EP1244094A1 - Method and apparatus for determining a quality measure for an audio signal - Google Patents

Abstract

The invention relates to a method for determining a quality measure (2) for an audio signal (1), whereby the speech signal component (4) is first extracted from the audio signal (1). A reference signal (6) is then generated using said signal, by means of noise suppression (7) and interrupt interpolation (8). The above is compared with the speech signal (4) and an intrusive quality value (10) thus determined. A further quality value (15) is determined, by determining and evaluating Codec-related signal distortions in the speech signal (4). A further quality value (17) is generated from information on the detected signal interruptions (8). The quality measure (2) is finally determined as a linear combination (16) of the various quality values (10, 15, 17, 18).

Description

Technical field

The invention relates to a method for determining a quality measure of an audio signal. The invention further relates to a device for performing this method and a noise reduction module and an interrupt detection and interpolation module for use in such a device.

State of the art

Assessing the quality of a telecommunications network is an important tool to achieve or maintain a desired quality of service. One way that Assessing the quality of service of a telecommunications network is quality to determine a signal transmitted over the telecommunications network. With audio signals, Various intrusive methods are known for this, particularly in the case of voice signals. In such methods, as the name suggests, the system under test is used intervened by occupying a transmission channel and transmitting a reference signal therein becomes. The quality assessment is then carried out by comparing the known one Reference signal with the received signal, for example, subjectively by or a large number of test subjects. However, this is complex and therefore expensive.

EP 0 980 064 describes a further intrusive method for machine-assisted quality assessment described an audio signal, being used to assess the transmission quality a spectral similarity value of the known source signal and the received signal is determined. This similarity value is based on a calculation of the covariance the spectra of the source signal and the received signal and a division of the covariance by the standard deviations of the two spectra mentioned.

However, intrusive methods generally have the disadvantage that, as already mentioned, this too testing system must be intervened. To determine the signal quality namely at least one transmission channel is occupied and a reference signal is transmitted therein become. This transmission channel cannot be used for data transmission during this time be used. It is also in a broadcasting system such as a broadcasting service in principle possible, the signal source for transmission of test signals, since this occupies all channels and the test signal this would be extremely impractical for all recipients. Intrusive Methods are also unsuitable to simultaneously control the quality of a variety of transmission channels to monitor.

Presentation of the invention

The object of the invention is to provide a method of the type mentioned above, which avoids the disadvantages of the prior art and in particular offers a possibility to assess the signal quality of a transmitted over a telecommunications network Signal without knowing the originally sent signal.

The solution to the problem is defined by the features of claim 1. In the inventive Process for machine-based determination of a quality measure of an audio signal, a reference signal is first determined from the audio signal. through Comparing the determined reference signal with the audio signal becomes a quality value determined, which is used to determine the quality measure.

The method according to the invention thus allows an assessment of the quality of an audio signal at any connection of the telecommunication network. That is, it allows thus the quality assessment of many transmission channels at the same time, even a simultaneous assessment of all channels would be possible. The quality assessment takes place solely on the basis of the properties of the received signal, d. H. without knowing the source signal or the signal source.

The invention thus not only enables monitoring of the transmission quality of the Telecommunications network, but also, for example, quality-based cost allocation, quality-based routing in the network, a test of the coverage ratio for example in the case of mobile radio networks, a QOS (Quality of Service) control of the network nodes or a quality comparison within a network or across networks.

An audio signal transmitted via a telecommunication network points next to the desired one Signal information typically also includes unwanted components such as different noise components, which are not in the original source signal were present.

In order to be able to carry out the best possible quality assessment, one is possible good estimate of the signal originally sent is necessary. To this reference signal There are different methods to reconstruct. One way is one Estimating the characteristics of the transmission channel and starting from the received signal to calculate backwards. Another option is one direct estimation of the reference signal based on the known information about the Receive signal and the transmission channel.

In the present method, the reference signal is determined by the in received signal estimated interference signal and then from the received signal can be removed. By removing the noise components from the audio signal first, a noisy audio signal is determined, which is preferred as Reference signal is used to assess the transmission quality.

There are various methods for removing noise components from the received audio signal. The audio signal could, for example, be passed through appropriate filters. A preferred method is to remove the noise components from the audio signal however, a neural network is used for this.

However, the audio signal is not used directly as an input signal. First up the audio signal applied a discrete wavelet transform (DWT). This transformation provides a plurality of DWT coefficients of the audio signal, which the neural Network as an input signal. The neural network delivers on Output a plurality of corrected DWT coefficients, from which with the inverse DWT the reference signal is obtained. This corresponds to the noiseless version of the Audio signal.

To achieve this, the coefficients of the neural network must be set in this way be that to the DWT coefficients of a noisy input signal provides the DWT coefficients of the corresponding noiseless input signal. In order to the neural network delivers the desired coefficients, it must first with a Set of corresponding noisy or noisy signal pairs trained become.

In this way, both stationary noise, such as white, thermal noise and suppress vehicle or road noise, as well as impulse noise. Echo interference and interference can also be suppressed with the neural network or eliminate.

When determining the quality measure, in addition to the quality value determined by the Comparison of the received audio signal with the reference signal determined therefrom any other information will also be taken into account. This can be both information contained in the audio signal, as well as information about the transmission channel or the telecommunications network itself.

It is advantageous to use information when determining the quality measure, which are obtained from the received audio signal with suitable means to let. For example, the quality of the received audio signal is determined by the quality of the Transmission influenced codec's (coder - decoder) influenced. It is difficult to do such Determine signal degradations, for example if the codec bit rates are too low part of the original signal information is lost. However, have too small Codec bit rates result in a change in the fundamental frequency (pitch) of the audio signal, which is why the course and dynamics of the fundamental frequency in the audio signal are advantageously examined becomes. Because such changes are easiest based on audio signal sections with vowels examined, signal components in the audio signal are preferred detected with vowels and then examined for pitch variations.

Back to the determination of the reference signal from the received audio signal. This can namely not only have unwanted signal components, it can also partially on the way desired information has been lost. So can the received audio signal for example, have more or less long signal interruptions.

However, the closer the reference signal generated from the audio signal to the original one Source signal, the more precise the assessment of the transmission quality. this is the Reason for replacing signal interruptions with suitable signals. For example suitable noise signals or signal sections already transmitted are used become.

However, in order to obtain the most accurate possible estimate of the reference signal, it is from The advantage of detecting such signal interruptions in the audio signal first and then the missing signal sections were achieved by interpolation as accurately as possible To replace estimates. The type of interpolation of the lost signal sections depends on the length of the signal interruption. With short breaks, d. H. in the case of interruptions of up to a few samples in the audio signal, preference is given to a polynomial and for medium-long breaks, d. H. from a few to a few A dozen samples are preferably model-based interpolation.

Longer signal breaks, i. H. Interruptions from a few dozen samples, can but can hardly be reconstructed meaningfully. Instead of making this information redundant consider and discard them, and in part the information about them short and medium-long signal interruptions, preferably when assessing the transmission quality considered. They flow into the determination of the quality measure the calculations.

The received audio signal can include various types of audio signals. So it can contain, for example, speech, music, noise or even quiet signal components. The quality assessment can of course be based on all or part of this Signal components take place. In a preferred variant of the invention, the assessment the signal quality, however, is limited to the speech signal components. With an audio discriminator the speech signal components are therefore first extracted from the audio signal and only these speech signal components for determining the quality measure, d. H. to Determination of the reference signal used. To determine the quality value is in In this case, of course, the determined reference signal does not match the received audio signal, but only compared with the extracted speech signal component.

The device according to the invention for machine-assisted determination of a quality measure an audio signal comprises first means for determining a reference signal the audio signal, second means for determining a quality value by means of comparisons the determined reference signal with the audio signal and third means for determination the quality measure taking into account the quality value.

The first means for determining a reference signal from the audio signal can be several Include modules. So is preferably a noise reduction module and / or a Interrupt detection and interpolation module provided.

With the noise reduction module, noise signal components can be received Suppress audio signal. It includes the means to carry out those already described Wavelet transformations and the neural network to determine the new one DWT coefficients. The interrupt detection and interpolation module has those Means on the one hand for the detection of signal interruptions in the audio signal and on the other hand for the polynomial interpolation of short and for model-based interpolation of medium-long signal interruptions are required. That determined so The reference signal thus corresponds to a noisy version of the received audio signal and typically only shows major signal interruptions.

The information about the signal interruptions of the audio signal is not only used to determine a better reference signal, they can also be used for determination of a better quality measure can be used. The third means of determination of the quality measure are therefore preferably designed such that information can be taken into account via signal interruptions in the audio signal.

The more information about the audio signal included when determining the quality measure the more accurate the quality assessment can be. The device therefore advantageously has fourth means for determining information about codec-related Signal distortion. These include, for example, a vowel detection module, with which signal components with vowels can be detected in the audio signal. These vowel signal components are passed on to an assessment module, which is based on this Signal components Information about codec-related signal distortions determines which can also be used to assess the signal quality. The third means are appropriate formed such that this information about the codec-related signal distortion can be taken into account when determining the quality measure.

However, it is advantageous not to use the entire audio signal, but only its voice signal components used for quality assessment. Corresponding to the one already described The device therefore has, in particular, fifth means for extracting the method Speech signal components from the audio signal. Accordingly, to determine the Reference signal not the audio signal itself, but only its voice signal component noisy and examined for interruptions. Likewise, of course, the audio signal, but only compared its voice signal component with this reference signal. In order to the quality measure is only determined on the basis of the information in the Speech signal component, whereby the information from the remaining signal components is not taken into account become.

From the following detailed description and the entirety of the claims there are further advantageous embodiments and combinations of features of the invention.

Brief description of the drawings

Fig. 1

ein schematisch dargestelltes Blockdiagramm des erfindungsgemässen Verfahrens;

Fig. 2

das Rauschunterdrückungsmodul im Betriebszustand;

Fig. 3

das Rauschunterdrückungsmodul im Trainingszustand;

Fig. 4

das neuronale Netzwerk des Rauschunterdrückungsmoduls und

Fig. 5

ein Beispiel für ein Audiosignal mit einem Unterbruch.

The drawings used to explain the exemplary embodiment show:

Fig. 1

a schematically illustrated block diagram of the inventive method;

Fig. 2

the noise reduction module in the operating state;

Fig. 3

the noise reduction module in the training state;

Fig. 4

the neural network of the noise reduction module and

Fig. 5

an example of an audio signal with an interruption.

In principle, the same parts are provided with the same reference symbols in the figures.

Ways of Carrying Out the Invention

FIG. 1 shows a block diagram of the method according to the invention. Here, for a Audio signal 1 determines a quality measure 2, which is also used for evaluation, for example of the telecommunications network used (not shown). The audio signal 1 is understood here to mean that signal which is a receiver after transmission over the telecommunications network. This audio signal 1 typically does not match the one sent by the (not shown) transmitter Signal coincides, because on the way from the transmitter to the receiver the transmission signal opens varied ways. For example, it goes through various modules such as voice encoders and decoders, multiplexers and demultiplexers or even voice enhancers and echo cancellers. But the transmission channel itself can also be a big one Influence the signal, which can take the form of interference, fading, Express transmission interruptions or interruptions, echo generation etc.

The audio signal 1 thus not only contains the desired signal components, i. H. the original Transmitted signal, but also unwanted interference signal components. It may also be that Signal components of the transmission signal are missing, i. H. lost during transmission are.

In the example shown, however, the assessment of the signal quality is not based on of the entire audio signal 1, but only based on the contained therein Speech portion. The audio signal 1 is initially based on an audio discriminator 3 Speech signal components 4 examined. Found speech signal components 4 become further Processing passed, whereas other signal components such as music 5.1, breaks 5.2 or strong signal interference 5.3 sorted out and processed further or can be discarded. To make this distinction, the audio signal 1 piecewise, d. H. to pieces a each about 100 ms to 500 ms, passed to the audio discriminator 3. This breaks these pieces down further individual buffers of about 20 ms in length, processes these buffers and then arranges them in each case one of the signal groups to be distinguished: voice signal, music, pause or strong interference to.

The audio discriminator 3 is used, for example, to assess the signal pieces an LPC (linear predictive coding) transformation, with which the coefficients of a adaptive filters corresponding to the human speech tract are calculated. The The assignment of the signal pieces to the different signal groups is based on the Form of the transmission characteristics of this filter.

In order to be able to assess the quality of the transmission, this speech signal component 4 now a reference signal 6, i. H. a best possible estimate of the original from the transmitter transmitted transmission signal, determined. This reference signal estimation is carried out in several stages.

In a first stage, a noise suppression module 7, undesirable components are initially created Signal components such as stationary noise or impulse interference from the voice signal component 4 removed or suppressed. This is done with the help of a neural network, which was previously used as an input and a plurality of noisy signals each trained the corresponding noise-free version of the input signal as the target signal has been. The noise-free speech signal 11 obtained in this way is sent to the second stage forwarded.

In the second stage, the interrupt detection and interpolation module 8, there are interruptions detected in the audio signal 1 or in its speech signal component 4 and if possible interpolated, d. H. the missing samples are replaced by suitably estimated values.

In the present example, signal interruptions are detected by means of an examination discontinuities in the fundamental signal frequency (pitch tracing). The interpolation is carried out depending on the length of the detected break. With short Interruptions, d. H. Interruptions of a few samples in length become a polynomial Interpolation such as a Lagrangian, Newton, Hermite, or Cubic Spline interpolation applied. With medium-long interruptions (a few to a few dozen Samples), model-based interpolations such as a maximum a posteriori, an autoregressive or a frequency-time interpolation is used. With longer ones Signal interruptions is an interpolation or other signal reconstruction in generally no longer possible in a meaningful way.

The whole thing is complicated by the fact that there are both different types of Interruptions - a distinction must be made between syllable and word breaks and correct ones Signal interruptions - as well as different types of techniques for processing such There are interruptions in the transmission channel. So from a terminal, for example depending on information about the transmission network, different on missing Frames to be responded to. In a first method, for example, lost frames simply replaced with zeros. A second method uses instead of lost frames other correctly received frames used and a third Instead of the lost frames, the method uses locally generated noise signals "comfort noise" used.

After the determination of the reference signal 6 with the noise reduction module 7 and the interruption detection and interpolation module 8 with the aid of the comparison module 9 compared with the speech signal component 4. An algorithm can be used for this comparison can be used as it is for example in intrusive methods for comparison of the known source signal is used with the received signal. Suitable are, for example, psychoacoustic models, the signals are perceptual, i. H. perceptible to compare. The result of this comparison is an intrusive quality value of 10. For determination this intrusive quality value 10 becomes the input signals, ie the Speech signal component 4 and the reference signal 6, in signal pieces of about 20 to 30 ms Length broken down and a partial quality value calculated for each signal piece. After about 20 to 30 signal pieces, which corresponds to a signal duration of 0.5 seconds, becomes intrusive Quality value 10 determined as the arithmetic mean of these partial quality values. The intrusive quality value 10 forms the output signal of the comparison module 9.

When determining the quality measure 2, however, in addition to the information about Interference signal components or signal interruptions also other information about the audio signal 1 are taken into account. For example, a speech encoder or speech decoder, which the transmitted signal passes on its way from the transmitter to the receiver has an influence on the audio signal 1. These influences exist, for example in that both the fundamental frequency and the frequencies of the higher harmonics of the signal vary. The lower the bit rate of the speech codecs used, the greater the frequency shifts and thus the signal distortions.

Such influences are easiest to examine in vowels, which is why Noisy speech signal 11 is first supplied to a vocal detector 12. This includes for example a neural network that was previously used for the detection of certain (individual or all) vowels have been trained. Vowel signals 13, i.e. H. Signal components which recognizes the neural network as vowels are forwarded to an evaluation module 14, other signal components are rejected.

The evaluation module 14 divides the vowel signal 13 into pieces of approximately 30 msec calculates a DFT (discrete Fourier transformation) with a frequency resolution of approximately 2 Hz at a sampling frequency of approximately 8 kHz. Leave with it then determine the fundamental frequency as well as the frequencies of the higher harmonics and examine for variations. Another feature to evaluate the codec-related Distortion forms the dynamics of the signal spectrum, with a smaller dynamic poor signal quality means. The reference values for the dynamic evaluation are obtained from the sample signals for the individual vowels. From the information on the influence of codecs on frequency shifts and spectrum dynamics the audio signal 1 or the noisy speech signal 11 becomes a codec quality value 15 derived.

When determining the quality measure 2 by the evaluation module 16 is additionally an interruption quality value for the intrusive quality value 10 and the codec quality value 15 17 considered. This value includes information about the length and the number of interruptions detected by the interruption detection and interpolation module 8, in a preferred embodiment of the invention only the information about the long breaks. In addition, of course also further quality information 18 about the received audio signal 1 or the noisy speech signal 11, which is determined with other modules or examinations are included in the calculations of quality measure 2.

The individual quality values are now scaled so that they are between 0 and 1 lie, with a quality value of 1 an undiminished quality and Values below 1 indicate a correspondingly reduced quality. The quality measure 2 is finally calculated as a linear combination of the individual quality values, whereby the individual weighting coefficients are determined experimentally and determined in such a way that their sum is 1.

Are there further quality-relevant information available via the telecommunications network? Available or new effects appear in the transmission channels, it is simple and way possible to add further modules for the calculation of further quality values and in determining the quality measure 2 in the manner described to consider.

Some of the modules are explained in more detail below with reference to FIGS. 2 to 5. figure 2 shows the noise reduction module 7. The speech signal component 4 of the audio signal 1 is first subjected to a known DWT 19 (discrete wavelet transformation). Similar to DFTs, DWTs are used for signal analysis. An essential one The difference is, however, in contrast to those used in a DFT, unlimited in time and thus sine or cosine waveforms not temporally localized, the use of so-called wavelets, i.e. H. limited and thus localized Average 0 waveforms.

The speech signal component 4 is divided into signal pieces of approximately 20 ms to 30 ms, which are each subject to DWT 19. The result of DWT 19 is a set of DWT coefficients 20.1, which are fed as input vector to a neural network 20 become. Its coefficients have previously been trained to match a given one Set of DWT coefficients 20.1 of a noisy signal a new set of DWT coefficients 20.2 deliver the noiseless version of this signal. This new set of DWT coefficient 20.2 is now the IDWT 21, i. H. subject to the DWT inverse to DWT 19. In this way, this IDWT 21 delivers a largely noiseless version of the Speech signal components 4, the desired, noiseless speech signal 11.

The training configuration of the neural network 20 is shown in FIG. 3. It will be with Trained pairs of noisy and noiseless versions of sample signals. On Noisy example signal 22.1 is subjected to DWT 19 and it becomes a first one Obtained set 20.3 of DWT coefficients. The noisy example signal 22.2 is also subjected to the same DWT 19 and a second set 20.4 of DWT coefficients generated, which is fed into the neural network 20. The output vector of the neural Network 20, the new DWT coefficients 20.5, is in a comparator 23 with the first Theorem 20.3 of DWT coefficients compared. Because of the differences between These two sets of DWT coefficients are corrected 24 for the coefficients of the neural network 20. This process is done with a variety of sample signal pairs repeated so that the coefficients of the neural network 20 perform the desired function perform more and more precisely. Advantageously, for training the neural Network 20 uses sample signals 22.1, 22.2, which human sounds from different Represent languages. It is also an advantage to do this for both women and women Use men's and children's voices. The mentioned size of the to be processed individually Signal pieces of 20 ms to 30 ms duration is selected so that the processing of the Speech signal portion 4 are carried out regardless of the language and the speaker can. Even pauses in speech and very quiet signal sections are trained with this these are also recognized correctly.

In the present exemplary embodiment, a multilayer perceptron was used as the neural network 20 with an input layer 25, a hidden layer 26 and one Output layer 27 used. The perceptron was trained with a back propagation algorithm. The input layer 25 has a plurality of input neurons 25.1, the hidden layer 26 a plurality of hidden neurons 26.1 and the Output layer 27 has a plurality of output neurons 27.1. Any input neuron 25.1 becomes one of the DWT coefficients 20.1 of the previous DWT 19 fed. After the input signals have passed through the neural network, where the respective values are determined with the set coefficients of the respective neurons and the value combinations are calculated in the individual neurons each output neuron 27.1 one of the new DWT coefficients 20.2. As before mentions, the audio discriminator 3 divides the signal pieces into individual buffers Length 20 ms. At a sampling rate of 8 kHz, this corresponds to 160 samples. For this For example, a neural network 20 with 160 input and output neurons each can be used 25.1, 27.1 and about 50 to 60 hidden neurons 26.1 can be used.

The interpolation of a signal interruption will be briefly described with reference to FIG. 5. For example, time-frequency interpolation is used for signal reconstruction. First of all, a short-term spectrum for signal frames with 64 samples Length (8 ms) calculated. This is done by placing the signal frames with Hamming windows be multiplied at an overlap of 50%.

The goal of interpolation is to treat this gap. First, a frequency time Transformation done. This leads to a three-dimensional signal display, which for each point in the time-frequency level (x-y level) the power spectrum in Direction of the z-axis. An interruption at a given time t is simply too recognize as zero points along the line x = t in the time-frequency plane.

FIG. 5 shows such a signal 28 of approximately 200 samples in length. About periodicity 5 shows the signal 28 in the temporal domain. On the abscissa axis 32 is the number of samples and the ordinate axis 33 the magnitudes applied. However, the interpolation takes place in the frequency-time domain. In Figure 5 The interruption 29 is easy to recognize as a gap of just under 10 samples in length.

A polynomial interpolation is now carried out for each frequency component, both for the phase, as well as the magnitude, with minimal phase and magnitude discontinuity he follows. For this purpose, the pitch period 30 of the signal 28 is first of all determined. For The interpolation will be information from the samples before and after the gap within this pitch period 30 is taken into account. The signal areas 31.1, 31.2 show those Areas of the signal 28 a pitch period before or after the interruption 29. This Signal areas 31.1, 31.2 are not identical to the original signal piece at break 29, but still show a high degree of similarity. For little ones Gaps up to about 10 samples are assumed to have enough signal information is present in order to be able to carry out correct interpolation. With longer gaps additional information from samples of the environment can be used.

In summary it can be stated that the invention allows the signal quality of a judge received audio signal without knowing the original broadcast signal. Of course, the signal quality can also affect the quality of the transmission channels used and thus concluded on the quality of service of the entire telecommunications network become. The fast response times of the method according to the invention, which are in the order of magnitude of approximately 100 ms to 500 ms, thus making different possible Applications such as general comparisons of the service quality of different ones Networks or subnetworks, a quality-based cost allocation or a quality-based one Routing in a network or across multiple networks using appropriate Control of the network nodes (gateways, routers etc.).

Claims (13)

Method for machine-based determination of a quality measure of an audio signal, characterized in that a reference signal is determined from the audio signal and a quality value is determined by comparing the reference signal with the audio signal, which is used to determine the quality measure. A method according to claim 1, characterized in that a noise-free audio signal is determined by removing noise signal components from the audio signal and this is used as a reference signal. A method according to claim 2, characterized in that the noisy audio signal is determined by subjecting the audio signal to a discrete wavelet transformation, the coefficients of which are fed into a previously trained neural network and the output signals of which are subjected to the inverse, discrete wavelet transformation. Method according to Claim 2 or 3, characterized in that signal components with vowels are detected in the noisy audio signal, information about codec-related signal distortions is determined therefrom and these are taken into account when determining the quality measure. Method according to one of Claims 1 to 4, characterized in that signal interruptions in the audio signal are detected and the reference signal is determined by at least partially reconstructing it during the signal interruptions, the reference signal preferably having a polynomial in the case of short signal interruptions and preferably having a in the case of medium-length signal interruptions model-based interpolation is reconstructed. A method according to claim 5, characterized in that information about the signal interruptions is taken into account when determining the quality measure. Method according to one of claims 1 to 6, characterized in that before the determination of the reference signal from the audio signal, a voice signal component is extracted and the determination of the quality measure is limited to the voice signal component. Device for machine-based determination of a quality measure of an audio signal, characterized in that it has first means for determining a reference signal from the audio signal, second means for determining a quality value by comparing the reference signal with the audio signal, and third means for determining the quality measure taking into account the quality value. Apparatus according to claim 8, characterized in that the first means have a noise suppression module for suppressing noise signal components and / or an interruption detection and interpolation module for the detection and interpolation of signal interruptions in the audio signal, and the third means are designed such that signal interruptions in determining the Quality measures can be taken into account. Apparatus according to claim 8 or 9, characterized in that it has means for determining codec-related signal distortions, which comprise a vowel detection module for detecting vowel signal components in the audio signal and an evaluation module for determining the codec-related signal distortions, the third means are designed in such a way that the codec-related signal distortions can be taken into account when determining the quality measure. Device according to one of claims 8 to 10, characterized in that it has means for extracting a speech signal component from the audio signal and is designed to determine the quality measure of the speech signal component. Noise suppression module for use in a device according to claim 8, characterized in that it has means for performing a discrete wavelet transformation for calculating signal coefficients of an audio signal, a neural network for calculating corrected signal coefficients and means for performing an inverse wavelet transformation of the corrected signal coefficients for determining the audio signal without noise signal components. Interrupt detection and interpolation module for use in a device according to claim 8, characterized in that it has means for detecting signal interruptions in an audio signal and means for interpolating signal interruptions of the audio signal, these preferably for polynomial interpolation of short or for model-based interpolation of medium-long signal interruptions are formed.

Images