EP2153438B1 - Post-processing for reducing quantification noise of an encoder during decoding - Google Patents

Abstract

The invention relates to the processing of a signal that is compression encoded (COD) according to a predetermined encoding type applying a quantification operation (Q) and then decoded (DEC) so that the quantification noise is present in the decoded signal (S*). The signal processing of the invention comprises applying a quantification noise reduction (TBQ) to the decoded signal (S), preferably in the following manner: first obtaining information (INF) on the type of compression encoding, selecting a model for the reduction of the quantification noise adapted to said information by estimating the quantification noise (BQ) that the encoding may have generated; and applying to the decoded signal (S*) a processing for reducing the quantification noise (FIL) according to the selected model.

Description

The present invention relates to signal processing, in particular digital audio signals in the field of telecommunications, these signals being for example speech, music, or other signals.

Generally, the rate needed to pass an audio and / or video signal with sufficient quality is an important parameter in telecommunications. In order to reduce this parameter and then increase the number of possible communications via the same network, audio coders have been developed in particular to compress the amount of information necessary to transmit a signal.

Some encoders achieve particularly high information compression rates. Such coders generally use advanced information modeling and quantification techniques. Thus, such encoders transmit only models or partial data of the signal.

The decoded signal, although not identical to the original signal (since some of the information has not been transmitted due to the quantization operation), nevertheless remains very close to the original signal. The difference, from a mathematical point of view, between the decoded signal and the original signal is then called "quantization noise". We can also speak of "distortion" introduced by the coding / decoding.

Signal compression processes are often designed to minimize quantization noise and, in particular, to make this quantization noise as audible as possible when dealing with an audio signal. There are then techniques taking into account the psycho-acoustic characteristics of hearing, in order to "hide" this noise. However, to obtain flow rates as low as possible, the noise may remain audible, sometimes, which in some circumstances degrades the intelligibility of the signal.

In order to reduce this noise, two families of techniques are usually used.

It is possible, first of all, to use a perceptual post-filter, of the type used, for example, in CELP type speech decoders (for "Code Excited Linear Prediction"). This is to perform a filtering that improves the subjective quality at the price of a distortion. Indeed, a signal attenuation is applied in the areas where the quantization noise is most audible (especially between the formants). Current perceptual post-filters provide good results for speech signals, but poorer results for other types of signals (music signals, for example).

• " Adaptive Postfiltering for Quality Enhancement of Coded Speech", Chen J.H., Gersho A., IEEE Trans. On Speech and Audio Proc., (janvier 1995 ).

Indeed, a post-filter for improving the coded speech is described in particular in the document Chen et al:

• " Adaptive Postfiltering for Quality Enhancement of Coded Speech, "Chen JH, Gersho A., IEEE Trans., Speech and Audio Proc., (January 1995). ).

• une section à « long terme » renforce les harmoniques (harmoniques de la fréquence fondamentale) et creuse les vallées spectrales entre ces harmoniques, et
• une section à « court terme » renforce les formants et creuse également les vallées spectrales entre ces formants.

The model described is based on a division into two sections:

• a "long-term" section reinforces the harmonics (harmonics of the fundamental frequency) and digs the spectral valleys between these harmonics, and
• a "short-term" section reinforces the formants and also widens the spectral valleys between these formants.

Harmonics and formants are well-known spectral characteristics of speech but applying this type of processing to a signal other than speech generates strong distortions. For example, the spectral richness of a music signal can not be handled with such a simple signal model.

Thus, perceptual post-filters can generate distortions, because they rely on a model that is not precise enough. Moreover, the perceptual post filter is usually ineffective in periods of silence. These problems have been observed experimentally by the Applicant who first sought to integrate this type of perceptual post-filters in decoders that are not CELP type, for example in decoders within the meaning of the G.711 standard. or the G.722 standard.

The document US2003 / 0182104 describes the modification of a digital audio signal in a decoding step based on a psychoacoustic model. Such a modification would be applicable to the perceptually encoded signals provided that the quantization noise distribution can be derived from the encoded data.

Another treatment family targets conventional noise reduction treatments to distinguish the useful signal from the noise. This type of processing therefore makes it possible to reduce the noise related to the environment of the signal capture and it is often used for speech signals. However, here, it is impossible to make transparent the treatment vis-à-vis the noise related to the environment of the sound recording, which poses a problem for music signal coding, in particular. Thus, in coding / decoding one may want to transmit the ambient noise and it is then desirable that the noise reduction does not apply to this type of noise.

The present invention improves the situation.

To this end, it proposes a method of processing a signal that has been coded in compression according to a predetermined type of coding, applying a quantization operation, and then decoding. The process according to the invention is defined in claim 1.

Here, the term "noise reduction processing" is understood to mean an operation of the type described above which consists in extracting the useful signal from a signal to be processed, by filtering the parasitic signals, for example by defining a gain function intervening in a filter applied to the decoded signal. Here, the quantization noise is filtered.

It is therefore a classic denoising but applied here to reduce the quantization noise. This denoising is in no way related to a perceptual post-filter of the type described in Chen et al, which relies completely on the characteristics and the dynamics of the signal, while the noise reduction processing in the sense of the The invention relies instead on the determination of the quantization noise.

Thus, there is provided a type of noise reduction processing specific to each type of compression coding performed. The very way of estimating the characteristics of the noise reduction filter (type of gain function, gain function parameters, etc.) depends on the type of coding performed.

It will be seen in particular in the embodiment examples given below that the quantization noise itself strongly depends on the type of coding performed. It will be seen that it is possible to establish a variation of the quantization noise as a function of a variation of the decoded signal, and that this variation of the quantization noise is specific to the type of coding implemented.

• on estime, à partir des informations sur le type de codage, une variation du bruit de quantification en fonction d'au moins un paramètre du signal décodé, et
• en fonction d'une valeur courante de ce paramètre dans le signal décodé, on estime le bruit de quantification pour déterminer la fonction de filtrage à appliquer au signal décodé ayant cette valeur courante de paramètre.

So :

• it is estimated from the information on the type of coding a variation of the quantization noise according to at least one parameter of the decoded signal, and
• based on a current value of this parameter in the decoded signal, the quantization noise is estimated to determine the filtering function to be applied to the decoded signal having this current parameter value.

• un modèle de variation d'un rapport signal à bruit de quantification, en fonction d'au moins un paramètre du signal décodé, et/ou
• une coloration spectrale du bruit de quantification (c'est-à-dire une variation spectrale du bruit de quantification en fonction des caractéristiques du signal décodé).

It will therefore be understood that the information on the type of compression coding is a priori information , independent of the characteristics of the signal and that advantageously, it can be deduced:

• a variation model of a signal-to-quantization noise ratio, as a function of at least one parameter of the decoded signal, and / or
• spectral staining of the quantization noise (i.e. spectral variation of the quantization noise as a function of the characteristics of the decoded signal).

In one possible embodiment, the prior information on the type of compression coding is obtained during an encoder declaration procedure.

The invention is particularly suitable in the case where the type of compression coding is a coding according to the G.711 standard.

The present invention also provides a device for processing a signal initially coded in compression according to a predetermined type of coding, and then decoded. The device is defined in claim 6.

More generally, the device advantageously comprises means for implementing the method described above.

It is advantageous for such a device to be integrated in a decoder, downstream from a decoding unit, as illustrated in FIG. figure 1 representing a TBQ device of the aforementioned type downstream of the DEC decoding unit. This figure 1 will be described in detail later.

The present invention also relates to a computer program, intended to be stored in memory of a processing device of the aforementioned type, and comprising instructions for calculating the quantization noise, as well as parameters of a quantization noise reduction filter. when these instructions are executed by a processor of the processing device.

• le bruit de quantification présent dans le signal décodé est calculé,
• et les paramètres du post-filtre sont calculés en correspondance de ce bruit de quantification, pour limiter, voire supprimer, ce bruit.

An advantageous embodiment may consist in providing a set of instructions for each type of coding implemented and, in each set of instructions, defining a variation of the quantization noise as a function of the decoded signal. Thus, upon receipt of the information a priori, a set of appropriate instructions is selected. With this instruction set:

• the quantization noise present in the decoded signal is calculated,
• and the parameters of the post-filter are calculated in correspondence of this quantization noise, to limit or even eliminate this noise.

The instructions on the variation of the quantization noise can be programmed offline, on the basis of observations (theoretical or experimental according to the exemplary embodiments which will be described later) made on the type of coding used. The way in which these instructions are carried out, itself, will be described in detail later, with reference to figures 2 and 5 which can then constitute flowcharts of a computer program within the meaning of the invention.

Thus, the invention proposes a post-processing performed after decoding and which uses information a priori on the characteristics of the quantization operation performed by the encoder. The type of processing (or "processing model" according to the generic terms above) that will be chosen to process the signal is independent of the characteristics of the signal itself. Of course, the processing itself (in particular the estimation of the gain function) may depend on the signal, for example its energy or its power. On the other hand, whether dealing with a music signal, a speech signal or any other signal (harmonic, impulsive, etc.), the type of treatment is the same and is not based, for example, on only on the energy of a decoded frame received. Indeed, it is possible to know theoretically the characteristics of the quantization noise, in particular according to different families of encoders. For the purposes of the invention, this information is then used to estimate quantities that are exploited to define at least one gain function of a noise reduction unit that operates downstream of a decoding unit.

Thus, the invention makes it possible to reduce the quantization noise (and therefore the distortion) that a compression encoder of the signal implementing a quantization operation usually introduces.

According to one of the advantages offered by the present invention, it is possible to keep the same coding / decoding structure without making any changes and yet ensure a better quality of the decoded signal, without increasing the amount of information to be transmitted by the coder.

According to another advantage, the invention advantageously reduces the quantization noise alone, even during periods of silence, and this, for any type of signal.

According to yet another advantage, the implementation of the invention does not perform a conventional noise reduction and therefore does not modify the noise related to the environment of the capture of the signal.

We especially remembered that the implementation of the invention can reduce or eliminate quantization noise without distorting the signal and this for any type of signal, just using a priori information on the type of encoder used (for example the characteristics of the encoder compression model, the characteristics of the quantizer, or other).

The present invention finds an advantageous application to the field of speech and music processing, and more generally to the processing of the signal, especially images, since any encoder is required to introduce a quantization noise.

More generally, the invention applies to all areas where it is sought to reduce a quantization noise of a signal.

• la figure 1 illustre schématiquement la structure générale d'une unité de traitement au sens de l'invention,
• la figure 2 illustre schématiquement les étapes d'un procédé au sens de l'invention,
• la figure 3 illustre une variation de la loi de compression (dite « loi A ») des amplitudes, dans un codage selon la norme G.711 pour illustrer un exemple de réalisation de l'invention,
• la figure 4 illustre la variation du rapport signal à bruit de quantification RSB en fonction du facteur de charge, cette variation étant tirée de la variation illustrée sur la figure 3,
• la figure 5 illustre les étapes d'un exemple de traitement dans le cas d'un codage selon la norme G.711, basé notamment sur les observations des variations des figures 3 et 4,
• la figure 6 illustre un exemple du spectre du signal (courbe en pointillés) et du spectre du bruit de quantification (courbe continue) pour un codage selon la norme G.722,
• la figure 7 illustre un exemple de forme d'onde d'un signal de parole S^* (courbe de dessus) et le rapport signal à bruit de quantification correspondant RSB (courbe de dessous), pour un codage/décodage selon la norme G.722,
• la figure 8 est un nuage de points illustrant pour chaque segment de 80 échantillons la corrélation entre le rapport signal à bruit RSB et l'énergie du signal, dans une application à un codage/décodage selon la norme G.722,
• la figure 9 montre les segments de signal (en noir) où l'erreur de l'estimation du rapport signal à bruit de quantification RSB est supérieure à 6 dB tandis que le rapport RSB est inférieur à 25 dB, dans l'application à un codage/décodage selon la norme G.722,
• la figure 10 reprend le nuage de point représentant, pour chaque segment, l'énergie du bruit en fonction de l'énergie du signal, en illustrant ici l'estimation du niveau de bruit (ligne en traits mixtes), la zone ou l'erreur de l'estimation est inférieure à 6 dB (lignes en traits pointillés), et la délimitation pour laquelle le rapport RSB est supérieur 25 dB (ligne en trait plein).

Other features and advantages of the invention will appear on examining the detailed description below, and the attached drawings in which:

• the figure 1 schematically illustrates the general structure of a processing unit within the meaning of the invention,
• the figure 2 schematically illustrates the steps of a process within the meaning of the invention,
• the figure 3 illustrates a variation of the compression law (called "law A") of the amplitudes, in a coding according to the G.711 standard to illustrate an embodiment of the invention,
• the figure 4 illustrates the variation of the signal-to-quantization ratio RSB as a function of the load factor, this variation being drawn from the variation illustrated in FIG. figure 3 ,
• the figure 5 illustrates the steps of an example of processing in the case of a coding according to the G.711 standard, based in particular on the observations of the variations of the Figures 3 and 4 ,
• the figure 6 illustrates an example of the signal spectrum (dotted curve) and the quantization noise spectrum (continuous curve) for G.722 coding,
• the figure 7 illustrates an example of a waveform of a speech signal S ^* (top curve) and the corresponding quantization noise-to-noise ratio RSB (bottom curve), for coding / decoding according to the G.722 standard,
• the figure 8 is a scatterplot illustrating for each segment of 80 samples the correlation between the SNR signal-to-noise ratio and the signal energy, in an application to G.722 coding / decoding,
• the figure 9 shows the signal segments (in black) where the error of the RSB quantization signal to noise ratio estimate is greater than 6 dB while the RSB ratio is less than 25 dB, in the coding / decoding application according to G.722,
• the figure 10 resumes the point cloud representing, for each segment, the noise energy as a function of the signal energy, illustrating here the estimate of the noise level (phantom line), the zone or the error of the noise. estimate is less than 6 dB (dashed lines), and the bound for which the RSB is greater than 25 dB (solid line).

• codé en compression par un codeur COD de type connu et appliquant notamment une opération de quantification Q au signal S,
• transmis via un canal de transmission CA, puis
• décodé par un décodeur DEC homologue du codeur COD.

We first refer to the figure 1 on which a signal S is:

• encoded in compression by a COD encoder of known type and in particular applying a quantization operation Q to the signal S,
• transmitted via a CA transmission channel, then
• decoded by a decoder DEC homologous coder COD.

The signal thus decoded, denoted S ^* , then has a quantization noise which is defined mathematically as a difference (S ^* - S) with respect to the original signal S.

With reference again to the figure 1 for the purposes of the invention, a quantization noise reduction processing unit TBQ is provided, downstream of the decoder DEC, for suppressing or at least limiting the quantization noise in the signal S ^* .

For this purpose, the unit TBQ comprises at least one input E to receive decoder DEC information INF on the type of coding / decoding implemented, which then allows to choose a noise reduction treatment model to be implemented. artwork. In particular, it is estimated from the signal received and decoded S ^* and, depending on the type of coding / decoding that has been implemented, the influence of the quantization noise in the received signal S ^* . For this purpose, a calculation module is provided to give an estimate of the quantization noise BQ, on the basis of the model chosen and as a function of the received signal S ^* . This calculation module can typically be in the form of a combination of a processor and a working memory (not shown). From the estimated quantization noise BQ, the estimated noise BQ is simply processed by applying a conventional filtering FIL to the signal S ^* to finally output a processed signal S ^* _T. It should be further emphasized that the parameters PAR of the FIL filter applied to the signal S ^* (for example a gain function for the signal filtering) are determined to reduce in particular the estimated quantization noise BQ.

Indeed, with reference to the figure 2 from the INF information received on the type of coding / decoding employed (step S2), determining a model (step S3) of noise reduction processing. It will be seen in the embodiments described below that the quantization noise reduction model chosen may be different, for example depending on whether the signal has been coded / decoded according to the G.711 standard or coded / decoded according to the standard. G.722.

Thus, when the signal is received in successive blocks (or frames marked TRi in step S1), it is estimated (step S4) a quantization noise level specific to the chosen model. As will be seen in the examples below, it is advantageous to estimate the quantization noise level from the calculation of the signal-to-quantization noise ratio (denoted RSB). This RSB information depends on the decoded signal S ^* , but also the type of coding implemented. Thus, prior knowledge of the coding, by obtaining the information INF allows, together with certain statistical characteristics of the signal S ^* , to estimate here the signal-to-noise ratio RSB.

This step S4 therefore requires knowing a priori the type of encoder that has been used, information that can be obtained for example during a declaration procedure of the encoder called "encoder transaction", which is supposed to be acquired.

The type of encoder, the characteristics of its compression model and its quantizer Q make it possible to estimate an evolution of the signal-to-quantization noise ratio, as a function of certain statistical parameters of the signal, for example its variance, its spectral density of power, or others. This relationship between the signal-to-quantization noise ratio and the statistical parameters of the signal involves coder-specific laws which will be described later, for some exemplary embodiments.

The necessary statistical parameters can be calculated by classical quantity estimators (eg variance). Based on these estimates, an estimate of the signal to noise quantization ratio can be extrapolated. The estimates can be made indifferently in the time domain, frequency, or any other time-frequency domain (wavelet for example).

Again with reference to figure 2 the next step S5 consists in calculating the filter parameters for the reduction of the quantization noise in the received signal S ^* . Knowledge of the signal-to-noise ratio makes it possible to deduce the expression of a quantization noise reduction filter, this filter being hereinafter called "post-filter" (downstream of the decoder). It is indeed possible to deduce the expression of a digital filter whose purpose is to reduce a noise whose most characteristics are known a priori (its power spectral density for example) and whose level is determined from the estimation of the quantization signal-to-noise ratio obtained at step previous S4. For example, the calculation of the filter can be carried out in the frequency domain and implement any short-term spectral attenuation technique (a spectral subtraction, a Wiener filter, or other). The calculation of the post-filter in step S5 can be performed in the time domain, frequency domain, or any other time-frequency domain.

Finally, the noise reduction processing step S6, itself, here amounts to filtering the decoded signal S ^* by the post-filter calculated in step S5. This step S6 can be performed in the time or frequency domain, according to the constraints related to the implementation and the estimation domain of the PAR parameters and the RSB ratio in the previous steps. Finally, a TRi 'frame processed by denoising the quantization noise in step S7 is obtained.

An exemplary implementation of the invention for coding / decoding according to the G.711 standard (according to the European law called "law A") is described below.

du signal, des niveaux de saturation x _max déterminés par la dynamique et bien entendu du nombre de bits b utilisés pour la représentation des échantillons, selon une expression du type : $$\mathit{RSB}=3\frac{{\sigma }_x^2}{x_{\max }^2}{2}^{2b},$$

soit, en dB : $$\mathit{RSB}=10\log \left(3\frac{{\sigma }_x^2}{x_{\max }^2}{2}^{2b}\right)=\left(20\log 2\right)b+10\log 3-20\mathrm{log\; \Gamma }\left[\mathrm{dB}\right]$$

The traditional digital representation of one-dimensional signals involves uniform quantization of samples. Thus, in the absence of quantizer overflow, the quantization signal-to-noise ratio (RSB) depends on the variance ${\mathrm{\sigma }}_x^2$

of the signal, saturation levels x _max determined by the dynamics and of course the number of bits b used for the representation of the samples, according to an expression of the type: $$\mathit{RSB}=3\frac{{\sigma }_x^2}{{\mathrm{<figure-callout\; id="2"\; label="x\; max"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">x\; </figure-callout>}}_{\max }^{}}{2}^{2b},$$

in dB: $$\mathit{RSB}=10\mathrm{<figure-callout\; id="3"\; label="log"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">log</figure-callout>}\left(3\frac{{\sigma }_x^2}{{\mathrm{<figure-callout\; id="2"\; label="x\; max"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">x\; </figure-callout>}}_{\max }^{}}{2}^{2b}\right)=\left(20\mathrm{<figure-callout\; id="2"\; label="log"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">log</figure-callout>}2\right)b+10\log 3-20\mathrm{log\; \Gamma }\left[\mathrm{dB}\right]$$

représente un paramètre dit "facteur de charge", qui détermine la qualité d'utilisation de la dynamique du quantificateur disponible par le signal, où :

• x _max est le niveau numérique d'amplitude maximum possible d'un échantillon selon le quantificateur choisi, et
• σ _x est l'écart-type du signal (la racine carrée de la variance) qui, pour un bloc complet d'échantillons (ou « trame »), peut être estimé par la racine carrée de la puissance moyenne Pm du signal sur ce bloc.

The height $\mathrm{\Gamma }=\frac{x_{\max }}{{\mathrm{\sigma }}_x}$

represents a parameter called "load factor", which determines the quality of use of the dynamics of the quantizer available by the signal, where:

• x _max is the maximum possible numerical level of amplitude of a sample according to the quantizer chosen, and
• σ _x is the standard deviation of the signal (the square root of the variance) which, for a complete block of samples (or "frame"), can be estimated by the square root of the average power Pm of the signal on that block.

Expression (1) is strongly dependent on the value of this parameter Γ. In particular, it can be seen that the maximum signal-to-noise ratio is obtained for a full-scale signal and decreases rapidly if the signal amplitude decreases.

The low bit rate limits of the uniform quantization law led to the development of a quantization law whose signal to quantization noise ratio was approximately independent of the signal variance for wide signal dynamics. This is what the logarithmic log-quantization law of the G.711 standard (called "A-law" in Europe, or "μ-law" in North America) achieves.

Law A in use in Europe is defined by a dependent expression of the x value of the quantized sample, as follows: $$F\left(x\right)=\{\begin{array}{c}\frac{\mathrm{AT}\left|x\right|/{\mathrm{<figure-callout\; id="1"\; label="x\; max"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">x\; </figure-callout>}}_{\max }}{1+\ln \mathrm{AT}}\mathrm{sgn}x,0\le \left|x\right|/x_{\max }<{\mathrm{AT}}^{-1}\\ x_{\max }\frac{1+\ln \left(\mathrm{AT}\left|x\right|/{\mathrm{<figure-callout\; id="1"\; label="x\; max"\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">x\; </figure-callout>}}_{\max }\right)}{1+\ln \mathrm{AT}}\mathrm{sgn}x,{\mathrm{AT}}^{-1}\le \left|x\right|/{\mathrm{<figure-callout\; id="1"\; label="x\; max\; \le "\; filenames="imgf0002.png,imgf0004.png"\; state="\{\{state\}\}">x\; </figure-callout>}}_{\max }\le 1\end{array}$$

With reference to the figure 3 , the first variation of the compression law (0≤ | x | / x _max < A ^-1 ) is linear, generates a uniform law of quantification and is hereinafter called "uniform variation", while the second variation of the compression law ( A ^{- 1} ≤ | x | / x _max ≤ 1) is logarithmic, and hereinafter called "logarithmic variation".

European law uses a value of A = 87.56 (which numerically satisfies the equation A / (1 + 1n A ) = 16).

From these observations, it is possible to calculate the signal-to-quantization noise ratio for A-law compression as follows.

For low intensity signals (uniform part of the compression law), law A provides a higher quantization signal-to-noise ratio (in dB) of 10log ( A / (1 + ln A )) than that obtained by quantization uniform on the same number of levels, the expression of which is given by: $$\begin{array}{ll}{\mathit{RSB}}_{\mathit{Uni}}& \approx \left(20\mathrm{<figure-callout\; id="2"\; label="log"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">log</figure-callout>}2\right)b+10\mathrm{<figure-callout\; id="3"\; label="log"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">log</figure-callout>}3+10\log \left[\mathrm{AT}/\left(1+\ln \mathrm{AT}\right)\right]-20\log \left(\mathrm{\Gamma }\right)\left[\mathrm{dB}\right]\\ {\mathit{RSB}}_{\mathit{Uni}}& \approx 6.02b+4.77+10\log \left[\mathrm{AT}/\left(1+\ln \mathrm{AT}\right)\right]-20\log \left(\mathrm{\Gamma }\right)\left[\mathrm{dB}\right]\\ {\mathit{RSB}}_{\mathit{Uni}}& \approx 67.97-20\log \left(\mathrm{\Gamma }\right)\left[\mathrm{dB}\right]\mathrm{for}b=8\end{array}$$

For signals of greater amplitude (logarithmic part of the compression law), the quantization signal-to-noise ratio is constant and equals 38.16 dB (for b = 8 bits): $$\begin{array}{ll}{\mathit{RSB}}_{\log }& =\left(20\mathrm{<figure-callout\; id="2"\; label="log"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">log</figure-callout>}2\right)b+10\log 3-20\log \left(1+\ln \mathrm{AT}\right)\left[\mathrm{dB}\right]\\ {\mathit{RSB}}_{\log }& \approx 6.02b-10\left[\mathrm{dB}\right]\\ {\mathit{RSB}}_{\log }& \approx 38.16\mathrm{dB\; for}b=8\end{array}$$

• une première partie croissante, correspondant à la variation uniforme de la loi de compression, et
• une partie suivante, constante, correspondant à la variation logarithmique de cette loi.

The figure 4 represents the evolution of the signal-to-quantization noise ratio RSB for a law A with b = 8 bits. We immediately identify:

• an increasing first portion, corresponding to the uniform variation of the compression law, and
• a following part, constant, corresponding to the logarithmic variation of this law.

• le rapport signal à bruit de quantification qui est donné par les équations (3) et (4) précédentes, et
• l'information bien connue selon laquelle ce bruit est "blanc" pour ce type de codage.

To deal with the reduction of quantization noise introduced by a coding according to the G.711 standard, two information is used here:

• the signal-to-noise ratio of quantization given by equations (3) and (4) above, and
• the well known information that this noise is "white" for this type of coding.

The implementation of the quantization noise reduction processing is based on the exploitation of this information a priori. In particular, it requires an estimation of the load factor Γ, the parameter on which the power of the quantization noise depends, as follows.

• si le facteur de charge Γ est tel que -20.log(Γ) > -20.log(Γ_s) = 38.16-64.97∼=-27dB (flèche o en sortie du test T54), alors le rapport signal à bruit de quantification est constant et vaut RSB_M ∼= +38dB (plateau de la figure 4), comme fixé à l'étape S55,
• sinon (flèche n en sortie du test T54), alors le rapport signal à bruit de quantification RSB peut être calculé selon une variation linéaire en fonction du facteur de charge tirée de l'équation (3) : $$\mathrm{RSB}=f\left(\mathrm{\Gamma }\right)=65-20\log \left(\mathrm{\Gamma }\right)\left[\mathrm{dB}\right]$$
  

comme fixé à l'étape S56.With reference to the figure 5 the average power Pm of a current block TRi (step S52) is estimated, and hence the load factor Γ, varying as the inverse of the square root of the average power (step S53). It is considered that the numerator x _max of the load factor is here constant (at constant saturation level). At the T54 test, the found value of the load factor Γ is compared with that of a threshold Γ _s defining the point of inflection of the compression law ( figure 4 ), as following :

• if the load factor Γ is such that -20.log (Γ)> -20.log (Γ _s ) = 38.16-64.97~ = -27dB (arrow o at the output of the T54 test), then the signal-to-noise ratio of quantization is constant and is worth RSB _M ~ = + 38dB (plateau of the figure 4 ), as set in step S55,
• if not (arrow n at the output of the test T54), then the signal-to-quantization noise ratio RSB can be calculated according to a linear variation as a function of the load factor derived from equation (3): $$\mathrm{RSB}=f\left(\mathrm{\Gamma }\right)=65-20\log \left(\mathrm{\Gamma }\right)\left[\mathrm{dB}\right]$$
  

as set in step S56.

• g(RSB) =f _w = RSB / (RSB + 1), où, ici, la valeur RSB ne s'exprime pas en dB mais en valeur naturelle.

The gain function (step S57) is then evaluated for post-filter application (step S58). By way of purely illustrative example, a Wiener filter can be provided as gain function g (RSB). The expression of the Wiener filter f _w can be given by the value of the RSB quantization signal-to-noise ratio calculated previously, taking into account, of course, its frequency dependence with:

• g (RSB) = f _w = RSB / (RSB + 1), where, here, the RSB value is not expressed in dB but in natural value.

• un seuillage du post-filtre, et/ou
• un détecteur d'activité vocale pour des signaux de parole (avec un traitement de réduction de bruit de quantification plus léger pendant les périodes d'inactivité vocale).

Advantageously, it is possible to reduce the noise reduction processing, in particular for signals of low signal-to-quantization noise ratio, and therefore of low amplitude (for load factors such as -20.log (Γ) < -50dB on the figure 4 ), possibly including:

• a threshold of the post-filter, and / or
• a voice activity detector for speech signals (with lighter quantization noise reduction processing during periods of voice inactivity).

It is indicated that a variant of the treatment presented here is to reduce the quantization noise, sample by sample, rather than a treatment by successive blocks. In this case, the load factor is directly given by the amplitude level of the sample (inverse of the square root of the amplitude) and the continuation of the treatment is similar to that presented above.

Another possible application of the invention to a different type of coding, here the coding according to the G.722 standard, is described.

The ITU-T G.722 coding, standardized in 1988 for 64 kbit / s digital audio conferencing applications, is still very widely used. It is a hierarchical coding / decoding at three rates: 64, 56 and 48 kbit / s. The signal is divided into two subbands by a filter called QMF (for "Quadrature Mirror Filter"). The two bands obtained are encoded with an ADPCM encoder (for "Adaptive Differential Pulse Code Modulation").

The high band is coded on 2 bits per sample. The difference between the three bit rates comes from the low band which is coded on 6 bits per sample for the highest bitrate, but it is possible to reserve the last or the last two bits for data transmission.

The quality of the higher bit rate is very good, but the coding noise becomes very audible and annoying for the lowest bit rate at 48 kbit / s. The quantization noise reduction processing in the sense of the invention can be advantageously applied in this case.

Already, the characteristics of the quantization noise can be efficiently estimated from the decoded signal. As illustrated by figure 6 , the quantization noise spectrum (solid line curve) is always flat, regardless of the signal spectrum (dotted line curve). The signal to quantization noise ratio depends on the average signal strength and its nature. On the figure 7 it can be observed that the signal-to-quantization noise ratio (RSB) correlates well with the average power of the signal S ^* . In the example shown, the RSB ratio was estimated on segments of 80 samples (5 ms for a sampling frequency of 16 kHz).

The representation in the form of point clouds of the figure 8 illustrates even better the correlation between the average power of the signal (abscissa axis) and the signal-to-quantization noise ratio (y-axis), calculated by segments of 80 samples.

où CST est une constante qui vaut, dans l'exemple de la figure 8, environ 10 dB.From this observation, we can deduce a first simple rule for estimating the RSB ratio as a function of the average power P _moy of the segment (correlation line represented in dashed lines on the figure 8 ), given by: $$\mathrm{RSB}\mathrm{=}{\mathrm{P}}_{\mathrm{Avg}}\mathrm{-}\mathrm{CST}\left[\mathrm{dB}\right]$$

where CST is a constant that is worth, in the example of the figure 8 , about 10 dB.

It will be understood from this expression that the average noise power, determined experimentally here, is constant CST = 10 dB, and this, independently of the average signal power, so that the RSB ratio increases well with the average power of the signal.

The best estimate of the RSB quantization signal-to-noise ratio is obtained for the low signal levels, ie when the RSB is low (and therefore when the noise is most audible). However, some segments have points well below the dashed line and the use of this simple rule is suboptimal. It has been observed, however, that these zones correspond to strong RSB ratios, where the quantization noise is already probably masked by the useful signal.

In general, it has been observed that the treatment according to the invention applied here nevertheless achieves an advantageous reduction of the quantization noise.

In the case where the simple rule of equation (5) is used, the figure 9 represents in black on a gray background the areas of the signal where the estimation error of the RSB ratio is greater than 6 dB, and the RSB ratio itself is less than 25 dB, that is to say the signal areas wherein the estimator underestimates the quantization noise, resulting in a lower efficiency of the quantization noise reduction processing. It can nevertheless be noted that these zones correspond to unvoiced signal segments, for which the quantization noise is less troublesome because of the intrinsically noisy nature of the signal.

We have shown on the figure 10 a power diagram of the noise with respect to a signal power, in accordance with the empirical equation (5). The dashed line represents the estimate of the noise power. The dashed lines delimit the area where the error of the estimate is less than 6 dB. Below the solid line, the RSB is greater than 25 dB. The black dots (relative to the other gray dots) correspond to the black segments of the figure 9 .

It is thus shown that a very simple estimate of the RSB ratio based solely on the energy of the decoded signal can give good results for ADPCM type coding / decoding. The estimation of the RSB ratio can be further refined by taking into account, for example, the prediction gain of the ARMA (autoregressive) filters which intervene in the G.722 decoder.

Knowing the spectral shape of the quantization noise and its energy, the quantization noise reduction process of the invention can be effectively applied for this type of coding / decoding. This example is obviously valid for other types of coding / decoding of the same family as those standardized G.726 or G.727.

Of course, the present invention is not limited to the embodiment described above by way of example; it extends to other variants.

Thus, it has been shown above that an advantageous application of the invention can for example aim to reduce the quantization noise of a standardized ITU-G.711 encoder by using the properties of the quantization law implemented. , especially according to law A in Europe. Indeed, in this application, the quantization noise is white and it is possible to estimate the quantization signal-to-noise ratio and hence a gain function that makes it possible to reduce this noise. An advantageous application of the invention thus aims at the reduction of quantization noise in the processing at the extended band extension of the G.711 coder (ITU-T SG16, G.711WB).

However, the treatment of the case of law A has been given above as an example. Similarly, it could have been described the example of the law μ (part of the G.711 standard applied in the United States).

More generally, the invention applies to any type of coding / decoding as long as its intrinsic characteristics are known.

Claims (8)

1. Method for processing a digital audio signal,
   said signal having been:

   - compression encoded (COD) according to a predetermined encoding type, applying a quantization operation,

   - then decoded (DEC),

   the processing method comprising:

   - an estimate (S4) of a quantization noise introduced by the compression encoding based on information (INF) obtained a priori on the type of compression encoding, and

   - a determination (S5) of a filtering function to be applied to the decoded signal in order to apply (S6) an estimated quantization noise reduction process (TBQ),

   characterized in that:

   - based on said information (INF), a variation (fig. 4) of the quantization noise (RSB) is estimated as a function of at least one parameter relative to a load parameter (Γ) of the decoded signal, and

   - according to a current value of said parameter (Γ) in the decoded signal (S52, S53), the quantization noise is estimated (S55; S56) in order to determine the filtering function (S57) to be applied (S58) to the decoded signal having said current parameter value (Γ).
2. Method according to Claim 1, characterized in that deduced from said a priori information is a variation model (fig. 4) of a signal-to-quantization noise ratio (RSB), as a function of said parameter (Γ) of the decoded signal.
3. Method according to Claim 2, characterized in that a spectral coloration of the quantization noise is deduced from said a priori information and account is also taken of said spectral coloration in order to determine the filtering function to be applied to the decoded signal.
4. Method according to one of Claims 1 to 3, characterized in that said a priori information is obtained during an encoder declaration procedure.
5. Method according to one of Claims 1 to 4, characterized in that the compression encoding type is an encoding according to the G.711 standard.
6. Device (TBQ) for processing a digital audio signal that is initially compression encoded according to a predetermined encoding type, then decoded, the processing device (TBQ) comprising:

   - means for estimating a quantization noise (BQ) introduced by the compression encoding, based on the decoded signal and information (INF) obtained a priori on the type of compression encoding, and

   - means for determining a filtering function to be applied to the decoded signal in order to apply (S6) an estimated quantization noise reduction process (FIL),

   characterized in that means for estimating estimate:

   - based on said information (INF), a variation (fig. 4) of the quantization noise (RSB) as a function of at least one parameter relative to a load parameter (Γ) of the decoded signal, and

   - according to a current value of said parameter (Γ) in the decoded signal (S52, S53), the quantization noise in order to determine the filtering function (S57) to be applied (S58) to the decoded signal having said current parameter value (Γ).
7. Device according to Claim 6, characterized in that it is incorporated into a decoder, downstream of a decoding unit (DEC).
8. Computer program, designed to be stored in the memory of a device (TBQ) for processing a digital audio signal that is initially compression encoded according to a predetermined encoding type, then decoded, characterized in that it comprises instructions for implementing the method according to one of Claims 1 to 5, when these instructions are executed by a processor of the processing device.

Images