Noise reduction methods in a hearing instrument

Abstract

The invention describes a method for noise reduction in a hearing instrument (1), wherein an acousto-electrical input transducer (E1) of the hearing instrument (1) generates an input signal (E1) from ambient sound (2), wherein a frequency-bandwise noise reduction (21) is applied to a processing signal (V1) derived from the input signal (E1), in which stationary noise in the respective frequency band (Bj, Bk) is detected, and depending on the detected stationary noise, a gain factor (Gk) of the processing signal (V1) for the respective frequency band (Bj, Bk) is set, wherein an analysis (12) is applied to the processing signal (V1), which is configured to detect both stationary and non-stationary noise, wherein said analysis (12) has a lower frequency resolution than the frequency-bandwise noise reduction (21).and wherein, in the case of a detected presence (Y) of stationary and / or non-stationary noise in said analysis (12) of the processing signal (V1), the stationary noise in the frequency-bandwise noise suppression (21) is assumed to be detected in each frequency band (Bj, Bk), and the corresponding gain factors (Gk) are set.

Description

[0001] The invention relates to a method for noise suppression in a hearing instrument, wherein a frequency-band-wise noise suppression is applied to a processing signal derived from an input signal of the hearing instrument, in which stationary noise is detected in the respective frequency band.

[0002] The term "hearing instrument" generally refers to devices that output sound signals to the ear, or more generally, to the auditory center of a user of the device. Hearing aids, in particular, fall under this definition. Hearing aids serve to at least partially compensate for the hearing loss resulting from hearing impairment in individuals with this condition. Typically, hearing aids have at least one electroacoustic input transducer, usually in the form of a microphone, to detect acoustic (ambient) noise and convert it into an electrical input signal. Furthermore, such hearing aids regularly include a signal processing unit designed to filter out interference from the input signal(s) (e.g., background noise).to analyze noise, ambient noise and the like, to filter and / or attenuate these interfering components and to amplify the remaining signal components as the 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 converter, for example in the form of a loudspeaker (also called a receiver), by means of which the processed input signal is converted into an output sound signal and delivered to the ear of the hearing aid wearer. Alternatively, hearing aids have a cochlear or bone conduction receiver to deliver an output signal to the ear in electrical or mechanical form.

[0004] Signal processing in hearing aids can vary drastically depending on the overall acoustic environment. Individual subprocesses of signal processing, such as the detection of speech activity (especially the user's own speech) or noise reduction, can vary in complexity and require different resources depending on the acoustic environment in order to reliably provide adequate signal processing parameters (such as frequency-band-dependent gain factors for noise reduction). The quality of the applied signal processing depends significantly on the "quality" of the signal analysis, i.e., the precision of its detection.Since noise reduction is a process that is performed in many situations in addition to other signal processing algorithms, a resource-efficient implementation is desired.

[0005] WO 2011 / 041 738 A2 discloses an electronic device for suppressing noise in an audio signal. The electronic device comprises a processor and instructions stored in memory. The electronic device receives an input audio signal and calculates an estimate of total noise based on estimates of steady-state noise, non-steady-state noise, and excess noise. The electronic device also calculates a matching factor based on an input signal-to-noise ratio (SNR) and one or more SNR limits. Furthermore, based on the total noise estimate and the matching factor, a set of gain factors is calculated, which are applied by the electronic device to the input audio signal to produce a noise-suppressed audio signal.

[0006] The object of the invention is therefore to allow the simplest and most precise noise suppression possible for the signal processing of a hearing instrument.

[0007] The aforementioned problem is solved according to the invention by a method for noise suppression in a hearing instrument, wherein an acousto-electrical input transducer of the hearing instrument generates an input signal from an ambient sound, wherein a frequency-band-wise noise suppression is applied to a processing signal derived from the input signal, in which stationary noise in the respective frequency band is detected, and depending on the detected stationary noise, a gain factor of the processing signal for the relevant frequency band is set.Here, the stationary noise in the respective frequency band is detected by means of a first level measurement and a second level measurement, wherein the first level measurement is parameterized to have a short first settling time and a long first decay time, and the second level measurement is parameterized to have a long second settling time and a long second decay time, and wherein the stationary noise is detected by means of the difference between the first level measurement and the second level measurement.

[0008] The method involves applying an analysis to the processing signal, which is configured to detect both stationary and non-stationary noise. This analysis has a lower frequency resolution than the frequency-bandwise noise suppression. If stationary and / or non-stationary noise is detected in the analysis of the processing signal, the stationary noise is assumed to be detected in each frequency band by the frequency-bandwise noise suppression, and the corresponding gain factors are set. Preferably, an output signal is generated based on the processing signal amplified bandwise by these gain factors. Advantageous and, in some cases, inventive embodiments are the subject of the dependent claims and the following description.

[0009] In this context, a hearing instrument generally encompasses any device designed to generate an audio signal from an electrical signal—which may also be an internal signal of the device—and deliver it to the ear of a wearer of that device. This includes, in particular, headphones (e.g., as "earbuds"), headsets, smart glasses with speakers, etc. However, a hearing instrument also includes a hearing aid in the narrower sense, i.e., a device for treating a wearer's hearing impairment. In this device, an input signal generated from an ambient signal by means of a microphone is processed into an output signal, with amplification being particularly frequency-dependent. The output audio signal generated from the output signal by means of a loudspeaker or similar device is suitable for at least partially compensating for the wearer's hearing impairment, particularly in a user-specific manner.

[0010] An acousto-electrical input converter, in this context, includes in particular any device designed to generate a corresponding electrical signal from a sound signal. In particular, the generation of the first or second input signal by the respective input converter may also involve preprocessing, e.g., in the form of linear pre-amplification and / or analog-to-digital conversion. The input signal generated is, in particular, an electrical signal whose current and / or voltage fluctuations essentially represent the sound pressure fluctuations in the air.

[0011] A processing signal derived from the input signal is understood to mean, in particular, that signal components of the input signal are incorporated into the processing signal. The processing signal may also contain signal components of other signals (such as another input signal in the case of directional processing of the input signals). The signal components of the input signal may be amplified, particularly in a frequency-dependent manner, or band-limiting (i.e., complete suppression of certain frequency bands) of the input signal may also occur. However, the signal components of the input signal may also be incorporated into the processing signal completely and / or unchanged. Thus, the processing signal may also be directly derived from the (complete and unchanged) input signal.

[0012] Frequency-bandwise noise suppression means, in particular, that the processing signal is divided into a plurality of frequency bands in a suitable manner, preferably by means of an analysis filter bank (whereby, in particular, adjacent frequency bands may have a finite overlap at their respective band boundaries), and the signal components of the individual frequency bands are suppressed in the described dependence on the respective stationary noise in the relevant frequency band by applying a corresponding gain factor a priori to the signal components of the processing signal in the frequency band in order to lower or raise them relative to each other in said dependence.

[0013] Stationary noise, in this context, is defined in particular as noise that exhibits level variations of less than 5 dB, preferably less than 3 dB, in the respective frequency band over a timescale of at least 0.5 s, preferably at least 1 s, and particularly preferably at least 2 s. Specifically, any signal components that meet the required conditions for stationaryness (i.e., do not exceed a maximum variation in signal level over the described minimum period) can be considered noise. This is particularly true under the assumption that the desired signals are usually speech signals or music, which each exhibit a higher variation in signal level within a frequency band.

[0014] An analysis of the processing signal, designed to detect both stationary and non-stationary noise, means, in particular, that the processing signal, especially in an analysis path parallel to frequency-bandwise noise suppression, is examined for noise based on (especially temporal) features of the signal components, the information content of which preferably goes beyond the aforementioned examination of the variation of signal levels. This analysis has a lower frequency resolution than frequency-band-dependent noise suppression and can, in particular, make a broadband statement about the presence of stationary and / or non-stationary noise (i.e., a statement about whether any form of noise is present anywhere in the processing signal, regardless of the frequency range).If such a statement is made by analyzing individual frequency ranges separately, these ranges are each larger, and preferably significantly larger, than the respective bandwidths of the frequency-dependent noise reduction (i.e., preferably at least twice as large, and particularly preferably on average at least four times as large). This ensures, in particular, that the analysis has a lower latency than the frequency-dependent noise reduction and can therefore be used to control it as described.

[0015] In particular, the aforementioned analysis of the processing signal can be performed using an artificial neural network (ANN), whereby temporal signal strings and / or analysis features are generated from the processing signal and passed to the ANN as input for analysis. The signal strings can be short fragments of the processing signal, for example, with a length of > 1 ms, preferably > 5 ms, and most preferably > 30 ms. The analysis features can be temporal and / or spectral features of the processing signal, such as modulation properties (e.g., the modulation depth at 4 Hz, or a center frequency and its change, etc.).

[0016] The processing signal can also be analyzed, for example, using a set of wavelets, for each of which a correlation with signal components (especially the aforementioned signal strings) of the processing signal is determined. Based on the different correlations with these wavelets, potentially non-stationary noise in the processing signal can be detected. The analysis can also be performed using speech recognition, so that general (i.e., stationary and / or non-stationary) noise is present when no speech is recognized (and vice versa).

[0017] If noise is detected in the aforementioned analysis of the processing signal—regardless of whether it is stationary or non-stationary—the frequency-bandwise noise reduction considers the stationary noise detected in each frequency band and sets the corresponding gain factor for that frequency band. Thus, the frequency-bandwise noise reduction checks for a criterion in each frequency band: whether stationary noise is present. If so, noise reduction (via the corresponding gain factor) is applied. In contrast, the analysis checks for a different criterion: whether noise of any kind (stationary or non-stationary) is present. If so, the frequency-bandwise noise reduction intervenes, so that stationary noise is considered detected in all frequency bands.

[0018] In other words, the analysis of the processing signal is used here as an additional trigger or activator for frequency-bandwise noise reduction, while the actual noise suppression still occurs (at least also) via the frequency-bandwise gain factors. The analysis of the processing signal merely affects the trigger or activator of the frequency-bandwise noise reduction by identifying it (namely, the stationary noise) as detected for all frequency bands, and consequently setting the corresponding gain factors in the frequency bands to be applied to the processing signal.

[0019] A logical application of noise (stationary or non-stationary) detected by analyzing the processing signal as a trigger / activator for frequency-bandwise noise reduction would be to simply multiply an additional activation factor on the individual gain factors, depending on the analysis. However, under normal circumstances (and with standard parameter settings), the analysis would no longer detect noise during a so-called "speech onset" (i.e., the onset of speech). In this case, however, the attenuation in those frequency bands where non-stationary noise (with variations on a timescale of, for example, approximately 1 second, such as several short noise bursts from traffic or a machine, etc.) is present would be abruptly reduced.This problem persists because the frequency-bandwidth noise reduction in these signals does not have a corresponding gain factor set for noise reduction, as frequency-bandwidth noise reduction only reacts to stationary noise. These noise components would be directly "co-modulated" (i.e., effectively mixed in) with the output signal generated from the noise-suppressed processing signal, which would be clearly perceptible and is therefore undesirable.

[0020] The present invention therefore takes a different approach to integrating the trigger / activator from the analysis of the processing signal with respect to general noise into the frequency-band-wise noise suppression. Instead of relying on the application of gain factors, the invention intervenes in the frequency-band-wise detection of (stationary) noise and sets the stationary noise as present for each frequency band, so that the corresponding gain factor for suppressing (stationary) noise can be set in each frequency band and applied to the signal components of the processing signal. In the aforementioned situation of a speech onset (which represents a temporary, brief absence of noise), nothing would change for frequency bands with stationary noise (since the stationary noise for these would still be detected and suppressed accordingly via the associated gain factor).In frequency bands without stationary noise, noise suppression can be omitted during a speech onset (thus preserving the speech components) if, in the case of a detected speech onset, noise suppression is based solely on the detected stationary noise. This prevents the described co-modulation of the stationary noise.

[0021] Preferably, the onset of speech in the ambient noise is detected based on the input signal (i.e., by means of so-called "Voice Activity Detection" applied to the input signal, the processing signal, or another signal derived from the input signal). In this case, the frequency-bandwise noise reduction adjusts the gain of the processing signal in each frequency band depending on the detected stationary noise. In particular, upon detection of speech onset (i.e., when a speech onset is detected), the operation of the frequency-bandwise noise reduction can be reset to the state that was valid before the general (stationary and / or non-stationary) noise was detected by analyzing the processing signal.

[0022] Preferably, an output signal of the hearing instrument is generated from the processing signal to which the described noise reduction (and in particular the frequency-band-dependent gain factors set according to the above criteria) is applied. This output signal is then converted into an output sound signal by an (electro-acoustic) output converter of the hearing instrument, whereby voltage and / or current fluctuations in the output signal are particularly preferably converted into corresponding amplitude fluctuations of the output sound signal. The output converter can be, in particular, a loudspeaker, a so-called balanced metal case receiver, or even a bone conduction receiver.

[0023] According to the invention, for frequency-band-wise noise suppression, the stationary noise in the respective frequency band is detected by means of a first level measurement and a second level measurement. Thus, in each frequency band, two different level measurements with preferably different time constants are applied to the signal components of the processing signal on at least one edge (rising or falling), and stationary noise in the frequency band is detected there based on the two aforementioned level measurements, in particular based on a difference and / or a quotient of the two level measurements. The use of two level measurements has the advantage that, through the appropriate selection of a release constant, the level measurements can intrinsically carry the information about the stationary nature of the signal level.

[0024] According to the invention, the first level measurement is parameterized to have a short initial settling time and a long initial decay time, and the second level measurement is parameterized to have a long second settling time and a long second decay time, wherein the initial decay time and the second decay time are preferably identical. The stationary noise is detected based on the difference or quotient of the first and second level measurements, and the gain factor for the processing signal is preferably set depending on said difference. In other words, the first level measurement reacts quickly to short-term level changes in the frequency band and can therefore function as a so-called "peak tracker," while the second level measurement reacts more slowly to changes in the signal level than the first level measurement.However, the decay time is long for both level measurements and preferably identical, so that the level measurements gradually converge towards each other during their exponential decay behavior for stationary signal levels.

[0025] Advantageously, the gain of the processing signal for the relevant frequency band is also set monotonically as a function of the difference between the first and second level measurements. This means, in particular, that the gain of the processing signal is higher the greater the difference between the two level measurements (and thus the greater the influence of a peak that enters the first level measurement faster than the second), and that the gain of the processing signal is lower, or its attenuation is higher, the smaller the difference between the two level measurements (and thus the increasingly steady signal level in the frequency band).

[0026] Advantageously, in the event of the detection of stationary and / or non-stationary noise, the stationary noise is presumed to be detected in the frequency-bandwise noise suppression in each frequency band by changing the respective values ​​of the first settling and decay times towards the respective values ​​of the second settling and decay times, and in particular by setting them equal, so that the difference between the level measurements decreases, preferably by at least half, and in particular disappears. Specifically, the short first settling time of the first level measurement is adjusted to the higher value of the second settling time of the second level measurement.

[0027] It is further advantageous if the analysis of the processing signal detects the presence or absence of speech, and in the case of a detected absence of speech, stationary or non-stationary noise is assumed to have been detected. This particularly implies that the analysis recognizes only two distinct results: the presence of speech, which is equated with the absence of any kind of noise, and the absence of speech, which is equated with the presence of general noise. The detection of non-stationary noise, in particular, based on speech detection is especially advantageous because speech recognition algorithms are highly advanced and can be performed with sufficient precision, particularly using a broadband signal.

[0028] Particularly preferred is the modification of the input signal and / or the processing signal, specifically the modification of the values ​​of the first and second settling times, to differentiate them. This modification particularly restores the original values ​​of the first and second settling and decay times that were valid before the presence of stationary and / or non-stationary noise was detected by the aforementioned analysis of the processing signal. In other words, speech recognition / VAD can be applied to the input signal and / or the processing signal, thus enabling the detection of speech in the ambient sound. In this case, the settling time values ​​are modified to increase their difference.In particular, the original values ​​for both the onset and decay times, which were valid before the detection of the present speech, can be restored. The onset of speech can also be detected in the aforementioned analysis of the processing signal.

[0029] The described procedure makes it possible for the frequency-band-wise noise suppression to resume normal operation when speech is detected, thus suppressing stationary noise in the relevant frequency bands, while in the frequency bands containing speech components, these are no longer suppressed.

[0030] The invention further describes a method for noise suppression in a hearing instrument, wherein an acousto-electrical input transducer of the hearing instrument generates an input signal from ambient sound, wherein a frequency-bandwise noise suppression is applied to a processing signal derived from the input signal, in which stationary noise is detected in the respective frequency band, and depending on the detected stationary noise, a gain factor of the processing signal for the relevant frequency band is set, wherein an analysis is applied to the processing signal which is configured to detect both stationary and non-stationary noise, wherein said analysis has a lower frequency resolution than the frequency-bandwise noise suppression.

[0031] According to the invention, the processing signal is analyzed using an ANN, wherein one output class of the ANN is the presence of stationary and / or non-stationary noise. In the case of a detected presence of stationary and / or non-stationary noise in said analysis of the processing signal, the stationary noise is assumed to be detected in the frequency-bandwise noise suppression in each frequency band, and the corresponding gain factors are set. An ANN is particularly suitable for performing said analysis, especially when the output classes are simplified as described.

[0032] Preferably, a plurality of acoustic features are generated from the input signal or the processing signal, which are used as input variables for the artificial neural network. These acoustic features are selected from the following set: center frequency of an overall signal and / or a noise background, modulation depth at at least one given modulation frequency, stationarity, signal level, noise level for at least one given frequency range, and autocorrelation value for at least one given time delay. These features are highly informative regarding the presence of noise and can be easily determined for a signal without significantly increasing the signal delay.

[0033] The invention further describes a hearing instrument comprising: an acousto-electrical input converter for generating an input signal from ambient sound, and a signal processing device, wherein the hearing instrument is configured to perform the method described above. Preferably, the signal processing device is equipped with appropriate signal processors and / or ASICs to implement the individual process steps of the signal processing.

[0034] The hearing instrument according to the invention shares the advantages of the method according to the invention. The advantages stated for the method and for its further developments can be transferred analogously to the hearing instrument.

[0035] Preferably, the hearing instrument includes means for implementing an ANN (e.g., appropriate signal processors and / or ASICs). Analyzing the processing signal to detect stationary and / or non-stationary noise can be implemented particularly effectively and with the highest possible real-time accuracy using an ANN.

[0036] An embodiment of the invention is explained in more detail below with reference to the drawings. The drawings schematically depict: Fig. 1 in a block diagram a hearing instrument, and Fig. 2 in a block diagram a method for noise reduction for the hearing instrument according to Fig. 1.

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

[0038] In Fig. Figure 1 schematically depicts a hearing instrument 1 in a block diagram, which has an input transducer M1. The input transducer M1 is represented here by a suitable microphone. The input transducer M1 is configured to generate an input signal E1 from an ambient sound 2 during operation of the hearing instrument 1. The hearing instrument 1 can also have a further input transducer (in Fig. (1 not shown) which is configured to generate a further input signal from the ambient sound 2 during operation of the hearing instrument 1. The first input signal E1 (and, if applicable, the further input signal) is fed to a signal processing unit 4, in which the input signal E1 (and, if applicable, the further input signal) is processed into an output signal A1, and in particular, is amplified and / or compressed frequency-wise. The aforementioned signal processing of the input signal and the further input signal to the output signal A1 can, in particular, be direction-dependent, i.e., contributions of individual sound sources from different directions in the ambient sound can be amplified to varying degrees. Moreover, the aforementioned signal processing can, in particular, be carried out according to the audiological requirements of a user 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 is shown in Fig. Figure 1 shows a so-called Behind-The-Ear (BTE) hearing aid with an earpiece 8 in which the output transducer L1 is arranged, however the hearing instrument 1 is also conceivable as a design type, in particular as an In-The-Ear (ITE), an In-The-Canal (ITC), a Completely-In-the-Canal (CIC), a Receiver-In-the-Canal (RIC), or in particular also as an earphone not exclusively or primarily intended for the treatment of hearing loss.

[0040] In Fig. Figure 2 is a schematic block diagram illustrating the process of noise reduction in the hearing instrument 1 according to Fig. Figure 1 shows that a processing signal V1 is derived from the input signal E1 by means of a preprocessing step 9 (not shown in detail, which may include downsampling or similar processes, but also band limiting). However, the processing signal V1 can also be identical to the input signal E1. In particular, the (in Fig. 1 and Fig. (2. Not shown) In the case of a further input signal, the processing signal V1 is derived from both input signals.

[0041] The broadband processing signal V1 is now subjected to an analysis 12 in an analysis path 10 to determine whether general (i.e., stationary and / or non-stationary) noise is present in the processing signal. For this purpose, various temporal and spectral features 14 are determined from the processing signal V1. The features 14 can be obtained, in particular, from the following (non-exhaustive) group: center frequency of a total signal and / or a noise background, modulation depth at at least one given modulation frequency, signal level, noise level, and autocorrelation value for at least one given time delay. The features 14 are then passed as inputs to an ANN 16, which is trained to detect the presence Y or absence N of general noise based on the features 14. In particular, the ANN 16 can also detect an absence orDetect the presence of speech, and use this to detect general noise, i.e., an absence of speech implies the presence (Y) of stationary and / or non-stationary noise (the converse is not necessarily true, but the absence of general noise can also be detected using other parameters / features).

[0042] Furthermore, frequency-band-wise noise suppression 21 is applied to the processing signal V1 in a processing path 20. For this purpose, the processing signal V1 is divided into a plurality of frequency bands Bj by means of a filter bank 22. In each of the frequency bands Bj, a first level measurement P1 and a second level measurement P2 are applied to the signal components Sk of the processing signal V1 in that band. The first level measurement P1 is performed with a short initial settling time Ta1 and a first decay time Tr1, and the second level measurement P2 with a long second settling time Ta2 > Ta1 and a second decay time Tr2 = Tr1 that is preferably identical to the first decay time Tr1. This is done for the sake of clarity in Fig.Figure 2 is shown only as an example for frequency band Bk (but this applies to all frequency bands Bj). In principle, and regardless of the present embodiment, the two level measurements can be implemented, in particular, using first-order recursive low-pass filters with different time constants for the rising and falling edges, respectively.

[0043] The result 32 of the second level measurement P2 is now subtracted from the result 31 of the first level measurement P1 at a node 26, and the difference 28 is monotonically mapped to a gain factor Gk for noise reduction in the frequency band Bk. This gain factor Gk is then multiplied at a node 36 on the signal components Sk of the processing signal V1 in the frequency band Bk, for which purpose these components are additionally bypassed 34 past the two level measurements P1 and P2.

[0044] The signal components SkG of the processing signal V1, amplified by the respective gain factors Gk, are then combined at a synthesis filter bank 40, and the output signal A1 is generated from this (any further signal processing of the signal resulting from the synthesis filter bank 40 to the output signal A1 is to be neglected here, but can be solved by defining a preliminary output signal as said resulting signal).

[0045] Due to the different settling times Ta1 < Ta2 of the level measurements P1, P2, the difference 28 of the results 31, 32 reacts to stationary noise in the respective frequency band Bj; that is, if the signal component Sk is more stationary, the results 31, 32 of the two level measurements P1, P2 converge towards each other with an exponentially decreasing slope, which is why the difference 28 becomes increasingly smaller. This stationary signal component Sk is then considered noise, which is why the difference 28 is monotonically mapped to the respective gain factor (here Gk): With decreasing difference 28 (i.e., increasing stationaryity), the gain factor Gk becomes smaller, and the amplified signal component SkG is increasingly suppressed.

[0046] However, during a short peak (i.e., the actual opposite of a stationary signal component Sk), the results 31, 32 of the level measurements P1 and P2 initially diverge abruptly due to the different settling times Ta1 < Ta2, so that their difference 28 also increases. Accordingly, the gain factor Gk also increases, so that the amplified signal components SkG are raised.

[0047] If the presence Y of stationary and / or non-stationary noise is detected by the analysis 12 of the broadband processing signal V1 in the analysis path 10, this results in the first settling time Ta1 being increased to the second settling time Ta2 of the respective second level measurement P2 in all first level measurements P1 (i.e., in the first level measurement P1 of each frequency band Bj) in the processing path 20 (i.e., Ta1 ↦ Ta2), and this preferably applies at least as long as the presence Y is detected in the analysis 12.

[0048] This allows noise, including non-stationary noise (which would otherwise be undetectable based on the two level measurements P1 and P2), to be detected in all frequency bands Bj and suppressed via the corresponding gain factors (Gk). Furthermore, if the analysis 12 detects the sudden presence of speech, the original values ​​of the two settling times Ta1 < Ta2 can be restored, so that the frequency-band-wise noise suppression 21 only actually works in those frequency bands Bj and Bk in which it detects (and thus suppresses) stationary noise, thereby preserving the speech components in the other frequency bands Bj and Bk without stationary noise. This assumes that, in addition to the speech components, no other non-stationary signal components, and in particular no significant amount of non-stationary noise, are present.

[0049] Although the invention has been illustrated and described in detail by the preferred embodiment, the invention is not limited by the disclosed examples and other variations can be derived by the person skilled in the art without leaving the scope of protection of the invention. Reference symbol list 1 hearing instrument 2 Ambient sound 4 Signal processing unit 6 Output sound signal 8 earpiece 9 Preprocessing 10 Analysis path 12 Analysis 14 features 16 ANN 20 Processing path 21 frequency-band-wise noise reduction 22 filter bank 26 Node 28 difference 31 Result (of the first level measurement) 32 Result (of the second level measurement) 34 Bypass 36 Node 40 Synthesis filter bank A1 Output signal Year of manufacture, frequency bands E1 Input signal Gk amplification factor L1 output converter M1 input converter N Absence of general noise P1 / 2 first / second level measurement Sk signal components (of the processing signal) in frequency band SkG amplified signal components Ta1 / 2 first / second settling time Tr1 / 2 first / second decay time V1 Processing Signal Y Presence of general noise

Claims

[1] Method for noise reduction in a hearing instrument (1), - wherein an acousto-electrical input transducer (M1) of the hearing instrument (1) generates an input signal (E1) from an ambient sound (2), - wherein a frequency-bandwise noise suppression (21) is applied to a processing signal (V1) derived from the input signal (E1), in which -- a stationary noise is detected in the respective frequency band (Bj, Bk), -- the stationary noise in the respective frequency band (Bj, Bk) is detected by means of a first level measurement (P1) and a second level measurement (P2), -- where the first level measurement (P1) is parameterized such that it has a short initial settling time (Ta1) and a long initial decay time (Tr1), -- where the second level measurement (P2) is parameterized such that it has a long second settling time (Ta2) and a long second decay time (Tr2), -- where the stationary noise is detected based on the difference (28) between the first level measurement (P1) and the second level measurement (P2), and -- depending on the detected stationary noise, a gain factor (Gk) of the processing signal (V1) is set for the relevant frequency band (Bj, Bk), - wherein an analysis (12) is applied to the processing signal (V1), which is configured to detect both stationary and non-stationary noise, wherein said analysis (12) has a lower frequency resolution than the frequency-bandwise noise suppression (21), and - wherein, in the case of a detected presence (Y) of stationary and / or non-stationary noise in said analysis (12) of the processing signal (V1), the stationary noise in the frequency-bandwise noise suppression (21) is assumed to be detected in each frequency band (Bj, Bk), and the corresponding gain factors (Gk) are set. [2] Method according to claim 1, wherein the onset of speech in the ambient sound is detected on the basis of the input signal (E1), and in this case the frequency band-wise noise suppression (21) in each frequency band (Bj, Bk) adjusts the gain factor (Gk) of the processing signal (V1) depending on the respective detected stationary noise. [3] Method according to claim 1 or claim 2, wherein the gain factor (Gk) of the processing signal (V1) for the relevant frequency band (Bj, Bk) is set monotonically depending on the difference (28) from the first level measurement (P1) and the second level measurement (P2). [4] Method according to one of the preceding claims, wherein in the case of a detected presence (Y) of stationary and / or non-stationary noise, the stationary noise is assumed to be detected in the frequency-bandwise noise suppression (21) in each frequency band (Bj, Bk) by changing the respective values ​​of the first settling and decay time (Ta1, Tr1) in the direction of the respective values ​​of the second settling and decay time (Ta2, Tr2). [5] Method according to claim 4, wherein in the aforementioned case the respective values ​​of the first and second settling and decay times (Ta1, Ta2, Tr1, Tr2) are equated. [6] Method according to one of the preceding claims, wherein the analysis (12) of the processing signal (V1) detects the presence or absence of speech, and in the case of a detected absence of speech, stationary or non-stationary noise is assumed to be detected. [7] Method according to one of claims 4 to 6, wherein in the case of a detected onset of speech in the input signal (E1) and / or in the processing signal (V1) the values ​​of the first and second settling and decay times (Ta1, Ta2, Tr1, Tr2) are each changed away from each other. [8] Method according to claim 7, wherein in the case of detected speech onset the original values ​​of the first and second settling and decay times (Ta1, Ta2, Tr1, Tr2) are restored before detection of the presence (Y) of stationary and / or non-stationary noise. [9] Method according to one of the preceding claims, wherein the analysis (12) of the processing signal (V1) for the detection of both stationary and non-stationary noise is performed broadband. [10] Method for noise suppression in a hearing instrument (1), - wherein an acousto-electrical input transducer (M1) of the hearing instrument (1) generates an input signal (E1) from an ambient sound (2), - wherein a frequency-bandwise noise suppression (21) is applied to a processing signal (V1) derived from the input signal (E1), in which -- a stationary noise is detected in the respective frequency band (Bj, Bk), and -- depending on the detected stationary noise, a gain factor (Gk) of the processing signal (V1) is set for the relevant frequency band (Bj, Bk), - wherein an analysis (12) is applied to the processing signal (V1) which is configured to detect both stationary and non-stationary noise, wherein said analysis (12) has a lower frequency resolution than the frequency-bandwise noise suppression (21), - wherein the analysis (12) of the processing signal (V1) is performed using an artificial neural network (16), wherein an output class of the artificial neural network is the presence (Y) of stationary and / or non-stationary noise, and - wherein, in the case of a detected presence (Y) of stationary and / or non-stationary noise in said analysis (12) of the processing signal (V1), the stationary noise in the frequency-bandwise noise suppression (21) is assumed to be detected in each frequency band (Bj, Bk), and the corresponding gain factors (Gk) are set. [11] Method according to claim 10, wherein a plurality of acoustic features (14) are generated from the input signal (E), which are used as input variables of the artificial neural network (16), and wherein the acoustic features (14) are taken from the following set: center frequency of an overall signal and / or a noise background, modulation depth at at least one given modulation frequency, stationarity, signal level, noise level for at least one given frequency range, autocorrelation value for at least one given time delay. [12] Method according to claim 10 or claim 11, wherein the analysis (12) of the processing signal (V1) detects the presence or absence of speech, and in the case of a detected absence of speech, stationary or non-stationary noise is assumed to be detected. [13] Method according to any one of claims 10 to 12, wherein the analysis (12) of the processing signal (V1) for the detection of both stationary and non-stationary noise is performed broadband. [14] Hearing instrument (1), comprising: - an acousto-electrical input transducer (M1) for generating an input signal (E1) from an ambient sound (2), and - a signal processing device (4) wherein the hearing instrument (1) is configured to perform the method according to any of the preceding claims. [15] Hearing instrument (1) according to claim 14, further comprising means for implementing an artificial neural network (16).

Images