first class (C1a-d) of ambient noise in the ambient sound (2), in particular independently of the respective other first probability values (w1a-d), - wherein, based on the first input signal (E1) and on at least one of the first probability values (w1a-d), a second classifier (DNN2) determines a second probability value (w2a-c) for a second plurality of second classes (C2a-c) of ambient noise, which are different from the ambient noise of the first plurality of first classes (C1a-d), and - wherein the at least one of the first probability values (w1a-d) is used as input for the second classifier (DNN2). [2] Method according to claim 1, wherein a parameter of a signal processing of the hearing instrument (1) for the first input signal (E1) is set on the basis of at least one of the determined second probability values (w2a-c). [3] Method according to claim 1 or claim 2, wherein the parameter of the signal processing of the hearing instrument (1) for the first input signal (E1) is also set on the basis of at least one of the determined first probability values (w1a-d). [4] Method according to one of the preceding claims, wherein a second input signal (E2) is generated from the ambient sound (2) by a second input transducer (M2) of the hearing instrument (1), and wherein the first classifier (DNN1) and / or the second classifier (DNN2) determines the first probability values (w1a-d) and the second probability values (w2a-c) also based on the second input signal (E2). [5] Method according to one of the preceding claims, wherein a first set of acoustic features (F1a-j) is determined based on the first input signal (E1), and the first classifier (DNN1) determines the first probability values (w1a-d) of the first classes (C1a-d) based on the first set of acoustic features (F1a-j) as input variables, and / or wherein a second set of acoustic features (F2a-m) is determined based on the first input signal (E1), and the second classifier (DNN2) determines the second probability values (w2a-c) of the second classes (C2a-c) based on the second set of acoustic features (F2a-m) as input variables. [6] Method according to claim 5, wherein the acoustic features (F1a-j, F2a-m) of the first set and / or the second set are selected from the following features: centroid frequency of an overall signal and / or a noise background and / or a band-limited signal, variance of the centroid frequency, modulation depth at at least one given modulation frequency, modulation phase difference of at least two different frequency bands, stationarity, signal level, noise level for at least one given frequency range, autocorrelation value for at least one given time delay, beat or accent detection, onset detection, coherence and / or correlation of the two input signals (E1, E2). [7] Method according to one of the preceding claims, wherein the first classifier (DNN1) is implemented by a first artificial neural network, and / or wherein the second classifier (DNN2) is implemented by a second artificial neural network. [8] Method according to one of the preceding claims, wherein the first classes (C1a-d) comprise at least the classes “noise” (C1a) and “speech” (C1b). [9] Method according to claim 8, wherein the first probability value (w1b) of the class “speech” (C1b) is used as input for the second classifier (DNN2). [10] Method according to claim 8 or claim 9, wherein on the basis of the first probability value (w1a) of the class “noise” (C1a) - the second classifier (DNN2) is activated, and / or - a feature value of an acoustic feature (F2a-m) of the second set is reset. [11] Method according to one of the preceding claims, wherein at least the classes “babble” (C2a) and “traffic noise” (C2b) are included as second classes (C2a-c). [12] Method according to claim 11, wherein for the class “babble” (C2a), at least for a range of the associated second probability value (w2a-c), an output signal of the hearing instrument (1), which is reproduced by an output transducer (L1), has a component of a directional signal that increases monotonically with the second probability value (w2a), and / or wherein for the class “traffic noise” (C2b), at least for a range of the associated second probability value (w2b), the output signal of the hearing instrument (1), which is reproduced by the output transducer (L1) of the hearing instrument (1), has a component of an omnidirectional signal that increases monotonically with the second probability value (w2b). [13] Method according to one of the preceding claims, wherein temporal smoothing is applied to at least one second probability value (w2a-c), and wherein a time constant of said smoothing is set as a function of the same second probability value (w2a-c) at a previous time. [14] Method according to one of the preceding claims, wherein a binaural hearing system (10) with two local hearing aids (HG1, HG2) is used as the hearing instrument (1), wherein in each of the two local hearing aids (HG1, HG2) a preliminary second probability value (w', w'') is determined for each of the second classes (C2a-c), and wherein for at least some of the second classes (C2a-c) the second probability value (w2a-c) is determined on the basis of the two associated preliminary second probability values (w', w'') of the two local hearing aids (HG1, HG2). [15] Hearing instrument (1), comprising - a first input transducer (M1) for generating a first input signal (E1) from an ambient sound (2), and - Signal processing means (4) which are configured to implement at least a first classifier (DNN1) and a second classifier (DNN2), wherein the hearing instrument (1) is configured to carry out the method according to one of the preceding claims.

Images