EP2495724B1 - Method and device for estimating an interference noise - Google Patents

Abstract

The method involves obtaining power density value of noise signal output from the hearing device. The total noise signal value is compared with preset interference value from current time window by multiplying a gain factor. The background noise from current time slot is estimated based on comparison result so as to provide estimate value of background noise in the current time window. Independent claims are included for the following: (1) device for estimating background noise; and (2) hearing device.

Description

The present invention relates to a method of estimating a noise by providing a value for the power density of an overall signal including a useful signal and the noise to be estimated in a current time window, comparing the value of the overall signal with an amplification factor-multiplied estimated noise value a time window preceding the current time window, and using the smaller of the two values of the comparison as a prediction value for the noise in the current time window. Moreover, the present invention relates to a device for estimating a noise of an input device for providing the value of the power density of the total signal and a recursive minimum estimator for comparing the value of the total signal with the estimate of the preceding time window. Furthermore, the present invention also relates to a hearing device with such a device for estimating a noise. A hearing device here means any sound-emitting device that can be worn in or on the ear, in particular a hearing device, a headset, headphones and the like.

Hearing aids are portable hearing aids that are used to care for the hearing impaired. To meet the numerous individual needs, different types of hearing aids such as behind-the-ear hearing aids (BTE), hearing aid with external handset (RIC: receiver in the canal) and in-the-ear hearing aids (IDO), for example Concha hearing aids or canal hearing aids (ITE, CIC). The hearing aids listed by way of example are worn on the outer ear or in the ear canal. In addition, bone conduction hearing aids, implantable or vibrotactile hearing aids are also available on the market. The stimulation of the damaged hearing takes place either mechanically or electrically. Hearing aids have in principle as essential components an input transducer, an amplifier and an output transducer. The input transducer is usually a sound receiver, z. As a microphone, and / or an electromagnetic receiver, for. B. an induction coil. The output transducer is usually used as an electroacoustic transducer, z. As miniature speaker, or as an electromechanical transducer, z. B. bone conduction, realized. The amplifier is usually integrated in a signal processing unit. This basic structure is in FIG. 1 shown using the example of a behind-the-ear hearing aid. In a hearing aid housing 1 for carrying behind the ear, one or more microphones 2 for receiving the sound from the environment are installed. A signal processing unit 3, which is also integrated in the hearing aid housing 1, processes the microphone signals and amplifies them. The output signal of the signal processing unit 3 is transmitted to a loudspeaker or earpiece 4, which outputs an acoustic signal. The sound is optionally transmitted via a sound tube, which is fixed with an earmold in the ear canal, to the eardrum of the device carrier. The power supply of the hearing device and in particular the signal processing unit 3 is effected by a likewise integrated into the hearing aid housing 1 battery. 5

In many applications, especially in hearing aids and mobile phones, the useful signal, which is usually language, is often disturbed by noise. While stationary noise usually does not pose a major problem for speech enhancement systems of known type, non-stationary noise is usually more of a challenge. Particularly affected are single-channel (ie, a single microphone is used), model-based speech enhancement systems that are designed to suppress even very transient noises. Such single-channel speech enhancement systems can relieve the listener by attenuating noise accordingly.

Single-channel noise reduction is typically performed by so-called "Wiener filters". When creating a Wiener filter, it is necessary to estimate at least the spectral noise power density (PSD). Conventional speech enhancement systems typically require the noise to be more stationary, i. H. the characteristic of the noise changes only slowly as a function of time. Thus, the noise characteristics during speech pauses can be estimated, but this requires robust voice activity detection (VAD).

Further developed methods work on the principle of "minimum statistics" or "minimum tracking". They are able to update the noise estimate even during voice activity and thus do not require VAD. In the minimum statistics method, noisy speech is broken down into subbands, and minima are searched for in those subbands at a certain time interval. Because of the high dynamics of the speech signal, the minima should correspond to the spectral noise power density if the noise or noise is sufficiently stationary. The minima are used as inputs for setting a gain factor in the respective frequency band. However, the method fails if the noise exceeds a certain degree of unstationarity. This means that its performance breaks down in very unsteady environments (eg chatter in a cafeteria). With regard to the noise reduction with so-called "Recursive Minimum Tracking" or "Minimum Statistics" is on the book of Eberhard Hänsler and Gerhard Schmidt: "Acoustic Echo and Noise Control: Appraisal Approach", Wiley-Interscience-Verlag, 2004 and on the article of R. Martin: "Noise Power Spectral Density Estimation Based on Optimal Smoothing and Minimum Statistics", IEEE Transactions on Speech and Audio Processing, 2001, 9 (5), pages 504-512 directed.

Recently, so-called "codebook-based" speech enhancement techniques have been developed. These use a prior knowledge of speech and noise. The main idea is to estimate the spectral envelope and wideband signal powers (gains) of speech and noise from the distorted signal. Typical spectral envelopes of speech and different noise classes are stored in codebooks. For the estimation, first, a pair (a speech entry and a noise input) of spectral envelopes are taken from the respective codebooks. The optimal gain factors (i.e., the wideband speech power and the wideband noise power) are estimated by maximizing a certain optimization criterion. By way of example, the criterion is that the sum of the speech and noise codebook entries corresponds as far as possible to the current disturbed signal. In a second step, either the pair (along with the associated estimated gain factors) that most likely corresponds to the current disturbed spectrum is selected, or each pair is weighted with the probability that it corresponds to the current disturbed sound spectrum, and all so weighted couples are summed up. This provides estimates of the speech and noise components of the disturbed sound spectrum. These estimates are used as input to subsequent noise reduction, such as through a "Wiener filter". This estimation procedure is performed in short time windows (eg 8 ms) so that rapid changes in the noise characteristic can be followed almost instantaneously. A minimum statistic estimator can only follow such changes with a delay in the range of a few seconds.

Such a codebook-based algorithm is known from the article of T. Rosenkranz, "Noise Codebook Adaptation for Codebook-Based Noise Reduction, in Proceedings of International Workshop on Acoustic Echo and Noise Control (IWAENC), Tel Aviv, August 2010 known.

However, there are also three serious drawbacks to the codebook-based approach. First, the noise estimate is limited to a predefined set of codebook entries. Since these entries represent spectral envelopes, they are smoothed along the frequency axis. This means, for example, that sharp spectral peaks are not modeled. Second, the ability of the codebook-based approach to respond instantaneously to noise changes means that the estimate varies widely. Since the estimation of the broadband level is by nature not perfect and therefore fluctuates relatively strongly around the true value, unpleasant artifacts occur in the noise-free signal. Third, this codebook-based approach can not handle noise classes that have not been trained.

From the document EP 2 109 329 A2 is a multi-stage estimation method for noise reduction known. Likewise, a corresponding hearing aid is described. Two estimation algorithms are used, one for parameterizing the other. Further, one of the estimation algorithms may include recursive smoothing.

The object of the present invention is therefore to propose a method and a device with which it is possible to be able to estimate unknown noise as quickly as possible.

• Bereitstellen eines Werts für die Leistungsdichte eines Gesamtsignals, das ein Nutzsignal und das zu schätzende Störgeräusch enthält, in einem aktuellen Zeitfenster,
• Vergleichen des Werts des Gesamtsignals mit einem mit einem Verstärkungsfaktor multiplizierten Schätzwert eines Störgeräusches aus einem dem aktuellen Zeitfenster vorausgehenden Zeitfenster und
• Ermitteln und Verwenden des kleineren der beiden Werte des Vergleichs als Vorschätzwert für das Störgeräusch in dem aktuellen Zeitfenster, sowie
• Bereitstellen eines Codebuchschätzwerts für das Störgeräusch in dem aktuellen Zeitfenster und
• Verwenden des größeren Werts von dem Vorschätzwert und dem Codebuchschätzwert als Schätzwert für das Störgeräusch in dem aktuellen Zeitfenster.

According to the invention this object is achieved by a method for estimating a noise by

• Providing a value for the power density of a total signal containing a useful signal and the noise to be estimated, in a current time window,
• Comparing the value of the total signal with a gain multiplied by a gain factor Noise from a time window preceding the current time window and
• Determining and using the smaller of the two values of the comparison as a prediction value for the noise in the current time window, as well
• Providing a codebook estimate for the noise in the current time window and
• Using the larger value of the prediction value and the codebook estimate as an estimate of the noise in the current time window.

• einer Eingangseinrichtung zum Bereitstellen eines Werts für die Leistungsdichte eines Gesamtsignals, das ein Nutzsignal und das zu schätzende Störgeräusch enthält, in einem aktuellen Zeitfenster,
• einer rekursiven Minimumschätzeinrichtung zum Vergleichen des Werts des Gesamtsignals mit einem mit einem Verstärkungsfaktor multiplizierten Schätzwert eines Störgeräusches aus einem dem aktuellen Zeitfenster vorausgehenden Zeitfenster und zum Ausgeben des kleineren der beiden Werte des Vergleichs als Vorschätzwert für das Störgeräusch in dem aktuellen Zeitfenster, sowie mit
• einer Codebuchschätzeinrichtung zum Bereitstellen eines Codebuchschätzwerts für das Störgeräusch in dem aktuellen Zeitfenster und
• einer Logikeinrichtung zum Ermitteln des größeren Werts von dem Vorschätzwert und dem Codebuchschätzwert als Schätzwert für das Störgeräusch in dem aktuellen Zeitfenster.

In addition, the invention provides a device for estimating a noise with

• an input device for providing a value for the power density of a total signal, which contains a useful signal and the noise to be estimated, in a current time window,
• a recursive minimum estimator for comparing the value of the overall signal with a gain multiplied estimate of noise from a time window preceding the current time window and outputting the smaller of the two values of the comparison as the noise estimate in the current time window, and
• a codebook estimator for providing a codebook estimate for the noise in the current time slot and
• a logic means for determining the larger value of the prediction value and the codebook estimate value as an estimate of the noise in the current time window.

Advantageously, according to the invention, the "recursive minimum tracking" is combined with the "codebook-based noise estimation" in order to achieve an improved reduction of non-stationary noise. As a result, the above-mentioned disadvantages of the recursive minimum search as The disadvantages of the codebook-based estimation taken by itself are essentially eliminated.

Preferably, the value of the total signal and the estimate of a noise are each spectral values. The signal processing in the method according to the invention is then carried out in the spectral range.

In particular, it is favorable if the method is used in parallel in a plurality of frequency channels. For this purpose, the input signal is advantageously decomposed in a filter bank into the individual spectral components.

Furthermore, it is advantageous if the estimated value for the noise in the current time window is smoothed with the estimated value from the preceding time window. This is favorable in that then no excessive jumps occur in the noise reduction.

It is also particularly advantageous if the codebook estimate can be set to zero. Equivalent to this is when the codebook estimator is turned off. This makes the whole algorithm less sensitive to the fact whether the noise is known or not.

In an advantageous application, the above-described method of estimating noise is used for reducing noise. Here, in turn, it is of particular advantage if such a method for reducing noise is used for operating a hearing device or is implemented in a hearing device. As a result, in particular hearing aid wearers can benefit from the improved, combined noise reduction method.

The above-mentioned noise-estimating apparatus may be integrated into a hearing apparatus. specially This hearing device can be designed as a hearing aid.

FIG 1

eine Prinzipskizze eines Hörgeräts gemäß dem Stand der Technik;

FIG 2

ein Schaltbild einer Signalverarbeitung in einem Hörgerät;

FIG 3

ein Schaltbild eines rekursiven Störgeräuschschätzers gemäß dem Stand der Technik;

FIG 4

ein Schaltbild eines kombinierten Störgeräuschschätzers gemäß der vorliegenden Erfindung und

FIG 5

Signalverläufe von Störgeräuschen und Störgeräuschschätzungen nach unterschiedlichen Algorithmen.

The present invention will now be explained in more detail with reference to the accompanying drawings, in which:

FIG. 1

a schematic diagram of a hearing aid according to the prior art;

FIG. 2

a circuit diagram of a signal processing in a hearing aid;

FIG. 3

a circuit diagram of a recursive noise estimator according to the prior art;

FIG. 4

a circuit diagram of a combined noise estimator according to the present invention and

FIG. 5

Waveforms of noise and noise estimates according to different algorithms.

The embodiments described in more detail below represent preferred embodiments of the present invention.

In an exemplary hearing aid, signal processing according to the sketch of FIG FIG. 2 instead of. A microphone 10 of the hearing aid supplies a noisy or disturbed signal x (k). This signal is spectrally decomposed into individual frequency bands by means of a filter bank 11. This is a spectral signal X (e ^jΩ ) ready. This spectral signal is supplied to a noise ^{estimation unit} 12, which ^obtains therefrom an estimated value Ŝ _nn (e ^jΩ ) for the noise ^{power density} . A noise reduction ^filter 13 determines therefrom spectral weights Ĝ (e ^jΩ ) In a multiplier 14, the weights Ĝ (e ^jΩ ) are then multiplied by the spectrum X (e ^jΩ ) of the total signal, resulting in an estimated value Ŝ (e ^jΩ ) for the useful signal (e.g. B. pure speech signal) arises. An inverse filter bank 15 produces an estimate ŝ (k) of the useful signal in the time domain.

According to the invention, the noise estimation in the noise estimation unit 12 is now optimized. According to the invention, a noise estimation algorithm based on recursive minimum statistics and an algorithm based on one or more codebooks are combined for this purpose. This results in a noise estimation method, which has the corresponding advantages combined. For example, a codebook-based algorithm is used, as described in the article by T. Rosenkranz described at the outset. The noise estimate of the codebook based algorithm is integrated into the recursive estimation algorithm based on the minimum statistics similar to the algorithm of Eberhard Hänsler and Gerhard Schmidt mentioned earlier.

For a better understanding of the invention is described below with reference to FIG. 3 presented a model of a recursive noise estimator. The method shown there takes place in several frequency (sub) bands independently of each other. The individual frequency bands are, for example, with the in FIG. 2 obtained filter bank 11 won. For the input signal X (e ^jΩ ) or | X | ² is, for example, a periodogram of noisy speech. The output signal Ŝ _nn (e ^jΩ ) corresponds to an estimate of the noise ^{power spectrum} . The input signal is smoothed in a smoothing unit 16. The smoothed input spectrum is compared in a comparator 17 with the estimated noise spectrum of a previous window. For this purpose, the estimated interference power spectrum of the preceding time window is multiplied in advance by a constant "noise estimate gain" which corresponds to the value 1 + ε, where ε <<1. For this multiplication, the amplifier 18 is provided. It receives its input signal from a delay element 19, which in turn is fed by the interference estimate Ŝ _nn (e ^jΩ ) of the current time window. For smoothing the output signal, the estimated value of the preceding time window (signal after the delay unit 19) is subtracted from the output signal of the comparator 17 in a subtracter 20. The difference signal is multiplied by a constant in a further amplifier 21. The resulting signal is finally added in an adder 22 with the estimate of the preceding time window, which finally results in the smoothed estimate Ŝ _nn (e ^jΩ ).

With the elements 19, 20, 21 and 22, a first order IIR (Infinite Impulse Response) smoothing of the estimated noise spectrum is thus performed.

In the comparator 17, the minimum of the two signals (in the current time window and in the preceding time window) is used. It is thus a kind of efficient implementation of the minimum statistics algorithm according to the article by R. Martin.

The behavior of this known estimator comes from the graph of FIG. 5 out. The curve 23 shows the actual noise present. For example, it is street noise with fast passing cars. The estimates are determined from a mixture of this noise with a speech signal at a distance (SNR) of 0 dB. The curve 24 shows the estimation of the recursive minimum tracking algorithm. For example, as can be seen from the first two seconds of the estimate, the estimator can not follow the rapid increase in noise. The slope of the estimator is limited by the constant ε. This constant ε must be small, otherwise the estimate would follow the noisy input spectrum too quickly and speech components are erroneously included in the noise estimate.

According to the invention is carried out according to the example of FIG. 4 improving the noise estimate by combining the recursive estimate with a codebook based estimate, the combined algorithm being able to quickly track fast noise fluctuations. This is FIG. 4 The signal flow diagram shown corresponds essentially to that of FIG. 3 , Therefore, the description of FIG. 3 Referenced. A codebook-based noise estimate is integrated into the estimator by a maximum operation in a second comparator unit 27 (logic device) immediately after the comparator unit 17 with the minimum operation. The comparison unit 27 obtains a codebook estimate Ŝ _nnCB from an in FIG. 4 not shown codebook estimation device. Thus, if the actual noise is significantly underestimated (the estimate of the recursive minimum tracking algorithm is below that of the codebook based algorithm), then the codebook based estimate is taken. The recursive part of the algorithm is then able to track the noise from a higher level. The combined algorithm of the present invention, therefore, can respond to changes in noise level as quickly as codebook estimates.

FIG. 5 shows this behavior of the combined estimate. The codebook estimate is shown by curve 25. The estimate of the combined algorithm is shown by curve 26. Compared with the codebook estimate 25, the combined estimate 26 follows the increase of the noise floor with a very small delay due to the smoothing part 20, 21, 22 of the algorithm. However, it can be seen that the combined algorithm tracks the increase in noise level much faster than the recursive algorithm 24 alone. In addition, it can be seen that the inventive combined algorithm provides a better estimate when the codebook based estimate 25 underestimates the actual noise 23. Namely, in the time range between 4 and 6 seconds, the codebook estimate 25 is significantly lower than the actual noise. However, since the recursive portion of the algorithm can track the noise, the combined estimate 26 is much closer to the real noise than the codebook based estimate 25 or the recursive estimate 24 alone.

Advantageously, therefore, a codebook-based noise estimate is combined with a recursive noise estimate. The benefits of each of these estimates are gained for the combination, while the disadvantages are minimized.

The advantages of the combination are that the combined algorithm can track fast noise fluctuations much faster than conventional recursive noise estimators. Another advantage is that by injecting the codebook based estimation algorithm in the proposed manner, the estimator becomes a conventional recursive estimator when the codebook based estimation is turned off or set to zero. This in turn improves the robustness of the algorithm. Further, an advantage of the proposed combination is that the algorithm may continue to track the noise if the codebook based algorithm underestimates the actual noise floor. The combined algorithm can therefore bridge areas in which the codebook-based estimate either underestimates or shuts down the noise. In addition, the noise estimate varies significantly less than the codebook-based estimate alone, resulting in much more pleasing sound reproduction with reduced artifacts. In addition, the proposed estimator can handle noise for which the codebook-based algorithm has not been trained. This is due to the recursive part of the algorithm, which is independent of the codebook-based estimate.

Claims (10)

1. Method for estimating interference noise by

   - providing a value for the power density of a total signal, containing a wanted signal and the interference noise to be estimated, in a current time window,

   - comparing the value of the total signal with an estimated value, multiplied by an amplification factor (18), of interference noise from a time window preceding the current time window and

   - using the smaller (17) of the two values from the comparison as a preliminary estimated value for the interference noise in the current time window,

   characterised by

   - providing a codebook estimated value for the interference noise in the current time window and

   - determining and using the larger (27) of the preliminary estimated value and the codebook estimated value as the estimated value for the interference noise in the current time window.
2. Method according to claim 1, wherein the value of the total signal and the estimated value for interference noise are each spectral values.
3. Method according to claim 1 or 2, wherein the estimated value for the interference noise in the current time window is smoothed with the estimated value from the preceding time window (20, 21, 22).
4. Method according to one of the preceding claims, wherein the codebook estimated value is set to zero.
5. Method according to one of the preceding claims that is applied in a plurality of frequency channels in parallel.
6. Method for reducing interference noise by estimating the interference noise in accordance with one of the preceding claims and reducing the interference noise according to the estimated value.
7. Method for operating a hearing aid, wherein interference noise is reduced according to claim 6.
8. Device for estimating interference noise, comprising

   - an input device (11) for providing a value for the power density of a total signal, containing a wanted signal and the interference noise to be estimated, in a current time window,

   - a recursive minimum estimation device (17, 18, 19) for comparing the value of the total signal with an estimated value, multiplied by an amplification factor, of interference noise from a time window preceding the current time window and for outputting the smaller of the two values from the comparison as a preliminary estimated value for the interference noise in the current time window,

   characterised by

   - a codebook estimation device to provide a codebook estimated value for the interference noise in the current time window and

   - a logic device (27) to determine the larger of the preliminary estimated value and the codebook estimated value as the estimated value for the interference noise in the current time window.
9. Hearing device into which a device according to claim 8 is integrated for the purpose of estimating interference noise.
10. Hearing device according to claim 9, which is embodied in the form of a hearing aid.

Images