Method and device for determining the quality of a speech signal

A.

BACKGROUND OF THE INVENTION

The invention lies in the area of quality measurement of sound signals, such as audio, speech and voice signals. More in particular, it relates to a method and a device for determining, according to an objective measurement technique, the speech quality of an output signal as received from a speech signal processing system, with respect to a reference signal according to the preamble of claim 1 and claim 10, respectively. Method and device of such type are known, e.g., from References [1,-,5] (for more bibliographic details on the References, see below under C. References). Methods and devices, which follow the ITU-T Recommendation P.861 and its recently accepted successor Draft New Recommendation P.862 (see References [6] and [7]), are also of such a type. According to the present known technique, an output signal from a speech signals-processing and/or transporting system, such as wireless telecommunications systems, Voice over Internet Protocol transmission systems, and speech codecs, which is generally a degraded signal and whose signal quality is to be determined, and a reference signal, are mapped on representation signals according to a psycho-physical perception model of the human hearing. As a reference signal, an input signal of the system applied with the output signal obtained may be used, as in the cited references. Subsequently, a differential signal is determined from said representation signals, which, according to the perception model used, is representative of a disturbance sustained in the system present in the output signal. The differential or disturbance signal constitutes an expression for the extent to which, according to the representation model, the output signal deviates from the reference signal. Then the disturbance signal is processed in accordance with a cognitive model, in which certain properties of human testees have been modelled, in order to obtain a time-independent quality signal, which is a measure of the quality of the auditive perception of the output signal.

The known technique, and more particularly methods and devices which follow the Draft Recommendation P.862, have, however, the disadvantage that severe distortions as caused by extremely weak or silent portions in the degraded signal, and which are not present in the reference signal, may result in a quality signal, which possesses a poor correlation with subjectively determined quality measurements, such as mean opinion scores (MOS) of human testees. Such distortions may occur as a consequence of time clipping, i.e. replacement of short portions in the speech or audio signal by silence e.g. in case of lost packets in packet switched systems. In such cases the predicted quality is significantly higher than the subjectively perceived quality.

B.

SUMMARY OF THE INVENTION

The main object of the present invention is to provide for an improved method and corresponding device for determining the quality of a speech signal, which do not possess said disadvantage.

The present invention has been based on the following observation. The gain of a system under test is generally not known a priori. Therefore in an initialisation or pre-processing phase of the main step of processing the output (degraded) signal and the reference signal a scaling step is carried out, at least on the output signal by using a scaling factor for an overall or global scaling of the power of the output signal to a specific power level. The specific power level may be related to the power level of the reference signal in techniques such as following Recommendation P.861, or to a predefined fixed level in techniques which may follow Draft Recommendation P.862. The scaling factor is a function of the reciprocal value of the square root of the power of the output signal. In cases in which the degraded signal includes extremely weak or silent portions, this reciprocal value increases to large numbers, which can be used to adapt the distortion calculation in such a manner that a much better prediction of the subjective quality of systems under test is possible. The present invention aims to provide better controllable scaling factor and overall scaling step.

To this end a method and a device of the above kinds are, according to the invention, characterised as in claim 1 and in claim 9, respectively.

Further preferred embodiments of the method and the device of the invention are summarised in the various subclaims.

C.

REFERENCES

<dl id="dl0001" compact="compact"><dt>[1]</dt><dd>Beerends J.G., Stemerdink J.A., "A perceptual speech-quality measure based on a psychoacoustic sound representation", J.Audio Eng. Soc., Vol. 42, No. 3, Dec. 1994, pp. 115-123;</dd><dt>[2]</dt><dd>WO-A-96/28950;</dd><dt>[3]</dt><dd>WO-A-96/28952;</dd><dt>[4]</dt><dd>WO-A-96/28953;</dd><dt>[5]</dt><dd>WO-A-97/44779;</dd><dt>[6]</dt><dd>ITU-T Recommendation P.861, "Objective measurement of Telephone-band (330-3400 Hz) speech codecs", 06/96;</dd><dt>[7]</dt><dd>ITU-T Pre-published Recommendation P.862, "Perceptual evaluation of speech quality (PESQ), an objective method for end-to-end speech quality assessment of narrow-band telephone networks and speech codecs", March 2001.</dd></dl>

All References are considered as being incorporated into the present application.

D.

BRIEF DESCRIPTION OF THE DRAWING

The invention will be further explained by means of the description of exemplary embodiments, reference being made to a drawing comprising the following figures: <dl id="dl0002" compact="compact"><dt>FIG. 1</dt><dd>schematically shows a known system set-up including a device for determining the quality of a speech signal;</dd><dt>FIG. 2</dt><dd>shows in a block diagram a detail of a known device for determining the quality of a speech signal;</dd><dt>FIG. 3</dt><dd>shows in a block diagram a similar detail as shown in FIG. 2 of another known device;</dd><dt>FIG. 4</dt><dd>shows in a block diagram a similar detail as shown in FIG. 2 or FIG. 3, according to the invention;</dd><dt>FIG. 5</dt><dd>shows in a block diagram a device for determining the quality of a speech signal according to the invention, including a variant of the detail as shown in FIG. 4.</dd></dl>

E.

DESCRIPTION OF EXEMPLARY EMBODIMENTS

FIG. 1 shows schematically a known set-up of an application of an objective measurement technique which is based on a model of human auditory perception and cognition, and which follows the ITU-T Recommendation P.861 or the pre-published Recommendation P.862, for estimating the perceptual quality of speech links or codecs. It comprises a system or telecommunications network under test 10, hereinafter referred to as system 10 for briefness' sake, and a quality measurement device 11 for the perceptual analysis of speech signals offered. A speech signal X₀(t) is used, on the one hand, as an input signal of the network 10 and, on the other hand, as a first input signal X(t) of the device 11. An output signal Y(t) of the network 10, which in fact is the speech signal X₀(t) affected by the network 10, is used as a second input signal of the device 11. An output signal Q of the device 11 represents an estimate of the perceptual quality of the speech link through the network 10. Since the input end and the output end of a speech link, particularly in the event it runs through a telecommunications network, are remote, for the input signals of the quality measurement device use is made in most cases of speech signals X(t) stored on data bases. Here, as is customary, speech signal is understood to mean each sound basically perceptible to the human hearing, such as speech and tones. The system under test may of course also be a simulation system, which simulates a telecommunications network. The device 11 carries out a main processing step which comprises successively, in a pre-processing section 11.1, a step of pre-processing carried out by pre-processing means 12, in a processing section 11.2, a further processing step carried by first and second signal processing means 13 and 14, and, in a signal combining section 11.3, a combined signal processing step carried out by signal differentiating means 15 and modelling means 16. In the pre-processing step the signals X(t) and Y(t) are prepared for the step of further processing in the means 13 and 14, the pre-processing including power level scaling and time alignment operations. The further processing step implies mapping of the (degraded) output signal Y(t) and the reference signal X(t) on representation signals R(Y) and R(X) according to a psycho-physical perception model of the human auditory system. During the combined signal processing step a differential or disturbance signal D is determined by the differentiating means 15 from said representation signals, which is then processed by modelling means 16 in accordance with a cognitive model, in which certain properties of human testees have been modelled, in order to obtain the quality signal Q.

Recently it has been experienced that the known technique, and more particularly the one of Pre-published Recommendation P.862, has a serious shortcoming in that severe distortions as caused by extremely weak or silent portions in the degraded signal, and which are not present in the reference signal, may result in quality signals Q, which predict the quality significantly higher than the subjectively perceived quality and therefore possess poor correlations with subjectively determined quality measurements, such as mean opinion scores (MOS) of human testees. Such distortions may occur as a consequence of time clipping, i.e. replacement of short portions in the speech or audio signal by silence e.g. in case of lost packets in packet switched systems.

Since the gain of a system under test is generally not known a priori, during the initialisation or pre-processing phase a scaling step is carried out, at least on the (degraded) output signal by using a scaling factor for scaling the power of the output signal to a specific power level. The specific power level may be related to the power level of the reference signal in techniques such as following Recommendation P.861. Scaling means 20 for such a scaling step has been shown schematically in FIG. 2. The scaling means 20 have the signals X(t) and Y(t) as input signals, and signals X_S(t) and Y_S(t) as output signals. The scaling is such that the signal X(t) = X_S(t) is unchanged and the signal Y(t) is scaled to Y_S(t) = S₁.Y(t) in scaling unit 21, using a scaling factor:<maths id="math0001" num="{1}"><math display="block"><mrow><msub><mrow><mtext>S</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><mtext> = S(X,Y) = √</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">X</mtext><mtext>)/</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">Y</mtext><mtext>)</mtext></mrow></math><img file="EP1241663A1_D0001.tif" /></maths> In this formula <i>P</i>_{<i>average</i>}<i>(X)</i> and <i>P</i>_{<i>average</i>}<i>(Y)</i> mean the time-averaged power of the signals X(t) and Y(t), respectively.

The specific power level may also be related to a predefined fixed level in techniques which may follow Pre-published Recommendation P.862. Scaling means 30 for such a scaling step has been shown schematically in FIG. 3. The scaling means 30 have the signals X(t) and Y(t) as input signals, and signals X_S(t) and Y_S(t) as output signals. The scaling is such that the signal X(t) is scaled to X_S(t) = S₂.X(t) in scaling unit 31 and the signal Y(t) is scaled to Y_S(t) = S₃.Y(t) in scaling unit 32, respectively using scaling factors:<maths id="math0002" num="{2}"><math display="block"><mrow><msub><mrow><mtext>S</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><msub><mrow><mtext> = S(P</mtext></mrow><mrow><mtext>f</mtext></mrow></msub><mtext>,X) = √</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">fixed</mtext></mrow></msub><mtext>/</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">X</mtext><mtext>)</mtext></mrow></math><img file="EP1241663A1_D0002.tif" /></maths> and<maths id="math0003" num="{3},"><math display="block"><mrow><msub><mrow><mtext>S</mtext></mrow><mrow><mtext>3</mtext></mrow></msub><msub><mrow><mtext> = S(P</mtext></mrow><mrow><mtext>f</mtext></mrow></msub><mtext>,Y) = √</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">fixed</mtext></mrow></msub><mtext>/</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">Y</mtext><mtext>)</mtext></mrow></math><img file="EP1241663A1_D0003.tif" /></maths> in which <i>P</i>_<i>fixed</i> (i.e. P_f) is a predefined power level, the so-called constant target level, and <i>P</i>_{<i>average</i>}<i>(X)</i> and <i>P</i>_{<i>average</i>}<i>(Y)</i> have the same meaning as given before.

In both cases scaling factors are used, which are a function of the reciprocal value of the square root of the power of the output signal, i.c. S₁ and S₃, or of the power of the reference signal, i.c. S₂. In cases in which the degraded signal and/or the reference signal includes extremely weak or silent portions, these reciprocal values may increase to very large numbers. This fact provides a starting point for making the used scaling factors and corresponding scaling operations adjustable and consequently better controllable.

In order to achieve such a better controllability at first an adjustment parameter Δ is added to each time-averaged signal power value as used in the scaling factor or factors, respectively in the first and second one of the two described cases. The adjustment parameter Δ has a predefined adjustable value in order to increase the denominator of each scaling factor to a larger value. The scaling factor(s) thus modified are used in the scaling step, hereinafter called first scaling step, of the initialisation phase in a similar way as previously described with reference to FIGs. 2 and 3. Secondly a further scaling factor is determined which equals to the modified scaling factor, as used for scaling the output signal, but raised to an exponent α. The exponent α is a second adjustment parameter having values between zero and 1. This further scaling factor is used in a further scaling step, hereinafter called second scaling step. It is possible to carry out the second scaling step on various stages in the quality measurement device. Hereinafter three different ways are described with reference to FIG. 4 and FIG. 5.

FIG. 4 shows schematically a scaling arrangement 40 for carrying out the first scaling step using modified scaling factors and the second scaling step. The scaling arrangement 40 have the signals X(t) and Y(t) as input signals, and signals X'_S(t) and Y'_S(t) as output signals. The first scaling step is such that the signal X(t) is scaled to X_S(t) = S'₂.X(t) in scaling unit 41 and the signal Y(t) is scaled to Y_S(t) = S'₃.Y(t) in scaling unit 42, respectively using modified scaling factors:<maths id="math0004" num="{1'}"><math display="block"><mrow><msub><mrow><mtext>S'</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><mtext> = S(Y+Δ) = √(</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">X</mtext><mtext>)+Δ)/(</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">Y</mtext><mtext>)+Δ)</mtext></mrow></math><img file="EP1241663A1_D0004.tif" /></maths> for cases having a scaling step in accordance with FIG. 2, in which X_S(t) = X(t) (i.e. S(X+Δ)=1 in FIG. 4), and<maths id="math0005" num="{2'}"><math display="block"><mrow><msub><mrow><mtext>S'</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><mtext> = S(X+Δ) = √</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">fixed</mtext></mrow></msub><mtext>/(</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">X</mtext><mtext>)+Δ)</mtext></mrow></math><img file="EP1241663A1_D0005.tif" /></maths> and<maths id="math0006" num="{3'},"><math display="block"><mrow><msub><mrow><mtext>S'</mtext></mrow><mrow><mtext>3</mtext></mrow></msub><mtext> = S(Y+Δ) = √</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">fixed</mtext></mrow></msub><mtext>/(</mtext><msub><mrow><mtext mathvariant="italic">P</mtext></mrow><mrow><mtext mathvariant="italic">average</mtext></mrow></msub><mtext>(</mtext><mtext mathvariant="italic">Y</mtext><mtext>)+Δ)</mtext></mrow></math><img file="EP1241663A1_D0006.tif" /></maths> for cases having a scaling step in accordance with FIG. 3. The second scaling step is such that the signal X_S(t) is scaled to X'_S(t) = S₄.X_S(t) in scaling unit 43 and the signal Y_S(t) is scaled to Y'_S(t) = S₄.Y_S(t) in scaling unit 44, using scaling factor:<maths id="math0007" num="{4}"><math display="block"><mrow><msub><mrow><mtext>S</mtext></mrow><mrow><mtext>4</mtext></mrow></msub><msup><mrow><mtext> = S</mtext></mrow><mrow><mtext>α</mtext></mrow></msup><mtext>(Y+Δ)</mtext></mrow></math><img file="EP1241663A1_D0007.tif" /></maths> The scaling factor S₄ may be generated by the scaling unit 42 and passed to the scaling units 43 and 44 of the second scaling step as pictured. Otherwise the scaling factor S₄ may be produced by the scaling units 43 and 44 in the second scaling step using the scaling factor S₃ as received from the scaling unit 42 in the first scaling step.

The values of the parameters α and Δ are adjusted in such a way that for test signals X(t) and Y(t) the objectively measured qualities have high correlations with the subjectively perceived qualities (MOS). Thus examples of degraded signals with replacement speech by silences up to 100% appeared to give correlations above 0.8, whereas the quality of the same examples as measured in the known way showed values below 0.5. Moreover there appeared indifference for cases for which the Pre-published Recommendation P.862 was validated.

The values for the parameters α and Δ may be stored in the pre-processor means of the measurement device. However, adjusting of the parameter Δ may also be achieved by adding an amount of noise to the degraded output signal at the entrance of the device 11, in such a way that the amount of noise has an average power equal to the value needed for the adjustment parameter Δ in a specific case.

Instead of in the pre-processing phase the second scaling step may be carried out in a later stage during the processing of the output and reference signals. However the location of the second scaling step does not need to be limited to the stage in which the signals are processed separately. The second scaling step may also be carried out in the signals combining stage, however with different values for the parameters α and Δ. Such is pictured in FIG. 5, which shows schematically a measurement device 50 which is similar as the measurement device 11 of FIG. 1, and which successively comprises a pre-processing section 50.1, a processing section 50.2 and a signal combining section 50.3. The pre-processing section 50.1 includes the scaling units 41 and 42 of the first scaling step, the unit 42 producing the scaling factor S4 indicated in the figure by S^αi(Y+Δ_i), in which i=1,2 for a first and a second case, respectively.

In the first case (i=1) the second scaling step is carried out, in the signal combining section 50.3, by scaling unit 51 and using the scaling factor S4 = S^α1(Y+Δ₁), thereby scaling the differential signal D to a scaled differential signal D'= S^α1(Y+Δ₁)·D. Alternatively, in the second case (i=2) the second scaling step is carried out, again in the signal combining section 50.3, by scaling unit 52 and using the scaling factor S4 = S^α2(Y+Δ₂), thereby scaling the quality signal Q to a scaled quality signal Q'= S^α2(Y+Δ₂)·Q. For the parameters α_i and Δ_i the same applies as what has been mentioned previously in relation to the parameters α and Δ.