EP1386307B2 - Method and device 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 area

The invention relates to a method for determining a quality measure of an audio signal. Furthermore, the invention relates to an apparatus for carrying out this method as well as a noise suppression module and an interruption detection and interpolation module for use in such a device.

State of the art

The assessment of the quality of a telecommunications network is an important tool for achieving or maintaining a desired service quality. One way to assess the quality of service of a telecommunications network is to determine the quality of a signal transmitted over the telecommunications network. In the case of audio signals, in particular speech signals, various intrusive methods are known for this purpose. In such methods, as the name suggests, the system to be tested intervenes by occupying a transmission channel and transmitting a reference signal therein. The quality assessment is then carried out by comparing the known reference signal with the received signal, for example subjectively by one or a plurality of test persons. However, this is expensive and therefore expensive.

In the EP 0 980 064 A further intrusive method for the machine-aided quality assessment of an audio signal is described, wherein a spectral similarity value of the known source signal and the received signal is determined in order to evaluate the transmission quality. This similarity value is based on a calculation of the covariance of 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, the system to be tested must be intervened. In order to determine the signal quality, at least one transmission channel must be occupied and a reference signal transmitted therein. This transmission channel can not be used for data transmission during this time. In addition, while it is in principle possible for a broadcasting system such as a broadcasting service to occupy the signal source for the transmission of test signals, as this would occupy all channels and the test signal would be transmitted to all receivers, this approach is extremely impractical. Intrusive techniques are also unsuitable for simultaneously monitoring the quality of a variety of transmission channels.

EP-A-644 526 discloses a non-intrusive noise reduction method which uses an estimate of noise energy to calculate the desired signal information. US-A-5,848,384 shows a method and apparatus for determining the quality of an audio signal.

Presentation of the invention

The object of the invention is to specify a method of the abovementioned type, which avoids the disadvantages of the prior art and, in particular, offers a possibility for assessing the signal quality of a signal transmitted via a telecommunication network without knowledge of the originally transmitted signal.

The solution of the problem is defined by the features of method claim 1 and the device claim. In the method according to the invention for machine-aided determination of a quality measure of an audio signal, a reference signal is first determined from the audio signal. By means of comparing the determined reference signal with the audio signal, a quality value is determined which is used to determine the quality measure.

The inventive method thus allows an assessment of the quality of an audio signal at any terminal of the telecommunications network. Ie. it also allows the quality assessment of many transmission channels simultaneously, even allowing simultaneous evaluation of all channels. The quality assessment is carried out solely on the basis of the properties of the received signal, d. H. without knowledge of the source signal or the signal source.

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

An audio signal transmitted via a telecommunication network typically has, in addition to the desired signal information, also undesired components, such as various noise components, which were not present in the original source signal.

In order to be able to carry out a quality assessment that is as good as possible, the best possible estimate of the originally transmitted signal is necessary. To reconstruct this reference signal, there are several methods. One possibility is to determine an estimate of the characteristics of the transmission channel and quasi backwards from the received signal. Another possibility is a direct estimation of the reference signal based on the known information about the received signal and the transmission channel.

In the presently applied method, the reference signal is determined by estimating the noise components present in the received signal and then removing them from the received signal. By removing the noise components from the audio signal, a denoised audio signal is first determined, which is preferably used as a reference signal for assessing the transmission quality.

There are several ways to remove noise from the received audio signal. The audio signal could, for example, be passed through appropriate filters. However, a preferred method of removing the noise from the audio signal uses a neural network.

However, the audio signal is not used directly as an input signal. First, a discrete wavelet transform (DWT) is applied to the audio signal. This transformation provides a plurality of DWT coefficients of the audio signal which are fed to the neural network as an input. The neural network provides at the output a plurality of corrected DWT coefficients, from which the reference signal is obtained with the inverse DWT. This corresponds to the noisy version of the audio signal.

To accomplish this, the coefficients of the neural network must be set to provide the DWT coefficients of a noisy input signal the DWT coefficients of the corresponding noisy input signal. For the neural network to provide the desired coefficients, it must first be trained with a set of corresponding noisy or noisy signal pairs.

In this way, both stationary noise such as white, thermal and vehicle or road noise, and impulse noise can be suppressed. Even echo interference and interference can be suppressed or eliminated with the neural network.

In determining the quality measure, any other information besides the quality value, which is determined by the comparison of the received audio signal with the reference signal determined therefrom, can also be taken into account. This may be information contained in the audio signal as well as information about the transmission channel or the telecommunication network itself.

It is advantageous in the determination of the quality measure to use information which can be obtained by suitable means from the received audio signal itself. Thus, the quality of the received audio signal is influenced, for example, by the codecs (coder-decoder) passed through during the transmission. It is difficult to detect such signal degradation because, for example, if the codec bit rates are too small, some of the original signal information is lost. However, too small codec bit rates result in a change in the fundamental frequency (pitch) of the audio signal, which is why it is advantageous to examine the course and dynamics of the fundamental frequency in the audio signal. Since such changes can be most easily examined on the basis of audio signal sections with vowels, signal components in the audio signal with vowels are preferably first detected 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 be gone partially desired information has lost. For example, the received audio signal may have more or less long signal interruptions.

However, the closer the reference signal generated from the audio signal is to the original source signal, the more accurate is the assessment of the transmission quality. This is the reason for replacing signal interruptions with suitable signals. For this purpose, for example, suitable noise signals or already transmitted signal sections could be used.

However, in order to obtain as accurate an estimate of the reference signal as possible, it is advantageous to first detect such signal discontinuities in the audio signal and then to replace the missing signal sections with the most accurate estimates achieved by interpolation. The type of interpolation of the lost signal sections depends on the length of the signal interruption. For short breaks, d. H. with interruptions up to a few samples in the audio signal is preferably a polynomial and medium-length interruptions, d. H. from a few to a few dozen samples, model-based interpolation is preferred.

Longer signal interruptions, d. H. Interruptions from a few tens of samples, however, can hardly be reasonably reconstructed. Instead of considering this information as redundant and discarding it, and in some cases also the information about the short and medium signal interruptions, are taken into account, preferably in the assessment of the transmission quality. They are included in the calculations when determining the quality measure.

The received audio signal may include various types of audio signals. For example, it can contain voice, music, noise or silence signals. Of course, the quality assessment may be based on all or part of these signal components. In a preferred variant of the invention, however, the assessment of the signal quality is limited to the speech signal components. With an audio discriminator, therefore, the audio signal components first become from the audio signal extracted and only these speech signal components for determining the quality measure, ie used to determine the reference signal. In this case, to determine the quality value, the determined reference signal is, of course, not compared with the received audio signal, but only with the speech signal component extracted therefrom.

The inventive device for machine-aided determination of a quality measure of an audio signal comprises first means for determining a reference signal from the audio signal, second means for determining a quality value by comparing the determined reference signal with the audio signal and third means for determining the quality measure taking into account the quality value.

The first means for determining a reference signal from the audio signal may comprise a plurality of modules. Thus, a noise suppression module and / or an interruption detection and interpolation module is preferably provided.

The noise suppression module suppresses noise signal components in the received audio signal. It includes the means for performing the wavelet transforms described above as well as the neural network for determining the new DWT coefficients. The interrupt detection and interpolation module has those means which are required on the one hand for detecting signal interruptions in the audio signal and on the other hand for polynomial interpolation of short as well as model-based interpolation of medium-length signal interruptions. The thus determined reference signal thus corresponds to a noisy version of the received audio signal and typically has only greater signal interruptions.

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

The more information about the audio signal is included in the determination of 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 distortions. These include, for example, a vocal detection module with which signal components with vowels can be detected in the audio signal. These vocal signal components are passed on to an evaluation module, which uses these signal components to determine information about codec-related signal distortions, which are also used to assess the signal quality. The third means are correspondingly designed such that this information about the codec-related signal distortions can be taken into account in the determination of the quality measure.

Advantageously, however, not the entire audio signal but only its voice signal components are used for quality assessment. According to the method already described, the device therefore has, in particular, fifth means for extracting the speech signal components from the audio signal. Accordingly, not the audio signal itself, but only its voice signal component is denoudized and examined for interruptions in order to determine the reference signal. Likewise, of course, not the audio signal, but only the voice signal component is compared with this reference signal. Thus, the determination of the quality measure takes place only on the basis of the information in the speech signal component, wherein the information from the remaining signal components is not taken into account.

From the following detailed description and the totality of the claims, further advantageous embodiments and feature combinations of the invention result.

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 embodiment show:

Fig. 1

a schematically illustrated block diagram of the inventive method;

Fig. 2

the noise suppression module in the operating state;

Fig. 3

the noise suppression module in the training state;

Fig. 4

the neural network of the noise suppression module and

Fig. 5

an example of an audio signal with an interruption.

Basically, the same parts are provided with the same reference numerals in the figures.

Ways to carry out the invention

FIG. 1 shows a block diagram of the inventive method. Here, a quality measure 2 is determined for an audio signal 1, which can be used, for example, for the evaluation of the used (not shown) telecommunications network. The audio signal 1 is understood here to mean the signal which a receiver receives after the transmission via the telecommunication network. This audio signal 1 is true typically not coincide with the signal transmitted by the transmitter (not shown) because on the way from the transmitter to the receiver, the transmission signal is varied in a variety of ways. For example, it goes through various modules such as speech coders and decoders, multiplexers and demultiplexers, as well as speech enhancers and echo cancellers. But even the transmission channel itself can have a large impact on the signal, which, for example, in the form of interference, fading, Übertragungsab- or interruptions, echoes, etc. express.

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

In the illustrated example, however, the assessment of the signal quality does not take place on the basis of the entire audio signal 1, but only on the basis of the speech component contained therein. The audio signal 1 is first examined with an audio discriminator 3 to speech signal components 4 out. Found speech signal components 4 are forwarded for further processing, whereas other signal components such as music 5.1, pauses 5.2 or strong signal interference 5.3 can be sorted out and otherwise processed or discarded. To make this distinction, the audio signal is 1 piecewise, d. H. to bits a each about 100 ms to 500 ms, passed to the audio discriminator 3. This splits these bits further into individual buffers of about 20 ms in length, processes these buffers and then allocates them to one of the different signal groups speech signal, music, pause or strong interference.

The audio discriminator 3 uses, for example, an LPC (linear predictive coding) transformation for the evaluation of the signal chips, with which the coefficients of an adaptive filter corresponding to the human voice tract are calculated. The assignment of the signal chips to the different signal groups is based on the shape of the transmission characteristics of this filter.

In order to be able to judge the quality of the transmission, a reference signal 6, d. H. a possible best estimate of the transmission signal originally transmitted by the transmitter, determined. This reference signal estimation takes place in several stages.

In a first stage, a noise suppression module 7, unwanted signal components such as stationary noise or impulse noise are first removed from the speech signal component 4 or suppressed. This is done with the aid of a neural network, which has been previously trained by means of a plurality of noisy signals as input and each of the corresponding noise-free version of the input signal as a target signal. The thus-obtained noisy voice signal 11 is forwarded to the second stage.

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

In the present example, the detection of signal interruptions by means of an examination of discontinuities of the signal fundamental frequency (pitch-tracing). The interpolation is carried out as a function of the length of the detected interruption. For short breaks, d. H. Interruptions of a few samples in length are polynomial-interpolated, such as Lagrange, Newton, Hermite, or Cubic Spline interpolation. For medium-length breaks (a few to a few tens of samples), model-based interpolations such as maximum a posteriori, autoregressive, or frequency-time interpolation are used. For longer signal interruptions, interpolation or other signal reconstruction is usually no longer possible in a meaningful way.

The whole thing is hampered by the fact that there are both different types of breaks - there is a difference between syllable breaks and proper signal breaks - as well as different types of techniques for handling such breaks in the transmission channel. Thus, a terminal, for example, depending on information about the transmission network, respond differently to missing frames. For example, in a first method, lost frames are simply replaced by zeros. In a second method, instead of the lost frames other, correctly received frames are used and in a third method locally generated noise signals, so-called "comfort noise" are used instead of the lost frames.

After determining the reference signal 6 with the noise suppression module 7 and the interruption detection and interpolation module 8, it is compared with the speech signal component 4 with the aid of the comparison module 9. For this comparison, an algorithm can be used, as used, for example, in intrusive methods for the comparison of the known source signal with the received signal. For example, psychoacoustic models that compare signals perceptually, ie perceptibly, are suitable. The result of this comparison is an intrusive quality value 10. To determine this intrusive quality value 10, the input signals, ie the speech signal component 4 and the reference signal 6, are decomposed into signal pieces of about 20 to 30 ms in length and a partial quality value is calculated for each signal piece. After about 20 to 30 signal pieces, which corresponds to a signal duration of 0.5 seconds, the intrusive quality value 10 is determined as the arithmetic mean of these partial quality values. The intrusive quality value 10 forms the output signal of the comparison module 9.

In determining the quality measure 2, however, other information about the audio signal 1 can be taken into account in addition to the information about interference signal components or signal interruptions. For example, a voice coder or speech decoder, which the transmitted signal has passed through on its way from the transmitter to the receiver, can have an influence on the audio signal 1. These influences include, for example, varying both the fundamental frequency and the higher harmonic frequencies of the signal. The smaller the bit rate of the speech codecs used, the greater the frequency shifts and thus the signal distortions.

Such influences can be examined most easily with vowels, which is why the noise signal 11 which has been deafened is first supplied to a vocal detector 12. This includes, for example, a neural network that has been previously trained for the recognition of particular (single or all) vowels. Vocal signals 13, d. H. Signal components that recognize the neural network as vowels are forwarded to an evaluation module 14, other signal components are discarded.

The evaluation module 14 divides the vocal signal 13 into signal pieces of about 30 ms and calculates thereon a DFT (discrete Fourier transform) with a frequency resolution of about 2 Hz at a sampling frequency of about 8 kHz. This allows the fundamental and higher harmonic frequencies to be determined and examined for variations. Another feature for evaluating codec-related distortions is the dynamics of the signal spectrum, with smaller dynamics implying poorer signal quality. The reference values for the dynamics evaluation are obtained for the individual vowels from example signals. A codec quality value 15 is derived from the information about the influence of codecs on the frequency shifts and the spectrum dynamics of the audio signal 1 and of the denoised speech signal 11.

In the determination of the quality measure 2 by the evaluation module 16, an interruption quality value 17 is taken into account in addition to the intrusive quality value 10 and the codec quality value 15. This value contains information about the length and the number of interruptions detected by the interruption detection and interpolation module 8, whereby in a preferred embodiment of the invention only the information about the long interruptions is taken into account. In addition, of course, further quality information 18 about the received audio signal 1 or the noisy speech signal 11, which are determined with other modules or examinations, can be included in the calculations of the quality measure 2.

The individual quality values are now scaled to lie in the range of numbers between 0 and 1, where a quality value of 1 denotes 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, wherein the individual weighting coefficients are determined experimentally and determined such that their sum is 1.

If further quality-relevant information is available via the telecommunications network or if new effects occur in the transmission channels, it is possible in a simple manner to add further modules for calculating further quality values and to take this into account when determining the quality standard 2 in the described manner ,

The following are based on the FIGS. 2 to 5 some of the modules explained in more detail. FIG. 2 shows the noise suppression module 7. The voice signal portion 4 of the audio signal 1 is first subjected to a known DWT 19 (Discrete Wavelet Transformation). DWT's are similar to DFT's used for signal analysis. An essential difference, however, is the use of so-called wavelets, ie temporally limited and temporally localized waveforms with a mean value of 0, in contrast to the sinusoidal or cosine wave forms used indefinitely and thus not temporally localized in a DFT.

The speech signal component 4 is divided into signal pieces of about 20 ms to 30 ms, which are each subjected to the DWT 19. The result of the DWT 19 is a set of DWT coefficients 20.1, which are input to a neural network 20 as an input vector. Its coefficients were previously trained to provide a new set of DWT coefficients 20.2 of the noisy version of this signal for a given set of DWT coefficients 20.1 of a noisy signal. This new set of DWT coefficients 20.2 will now be sent to IDWT 21, i. H. subjected to the DWT 19 inverse DWT. In this way, this IDWT 21 delivers a majority of the unencumbered version of the speech signal components 4, namely the desired, denoised speech signal 11.

The training configuration of the neural network 20 is in FIG. 3 shown. It is trained with pairs of noisy and noisy versions of sample signals. An unencumbered example signal 22.1 is subjected to the DWT 19 and a first set 20.3 of DWT coefficients is obtained. The noisy sample signal 22.2 is also subjected to the same DWT 19 and a second set 20.4 of DWT coefficients is generated, which is fed into the neural network 20. The output vector of the neural network 20, the new DWT coefficients 20.5, is compared in a comparator 23 with the first set 20.3 of DWT coefficients. Due to the differences between these two sets of DWT coefficients, there is a correction 24 of the coefficients of the neural network 20. This process is repeated with a plurality of example signal pairs so that the coefficients of the neural network 20 perform the desired function more and more precisely. Advantageously, the training of the neural network 20 uses example signals 22.1, 22.2, which represent human sounds from different languages. It is also an advantage to use both female and male and child voices. The mentioned size of the individually processed signal pieces of 20 ms to 30 ms duration is selected so that the processing of the speech signal component 4 can be performed independently of the speech and the speaker. Even pauses in speech and very quiet signal sections are trained so that they are recognized correctly.

In the present embodiment, as the neural network 20, a multi-layer perceptron having an input layer 25, a hidden layer 26, and an output layer 27 was used. The perceptron was trained with a backpropagation 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 a plurality of output neurons 27.1. Each input neuron 25.1 is in each case supplied to one of the DWT coefficients 20. 1 of the preceding DWT 19. After the input signals have passed through the neural network, determining the respective values with the set coefficients of the respective neurons and calculating the value combinations in the individual neurons, each output neuron 27.1 provides one of the new DWT coefficients 20.2. As already mentioned, the audio discriminator 3 splits the signal bits into individual buffers of length 20 ms. At a sampling rate of 8 kHz, this corresponds to 160 samples. For this case, for example, a neural network 20 each having 160 input and output neurons 25.1, 27.1 and about 50 to 60 hidden neurons 26.1 may be used.

Based on FIG. 5 The interpolation of a signal interruption should be briefly described. For example, a time-frequency interpolation is used for the signal reconstruction. First, a short-term spectrum for 64-sample-length (8 ms) signed frames is calculated. This is done by multiplying the signal frames by Hamming windows with a 50% overlap.

The goal of interpolation is to address this gap. First, a frequency-time transformation is performed. This results in a three-dimensional signal representation which provides the power spectrum in the z-axis direction for each point in the time-frequency plane (x-y plane). An interruption at a given time t is easy to recognize as zero points along the line x = t in the time-frequency plane.

FIG. 5 shows such a signal 28 of about 200 samples in length. In order to recognize the periodicity easier, shows FIG. 5 the signal 28 in the temporal domain. On the abscissa axis 32, the number of samples and on the ordinate axis 33, the magnitudes are plotted. However, the interpolation is done in the frequency-time domain. In FIG. 5 the interruption 29 is easy to recognize as a gap of almost 10 samples.

For each frequency component, a polynomial interpolation is now carried out for both the phase and magnitude, with minimal phase and magnitude discontinuity. For this purpose, again the pitch period 30 of the signal 28 is determined. Information from the samples before and after the gap within this pitch period 30 is taken into account for the interpolation. The signal areas 31.1, 31.2 show those areas of the signal 28 each before a pitch period before or after the interruption 29. These signal areas 31.1, 31.2 are not identical to the original signal piece at the interruption 29, but still show a high degree of similarity , For small gaps up to about 10 samples, it is assumed that there is still enough signal information to be able to perform a correct interpolation. For longer gaps, additional information from samples of the environment can be used.

In summary, it should be noted that the invention allows the signal quality of a received audio signal to be assessed without knowing the original transmission signal. From the signal quality can of course be concluded on the quality of the transmission channels used and thus on the service quality of the entire telecommunications network. The fast response times of the inventive method, which are in the order of about 100 ms to 500 ms, thus allow various applications such as general comparisons of service quality of different networks or subnets, quality-based cost allocation or quality-based routing in a network or across multiple networks by means of appropriate control of network nodes (gateways, routers, etc.).

Claims (10)

1. A method for the machine-assisted determination of a measure of quality of an audio signal, in which a reference signal that represents an estimate of an audio signal originally transmitted is determined from the audio signal, and a quality value, which is used for determining the measure of quality, is determined by means of comparing the reference signal with the audio signal, characterised in that a de-noised audio signal is determined by removing noise signal components from the audio signal and is used as the reference signal, and that signal components with vowels are detected in the de-noised audio signal information on codec-related signal distortions is determined therefrom and is taken into consideration in determining the measure of quality.
2. A method according to claim 1, characterised in that the de-noised audio signal is determined by subjecting the audio signal to discrete wavelet transformation, feeding the coefficients of the latter into a previously trained neural network and subjecting the output signals of the latter to inverse discrete wavelet transformation.
3. A method according to either of claims 1 and 2, characterised in that signal interruptions in the audio signal are detected and the reference signal is determined by at least partially reconstructing it in the case of the signal interruptions, the reference signal being reconstructed preferably by polynomial interpolation in the case of short signal interruptions and preferably by model-based interpolation in the case of medium-length signal interruptions.
4. A method according to claim 3, characterised in that information on the signal interruptions is taken into consideration in determining the measure of quality.
5. A method according to any one of claims 1 to 4, characterised in that, before the reference signal is determined, a speech signal component is extracted from the audio signal and the determination of the measure of quality is restricted to the speech signal component.
6. A device for the machine-assisted determination of a measure of quality of an audio signal, which 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 measure of quality while taking the quality value into consideration, the reference signal representing an estimate of an audio signal originally transmitted, characterised in that it has means for removing noise signal components from the audio signal and means for determining codec-related signal distortions, the latter means including a vowel detection module for detecting vowel signal components in the de-noised audio signal and an evaluation module for determining the codec-related signal distortions, the third means being so designed that the codec-related signal distortions can be taken into consideration in determining the measure of quality.
7. A device according to claim 6, characterised in that the first means have a noise suppression module for suppressing noise signal components and/or an interruption detection and interpolation module for detection and interpolation of signal interruptions in the audio signal, and the third means are so designed that signal interruptions can be taken into consideration in determining the measure of quality.
8. A device according to either of claims 6 and 7, characterised in that it has means for extracting a speech signal component from the audio signal and is designed for the purpose of determining the measure of quality of the speech signal component.
9. A device according to claim 7, wherein the first means have the noise suppression module, characterised in that the noise suppression module has means for performing discrete wavelet transformation for calculating signal coefficients of an audio signal, a neural network for calculating corrected signal coefficients and means for performing inverse discrete wavelet transformation of the corrected signal coefficients for determining the audio signal without noise signal components.
10. A device according to claim 7, wherein the first means have the interruption detection and interpolation module, characterised in that the interruption detection and interpolation module has means for detecting signal interruptions in an audio signal and means for interpolating signal interruptions of the audio signal, the latter means preferably being designed for the purpose of polynomial interpolation of short signal interruptions and model-based interpolation of medium-length signal interruptions.

Images