Method and device for coding an audio signal by "forward" and "backward" lpc analysis

The invention relates to a method and a device for coding an audio signal, such as a speech signal, by LPC analysis ".<i>before</i>"and"<i>back</i>".

At the present time, the techniques of coding of the audio-frequency signals, in particular the speech signals, aim to allow the transmission of these signals in digital form, under conditions of reduction of the transmission rate, in order, in particular, to ensure appropriate management of the networks for the transmission of these signals, taking into account the significant increase in transactions between users.

Among the coding techniques used, that designated by LPC analysis, for <i>"Linear Predictive Coding"</i> in Anglo-Saxon language, consists in carrying out a linear prediction of the audio frequency signal to be coded, the coding being carried out temporally by means of a linear prediction filtering applied to successive blocks of this signal.

In the aforementioned techniques, that known under the name of CELP coding, for <i>"Code Excited</i> Linear <i>Prediction</i>", is the most widespread and one of the most effective. Other techniques, such as the technique designated by MP-LPC, for"<i>Multi Pulse Linear Predictive Coding</i>", or the VSELP technique, for"<i>Vector Sum Excited Linear Prediction</i>"in Anglo-Saxon language, are relatively close to CELP coding.

The abovementioned coding techniques are said to be "by analysis by synthesis". In particular, they made it possible, for audiofrequency signals belonging to the telephone frequency band, to reduce the transmission rate of these signals from 64 kb / s (MIC coding) to 16 kb / s using the coding technique CELP, and even up to 8 kb / s in the case of coders implementing the most recent evolutions of this coding technique, without perceptible degradation of the quality of the speech restored after transmission and decoding.

A particularly important field of application of these coding techniques is, in particular, that of mobile telephony. In this field of application, the necessary limitation of the frequency band granted to each mobile telephone operator and the very rapid increase in the number of user subscribers make it necessary to correspondingly reduce the coding speed, while the requirements of the users in terms of speech quality keep growing. Other fields of application of these coding techniques relate, for example, to the storage of digital data representative of these signals on storage media, high-quality telephony for applications of video or audio conference, multimedia, or digital satellite transmissions.

The linear prediction filters used in the abovementioned techniques are obtained using an analysis module called "LPC analysis" operating on successive blocks of the digital signal. These filters are capable, according to the order of analysis, that is to say according to the number of coefficients of the filter, of modeling more or less faithfully the contours of the frequency spectrum of the signal to be coded. In the case of a speech signal, these contours are called formants. However, for good quality coding, required by most current applications, the filter thus defined is not sufficient to perfectly model the signal. It is then essential to proceed to the coding of the linear prediction residue. Such an operating mode relating to the linear prediction residue is notably implemented by the LD-CELP coding technique, for<i>Low Delay CELP</i> in Anglo-Saxon language, previously mentioned in the description. The residual signal is in this case modeled by a waveform extracted from a stochastic dictionary and multiplied by a gain value. The MP-LPC coding technique, for example, models this residue using variable position pulses assigned respective gain values, while the VSELP coding technique performs this modeling by a linear combination of pulse vectors extracts from appropriate directories. A didactic reminder of the operating mode of the LPC analysis and in particular of the LPC analysis "<i>back</i>"and LPC analysis <i>"before"</i> or LPC analysis <i>"backward"</i> and LPC analysis <i>"forward"</i> respectively in Anglo-Saxon language, will first be given below.

The general envelope of the frequency spectrum is modeled by means of a short-term synthesis filter, constituting the LPC filter, the coefficients of which are evaluated by means of a linear prediction of the speech signal to be coded. This LPC filter, autoregressive filter, has a shape transfer function, relation (1):<maths id="math0001" num=""><img file="EP0906613B1_D0001.tif" /></maths> where p denotes the number of coefficients a_i of the filter and the order of the linear prediction implemented, z denoting the variable of the z-transform of the frequency space.

A method for evaluating the coefficients a_i consists in applying a criterion for minimizing the energy of the prediction error signal of the speech signal over the analysis length of the latter. The analysis length for a digital speech signal formed by successive samples is in practice a number N of these samples, constituting a coding frame. The energy of the prediction error signal then checks the relation (2):<maths id="math0002" num=""><img file="EP0906613B1_D0002.tif" /></maths> where s (n) denotes the sample of rank n in the frame of N samples.

In a block coding process, the coding frame can advantageously be divided into several adjacent LPC sub-frames or blocks. The analysis length N then exceeds the length of each block in order to allow taking into account a certain number of past and, where appropriate, future samples, by means and at the cost of appropriate coding delays.

The analysis is called LPC <i>"before"</i> when the LPC analysis process is carried out on the block of the current frame of the speech signal to be coded, the coding at the level of the coder intervening "in real time", that is to say during the block of the current frame to the only processing delay introduced by the calculation of the filter coefficients. This analysis involves the transmission of the calculated values ​​of the coefficients of the filters to the decoder.

LPC analysis <i>"back"</i> implementation in the LD-CELP coder at 16 kb / s is the subject of standard ITU-T G728. This analysis technique consists in performing the LPC analysis, not on the block or the block of the current frame of the speech signal to be coded, but on the synthesis signal. It will then be understood that this LPC analysis is in fact carried out on the synthesis signal of the block preceding the current block, since this signal is available simultaneously at the level of the coder and the decoder. This simultaneous operation with the coder and the decoder thus makes it possible to avoid the transmission from the coder to the decoder of the value, obtained at the coder, of the coefficients of the LPC filter. For this reason, the LPC analysis "<i>back</i>"makes it possible to free up the transmission rate, the rate thus freed up being able to be used for example in order to enrich the excitation dictionaries in the case of CELP coding. LPC analysis"<i>back</i>"also authorizes an increase in the analysis order, the number of coefficients of the LPC filter being able to reach 50 in the case of an LD-CELP coder against 10 coefficients for most coders implementing an LPC analysis"<i>before</i>".

Thus, proper functioning of the LPC analysis "<i>back</i>"requires the following conditions:<ul id="ul0001" list-style="dash" compact="compact"><li>good quality of the synthesis signal, very close to the speech signal to be coded, which implies a sufficiently high coding rate, greater than 13 kb / s taking into account the current quality of CELP coders;</li><li>frame and block of reduced length due to the delay of a block between the analyzed signal and the signal to be coded. The frame and block length must therefore be small compared to the average stationarity time of the speech signal to be coded;</li><li>fidelity of the transmission and respect for the integrity of the data transmitted between coder and decoder, by the introduction of few transmission errors. As soon as the synthesis signals differ significantly from the speech signal to be coded, the coder and decoder no longer calculate the same filter and significant divergences may arise, in the absence of any chance of returning to a sensitive identity of the filters calculated at encoder or decoder.</li></ul>

Due to the respective advantages and disadvantages of the aforementioned types of LPC analysis <i>"back"</i> and <i>"before",</i> a technique of selectively combining LPC analysis <i>"back"</i> and <i>"before"</i> was proposed in the article entitled: <i>"Dual Rate Low Delay CELP Coding (8 kbits</i>/<i>s</i>/<i>16 kbits</i>/<i>s) using</i> a mixed <i>Backward</i>/<i>Forward Adaptive LPC Prediction "</i> published by S.PROUST, C.LAMBLIN and D.MASSALOUX, Proc.IEEE Workshop Speech Cod. Telecomm., Sept. 1995, pp 37-38. The conditions previously mentioned, relating to the proper functioning of the LPC analysis "<i>back</i>", reveal that this only type of analysis has obvious limits when operating at transmission rates significantly lower than 16 kb / s. In addition to the reduction in quality of the synthesis signal, which degrades the performance of the LPC filter, it is most often necessary, in order to reduce the transmission rate, to operate on a longer LPC frame length, of the order of 10 to 30 ms. It can then be seen that under these conditions, the degradation occurs before all during frequency spectrum transitions and more generally in areas with little stationary, while for signals which are generally very stationary such as those relating to music, LPC analysis <i>"back"</i> retains a very significant advantage over LPC analysis <i>"before.".</i>

The purpose of combining the two types of LPC analysis mentioned above is to overcome these drawbacks by taking advantage of the advantages inherent in each of them:<ul id="ul0002" list-style="dash" compact="compact"><li>LPC analysis "<i>before</i>"for coding transitions and non-stationary areas;</li><li>LPC analysis <i>"back"</i>, of higher order, for coding stationary areas.</li></ul>

In addition, the introduction of LPC frames encoded by LPC analysis "<i>before</i>"among LPC frames encoded by LPC analysis <i>"back"</i> allows the coder and the decoder to converge again towards the same synthesis signal in the event of transmission error and therefore offers a robustness to these errors which is far superior to coding by LPC analysis <i>"back"</i> pure.

Overall, the mixed LPC analysis "<i>before behind"</i> mentioned above consists in carrying out two LPC analyzes, an LPC analysis <i>"before"</i> on the speech or audio signal to be coded and an LPC analysis <i>"back"</i> on the synthesis signal. Two filters are calculated for each LPC block, these filters being designated by LPC filter "<i>before</i>"and LPC filter"<i>back</i>"respectively. A procedure for choosing the filter applied for the LPC block considered as a function of the stationarity of the signal is then implemented. This procedure uses two distinct criteria:<ul id="ul0003" list-style="dash" compact="compact"><li>a first criterion based on the filter prediction gains:</li></ul><ul id="ul0004" list-style="dash" compact="compact"><li>a second criterion based on a distance parameter between LPC filters "<i>before</i>"calculated successively. For each of these two criteria, fixed threshold values ​​are established.</li></ul>

First criterion:

Choosing the LPC filter <i>"back"</i> is retained if the difference between the prediction gain of the LPC filters <i>"back"</i> and <i>"before"</i> is greater than a first threshold value.

Second criterion:

For a current analysis in LPC analysis mode <i>"back",</i> prohibition of switching LPC analysis mode <i>"back"</i> in LPC analysis mode <i>"before"</i> if the distance calculated on the parameter vectors representing two LPC filters "<i>before</i>"consecutive is less than a second threshold value, a too small distance characterizing a substantially stationary zone for which it is advisable to avoid any change in LPC analysis mode. The distance calculated is a Euclidean distance between the spectral lines of the signal speech or audio frequency to be coded.

For a more detailed description of the above-mentioned mixed LPC analysis mode, one can usefully refer to the article published by S.PROUST, C.LAMBLIN and D.MASSALOUX previously cited.

Extensive investigations carried out on the procedure for the above-mentioned mixed analysis have made it possible to highlight the following significant drawbacks:<ul id="ul0005" list-style="dash" compact="compact"><li>for some signals, the values ​​of the prediction gains of the LPC filters <i>"before"</i> and <i>"back"</i> can oscillate on either side of the first threshold value. This phenomenon causes LPC filter changes<i>"back" -</i> LPC <i>"before",</i> or vice versa, abrupt and frequent. The filtering discontinuities then introduced constitute a source of significant degradation of the synthesis signal and are not, most of the time, linked to real spectral modifications of the speech or audiofrequency signal to be coded;</li><li>the optimal value of the first threshold, which should be fixed, varies very greatly as a function of the stationarity of the signal to be coded, all the more so when the coding rate is low. For a coding delay corresponding to an LPC frame of 10 to 30 ms, or when the transmission rate decreases, there is a very clear divergence in the coding mode of music and speech signals. For music signals, LPC analysis<i>"back"</i> is used almost permanently while for speech signals, LPC analysis "<i>before</i>"is used mostly.</li></ul> Whereas in the case of music signals, the stationarity being very high, the analysis in LPC mode <i>"back"</i> is retained even for a long LPC frame length, in the case of speech signals, on the contrary, the very stationary zones are of very limited duration and the transition to LPC analysis mode <i>"back"</i> consequently in short, which causes unwanted filter transitions which degrade the quality of the coding. The encoder is then no longer able to correct the phenomena caused by the discontinuity introduced by the tilting of the filters;<ul id="ul0006" list-style="dash" compact="compact"><li>the LPC filter which gives the best subjective quality and therefore best models the spectrum of the signal to be coded is not always the one with the best prediction gain. Certain switches from one LPC analysis mode to another, linked to an instant decision, are therefore unnecessary.</li></ul>

The object of the present invention is to remedy the aforementioned drawbacks by implementing a method and a device for coding an audio digital signal by LPC analysis "<i>before</i>"and <i>"back"</i> specific.

Another object of the present invention is also the implementation of a process of dynamic adaptation of the choice function between the LPC analysis. <i>"before"</i> and LPC analysis <i>"back"</i> depending on the degree of stationarity of the signal to be coded.

Another object of the present invention is also the implementation of a process of dynamic adaptation of the aforementioned choice function on the basis of a discrimination between strongly stationary signals, such as music or background noise, and other signals. , such as speech, to enable the most appropriate coding processing by LPC analysis <i>"back"</i> and <i>"before"</i> respectively.

Another object of the present invention is also, the choice of the aforementioned most appropriate coding having been carried out, for a signal to be coded of a given type or characteristics, to avoid any untimely switchover in the non-LPC analysis mode. retained, and thus avoid the appearance of LPC filter transitions "<i>before</i>" - "<i>back</i>"or vice versa likely to degrade the quality of the reproduced synthesis signal.

Another object of the present invention is finally the implementation of a process of dynamic adaptation of the aforementioned choice function for which the change in LPC analysis mode corresponds faithfully to a change in stationarity of the signal to be coded and therefore risks being much less linked to a simple occasional crossing of the first and second threshold values.

The method and the device for coding a digital audio signal, objects of the present invention, implement a double analysis on criteria of choice of LPC analysis. <i>"before"</i> and <i>"back"</i> respectively to generate a transmitted coded signal consisting of LPC filtering parameters accompanied by analysis decision information and a coding residue signal, not transmitted. The digital audio signal is subdivided into frames, a succession of blocks of a determined number of samples and the coding of this digital audio signal is carried out on this signal using LPC filtering "<i>before</i>"for non-stationary areas and on a synthesis signal respectively, this synthesis signal being obtained from the residual coding signal, from an LPC filtering <i>"back"</i> for stationary areas. They are remarkable in that they consist of and allow, respectively:<ul id="ul0007" list-style="dash" compact="compact"><li>determining the degree of stationarity of the digital audio frequency signal according to a stationarity parameter, the value of which lies between a maximum stationarity value and a minimum stationarity value;</li><li>establishing an analysis choice value, by applying an adaptive decision function to the stationarity parameter;</li><li>apply the analysis choice value to the LPC filtering to carry out the coding of the digital audio signal by "forward" LPC filtering for the non-stationary zones on the digital audio signal and by LPC filtering <i>"back"</i> for stationary zones on the synthesis signal.</li></ul> This operating mode makes it possible to favor maintaining in one of the LPC filtering modes <i>"before"</i> and <i>"back"</i> respectively, in connection with the degree of stationarity of the digital audio signal and to limit the number of switches from one to the other of the filtering modes and vice versa.

The method and the device, objects of the present invention, find application not only in the field of mobile telephony but also in the industry of creation and reproduction of phonograms, in satellite transmission and in high quality telephony for video or audio conference, multimedia applications.

They will be better understood on reading the description and on observing the drawings below, in which:<ul id="ul0008" list-style="dash" compact="compact"><li>FIG. 1 represents, in the form of a general flowchart, an illustrative diagram of the steps allowing the implementation of the coding method, object of the present invention;</li><li>FIG. 2a represents a general flow diagram of the steps for calculating the stationarity parameter for each current LPC block;</li><li>FIG. 2b represents a particular advantageous embodiment of the essential steps of the calculation of the stationarity parameter according to FIG. 2a;</li><li>FIG. 2c represents a detail of embodiment of FIG. 2b, more particularly a detail of the process of refining the value of the intermediate stationarity parameter for obtaining the stationarity parameter;</li><li>FIGS. 2d and 2e represent a first, respectively a second nonlimiting example of implementation of a refining function making it possible to calculate a refining value of the intermediate stationarity parameter as a function of the relative values ​​of the filtering gain LPC <i>"before"</i> and <i>"back"</i> ;</li><li>FIG. 2f represents, by way of illustrative example, a flow diagram of the steps allowing the implementation of the decision function and of the value of the LPC analysis choice "<i>before</i>" or <i>"back"</i> ;</li><li>FIG. 3 represents, in the form of functional blocks, the general diagram of an encoder making it possible to carry out the coding of an audio frequency signal in accordance with the object of the present invention;</li><li>FIG. 4 represents, in the form of functional blocks, the general diagram of a decoder making it possible to carry out the decoding of an audio-coded signal signal thanks to the implementation of an encoder as represented in FIG. 3.</li></ul>

A more detailed description of the coding process for a digital audio signal by double analysis, using the LPC analysis selection criterion "<i>before</i>" respectively <i>"back"</i> in a transmitted coded signal, object of the present invention, will now be given in connection with FIG. 1.

In general, it is indicated that the transmitted coded signal, noted s_c_not(t), partly consists of LPC filtering parameters accompanied by LPC analysis decision information. In addition, a coding residue signal res_not(t) not transmitted is available by the implementation of the coding method.

The digital audio signal is subdivided into LPC frames, succession of LPC blocks, each block, for the convenience of the description, being denoted B_not and provided with a determined number N of samples.

In accordance with one aspect of the coding method which is the subject of the present invention, it consists in carrying out the aforementioned coding on the digital audio frequency signal as defined above from LPC filtering "<i>before</i>"for non-stationary areas, respectively on a synthesis signal obtained from the residual coding signal from an LPC filtering <i>"back"</i> for stationary areas.

According to a particularly remarkable aspect of the process which is the subject of the present invention, it consists, in order to establish the LPC filtering selection criterion <i>"before"</i> or <i>"back",</i> on each current block of the succession of current blocks constituting a current frame, as shown in FIG. 1, each current block, denoted B_not, being available in a starting step 10, to be determined in a step 11 the degree of stationarity of the digital audio signal according to a stationarity parameter, denoted STAT (n). This stationarity parameter has a numerical value between a maximum stationarity value, denoted STAT_M, and a minimum stationarity value, denoted STAT_m.

By convention and without in any way detracting from the degree of generality of the coding method which is the subject of the present invention, it is indicated that the stationarity parameter has the maximum value STAT_M for a very strongly stationary signal, while this stationarity parameter has the minimum value STAT_m for a very strongly non-stationary signal.

Following the aforementioned step 11, the coding method which is the subject of the present invention consists in establishing, in a step 12, from the stationarity parameter STAT (n), an analysis choice value LPC, this choice value corresponding analysis of course, or the choice of LPC analysis <i>"before"</i>, or on the contrary to the choice of LPC analysis <i>"back".</i> The value of choice of analysis is noted d_not(n) and is obtained from a specific decision function, denoted D_not.

The aforementioned step 12 is then followed by a test step 13 allowing the application of the analysis choice value d_not(n), symbolized by C, with LPC filtering for coding the digital audio signal by LPC filtering <i>"before"</i> for non-stationary zones on the digital audio signal, respectively by LPC filtering <i>"back"</i> for stationary zones on the synthesis signal.

The implementation of the decision function D_not and the above-mentioned analysis choice values ​​d_not(n), in accordance with a particularly advantageous aspect of the coding method which is the subject of the present invention, makes it possible to favor the maintenance in one of the LPC filtering modes <i>"before"</i> respectively <i>"back",</i> in conjunction with the degree of stationarity of the audio signal, and to limit the number of switches from one to the other of the filtering modes, and vice versa.

In general, it is indicated that the decision function implemented in step 12, this decision function being denoted D_not, is an adaptive function updated for each current block B_not, from the stationarity parameter.

Updating the adaptive function makes it possible to favor maintaining in one of the LPC filtering modes <i>"before",</i> respectively <i>"back"</i>, depending on the degree of stationarity of the digital audio signal and thus limit the number of switches from one to the other of the filtering modes, and vice versa.

More specifically, we indicate that the choice of analysis value d_not(n) established from the decision function D_not above corresponds to an LPC filtering mode priority value "<i>before</i>" or <i>"back"</i> as well as another priority value representing in fact a value of no priority to return to the LPC filtering mode <i>"back"</i> or <i>"before".</i>

By LPC filtering mode priority value, we indicate that the analysis choice value d_not(n) can for example correspond to a logical value, the true value of this logical value, value 1 for example, corresponding to a choice of LPC filtering <i>"back"</i> whereas the value complemented by this true value, the value zero, corresponds to a choice of LPC filtering <i>"before".</i> It is thus understood that the test function in step 13 can be summed up as a test value on the logical value of the above-mentioned analysis choice value to ensure in LPC filtering step 14 <i>"back"</i> for stationary areas of the signal to be coded or LPC filtering <i>"before"</i> in step 15 for the non-stationary zones, the aforementioned steps 14 and 15 then being followed by steps 14a and 15a on returning to the next block marked B_n-1, for n = n + 1.

Although the value of choice of analysis d_not(n) is represented by a logical value, it is understood that this logical value can be associated with a priority and probability value of filtering mode established by the decision function D_not specifically. We understand in particular that this probability value can correspond, for each current block B_not, to the true logic value for a range of probability values ​​between zero and 1 for LPC filtering <i>"back"</i> whereas the complemented logic value, logic value zero for example, may correspond to the complement of the above-mentioned range of probability values ​​between zero and 1 of the first above-mentioned range. This probability is linked to the number of successive filtering decisions in the same filtering mode.

The operating mode of the decision function D_not actually allowing to associate with the logical variable d_not(n) the filtering mode priority value, is adaptive over time, for each current block B_not.

In general, we indicate that the adaptation of the decision function D_not aims to gradually favor the LPC filtering mode <i>"back"</i> or on the contrary the LPC filtering mode <i>"before"</i> which works best, taking into account the overall stationarity of the signal to be coded, in order to avoid as much as possible any unnecessary switching from one of the filtering modes to the other.

More specifically, we indicate that:<ul id="ul0009" list-style="dash" compact="compact"><li>the more stationary the signal to be coded, the more the decision function D_not favors LPC analysis "<i>back</i>"by limiting as much as possible the switch to LPC analysis mode"<i>before</i>",</li><li>on the contrary, the less stationary the signal to be coded, the more the decision function D_not favors LPC analysis "<i>before</i>"by limiting as much as possible any switch to LPC analysis mode <i>"back".</i></li></ul>

A more detailed description of the implementation of a specific decision function making it possible to carry out the adaptation of this decision function, as a function of the value of the stationarity parameter STAT (n), will be given later in the description.

A preferred method of calculating the stationarity parameter STAT (n) relative to each current LPC block B_not will now be given and described in connection with FIG. 2a.

According to the above-mentioned figure, it is indicated that step 11 consisting in determining the degree of stationarity of each current block B_not of the digital audio signal consists, starting from an arbitrary starting value of the stationarity parameter, as represented in step 110 of FIG. 2a, this arbitrary value being denoted STAT (O), to be calculated in a step 111 for this current block B_not an intermediate stationarity parameter value, denoted STAT * (n), as a function of a determined number of successive analysis choice values, these LPC analysis choice values, denoted d_n-1(n-1), ..., to d_np(np), being obtained for different successive blocks prior to the current block B_not the succession of LPC blocks, and the value of the stationarity parameter of the block preceding the current block, this stationarity value being denoted STAT (n-1). In step 111 represented in FIG. 2a, it is indicated that the function of the determined number of previous analysis choice values ​​is given in relation to these previous values, denoted by_n-1(n-1) to d_np(np). With regard to the arbitrary starting value of the stationary parameter STAT (0), it is indicated that this can, by way of nonlimiting example, be taken equal to the average value between the maximum value and the minimum value of the stationarity parameter previously mentioned in the description, STAT_M and STAT_m.

The aforementioned step 111 is then followed by a step 112, which consists in refining the value of the intermediate stationarity parameter as a function of the value of the prediction gains of the filters or LPC analysis mode. <i>"before"</i> and <i>"back"</i> of the frame preceding the current frame. In step 112 of FIG. 2a, it is indicated that the aforementioned function is denoted g (STAT * (n), Gpf, Gpb) where Gpf denotes the prediction gain of the LPC filter "<i>before</i>"and Gpb denotes the LPC filter's prediction gain <i>"back"</i> for the frame preceding the current frame. At step 112, that is to say at the end of the step consisting in refining the value of the intermediate stationarity parameter, to the value of stationary parameter STAT (n) of the current LPC block B_not the value is assigned, relation (3):<maths id="math0003" num=""><math display="block"><mrow><mtext>STAT (n) = g (STAT * (n), Gpf, Gpb)</mtext></mrow></math><img file="EP0906613B1_D0003.tif" /></maths> corresponding to the refined value of the intermediate stationarity parameter.

A more detailed description of the calculation step 111 of the intermediate stationarity parameter STAT * (n) and of the step 112 consisting in refining this parameter value will now be given in connection with FIG. 2b.

In accordance with the above-mentioned figure, step 111 consists, from an initialization step 1110 in which the value of the stationarity parameter STAT (n-1) and the value of analysis choice d_n-1(n-1) relating to the LPC B block_n-1 prior to current block B_not is available, to perform, in a step 1111, a step consisting in discriminating the LPC analysis mode "<i>before</i>"or LPC <i>"back"</i> from block B_n-1 preceding the current block B_not. This discrimination step 1111 can, as shown in FIG. 2b, consist of a test step on the value of choice of analysis d_n-1(n-1) with respect to the symbolic value <i>"fwd"</i> or to the logic value zero corresponding to the value complemented by the true logic value.

On negative response to the above test 1111, that is to say for any block B_n-1 preceding the current block LPC B_not analyzed in LPC analysis mode <i>"back",</i> the step of calculating the intermediate stationarity parameter value consists, in a step 1113, of determining the number of anterior frames analyzed consecutively in LPC analysis mode "<i>back</i>", number noted N_BWD, then, in a step 1114, to compare on criteria of comparison of superiority the number of frames prior to a first arbitrary value, noted Na, representative of a number of successive frames analyzed in LPC mode <i>"back".</i>

On a positive response to the comparison of superiority of test 1114, the calculation step then consists in assigning, in a step 1114b, to the value of intermediate stationarity parameter STAT * (n), the value of the stationarity parameter of the preceding block the current block, STAT (n-1), increased by a determined value as a function of the first arbitrary value representative of a number of successive frames analyzed, that is to say in fact of the number of anterior frames N_BWD analyzed consecutively in LPC analysis mode "<i>back</i>". In step 1114b, the value determined as a function of the first arbitrary value is noted f_at(N_BWD). During the aforementioned step, it is understood that the value of the intermediate stationarity parameter STAT * (n) for the current LPC block B_not is thus increased compared to the corresponding value of the same stationarity parameter for the previous block B_n-1.

On a negative response to the comparison of superiority to the comparison test 1114, the value of the intermediate stationarity parameter STAT * (n) is assigned, in a step 1114a, the value of the stationarity parameter STAT (n-1) of the preceding block the current block B_not.

On the contrary, for any previous block B_n-1 analyzed in LPC analysis mode <i>"before",</i> that is to say on a positive response to the test 1111, the step of calculating the intermediate stationarity parameter 111 consists, as shown in FIG. 2b, of determining in a step 1112, on the test criterion, the occurrence d a transition from LPC analysis mode <i>"back"</i> in LPC analysis mode "<i>before</i>"between the block before the block before the current block B_n-1, of rank n-2, i.e. the existence of an LPC analysis choice value d_n-2(n-2) = symbolic value "bwd", ie logical value zero as mentioned above. The positive response to test 1112 indicates the existence of such a transition from the analysis mode<i>"back"</i> for the LPC B block_n-2 preceding the block preceding the current block B_n-1, whereas a negative response to the aforementioned test 1112 indicates the absence of such a transition.

On a positive response to the aforementioned occurrence test 1112, the calculation step 111 then consists in comparing, on the basis of an inferiority comparison criterion, the number of above-mentioned anterior frames N_BWD with a second arbitrary value N_b representative of a number of successive frames analyzed in LPC mode <i>"back"</i> preceding block B_n-1 preceding the current block.

On a positive response to the comparison carried out in test 1118, this test is followed by a step 1118a consisting in assigning to the intermediate stationarity parameter value STAT * (n) the value of the stationarity parameter of the block preceding the current block, STAT (n-1) reduced by a determined value, function of the second arbitrary value N_b, this determined value being noted f_b(N_BWD). It is thus understood that during the allocation step 1118a, the value of the intermediate stationarity parameter is thus reduced accordingly.

On the contrary, on a negative response to the inferiority comparison carried out in test 1118, step 111 then consists in assigning, in a step 1118b, to the value of the intermediate stationarity parameter STAT * (n) the value of stationarity parameter of the block preceding the current block, that is STAT (n-1).

In FIG. 2b, it will be noted that the allocation steps 1118a and 1118b are then followed by a step of resetting to zero the number of successive blocks processed in LPC analysis mode. <i>"back",</i> this zeroing step carrying the reference 1118c and making it possible to update the whole process of calculating the value of the intermediate stationarity parameter.

On a negative response to the 1112 comparison test, no LPC analysis transition "<i>before</i>"not having appeared, the value of the intermediate stationary parameter STAT * (n) is assigned the value of the stationary parameter STAT (n-1) of the preceding block B_n-1 in one step 1119.

At the end of step 111, there is the value of the intermediate stationarity parameter STAT * (n) for the current block B_not.

As regards step 112 consisting in refining the value of the aforementioned intermediate stationarity parameter, it is indicated, with reference to FIG. 2b, that this can advantageously consist, in a step 1120, in discriminating the prediction gains from the LPC filtering "<i>back</i>"and LPC filtering"<i>before</i>t ", these gain values ​​being denoted Gpb and Gpf respectively. It is understood that the aforementioned discrimination step simply consists in memorizing and reading the gain values ​​calculated for the LPC filtering <i>"before"</i> respectively <i>"back"</i> cited above. In addition to the aforementioned gain values, step 1120 can consist in calculating the relative value of the prediction gains, denoted DGfb, such as the difference or the ratio between the prediction gains.<i>"before"</i> and <i>"back"</i> mentioned above.

As shown in addition in FIG. 2b, step 112 of FIG. 2a comprises, after the above-mentioned step 1120, a step 1121 consisting in modifying the value of the intermediate stationarity parameter STAT * (n) a refinement value ΔS, this refinement value in accordance with a particularly remarkable characteristic of the method which is the subject of the present invention being a function of the relative value of the LPC filtering prediction gains <i>"before"</i> and <i>"back".</i>

In general, it is indicated that the representative function of the refinement value ΔS is noted:<ul id="ul0010" list-style="dash" compact="compact"><li>ΔS = f_r(Gpf, Gpb) where Gpf and Gpb as previously designate the LPC filtering prediction gains <i>"before"</i> respectively <i>"back".</i></li></ul>

In general, we indicate that the function f_r(GPf, Gpb) making it possible to establish the refining value ΔS is an increasing respectively decreasing function of this relative value, according to the direction in which this relative value is considered. When the relative value designates the value of the LPC filtering gain<i>"back"</i> versus LPC filter gain <i>"before",</i> this choice can be arbitrarily retained without in any way detracting from the generality of the process, object of the invention, abovementioned relative value DGfb, the function f_r is then increasing. Otherwise, it decreases.

In other words, the modification, by increase or by decrease, of the value of the intermediate stationarity parameter of the ripening value ΔS is proportional to this relative value of the gains. Generally, this modification is written STAT (n) = STAT * (n) + kΔS. In practice we will take k = 1. More specifically, we indicate that the refinement value ΔS increases in algebraic value when the difference between the LPC filtering prediction gains<i>"before"</i> and <i>"back"</i> increases, the function f_r(GPf, Gpb) then being an increasing function, while this refinement value ΔS decreases in algebraic value when this same abovementioned difference decreases, the abovementioned difference being defined between the LPC filtering prediction gain <i>"back"</i> and the LPC filtering prediction gain "<i>before</i>". In fact, this function is increasing or decreasing depending on the definition of this difference.

Consequently, at the end of step 1121 as shown in FIG. 2b, the value of the intermediate stationarity parameter STAT * (n) can then, for k = 1, be corrected by the algebraic value of the refinement value ΔS above to calculate the value of the stationary parameter STAT (n).

Following step 1121, the value of the stationarity parameter STAT (n) is thus available in step 1122.

A more detailed description of step 1121 of FIG. 2b will now be given in conjunction with FIG. 2c in a preferred embodiment in which a plurality of test criteria are applied both to the refining value and to the values ​​of LPC prediction gain <i>"before"</i> and <i>"back"</i> with a view to optimizing the stationarity parameter calculation process.

As shown in FIG. 2c above, step 1121 can consist of a first step 1121a making it possible to calculate the refinement value ΔS from the function f_r(Gpf, Gpb) previously cited. Different examples of usable functions will be given later in the description.

Firstly, the refinement value ΔS is subjected to a comparison test of superiority to the value 0, in a step 1121b, this comparison test in fact making it possible to determine the increase in this refinement value ΔS.

On a positive response to the aforementioned test 1121b, the refinement value ΔS being positive and corresponding to an increase in the relative value of the LPC filtering prediction gains "<i>before</i>"and"<i>back</i>", the step of increasing the value of the intermediate stationarity parameter by the refinement value ΔS is also subject to a condition of superiority of the LPC filtering gain value"<i>back</i>", compared to a first positive value determined, in a step of comparing the superiority of the value of the LPC filtering gain <i>"back"</i> Gpb compared to this first determined positive value, noted S_i.

On a negative response to the aforementioned test 1121c, the value of the stationary parameter STAT (n) is assigned the value of the intermediate stationary parameter STAT * (n) in a step 1121g.

On a positive response to the aforementioned test 1121c, the increase in the value of the intermediate stationarity parameter of the ripening value ΔS is furthermore subject to a condition of inferiority of the value of the intermediate stationarity parameter STAT * (n) by relative to a second determined positive value STAT_i of course representing a stationarity value. This inferiority condition test is carried out in step 1121e.

On a negative response to the above test 1121e, the value of the intermediate stationarity parameter STAT (n) is assigned the value of the intermediate stationarity parameter STAT * (n) in the aforementioned step 1121g.

On a positive response to the inferiority condition test 1121e, the value of the intermediate stationarity parameter STAT (n) is assigned the value of the intermediate stationarity parameter STAT * (n) increased by the positive value ΔS by the ripening value at step 1121i.

On the contrary, on a negative response to the aforementioned test 1121b, the refinement value ΔS being negative, the step of decreasing the intermediate stationarity parameter by the refinement value ΔS, this value being negative, is also subjected to a test. LPC filter gain gain inferiority condition <i>"back"</i> Gpb with respect to a third determined positive value denoted S_d in a comparison step 1121d. This third determined positive value is of course representative of an LPC filtering gain value.

On a negative response to the aforementioned test 1121d to the stationarity parameter value STAT (n), the value of the intermediate stationarity parameter STAT * (n) is assigned in step 1121g.

On the contrary, on a positive response to the aforementioned test 1121d, the step of decreasing the value of the intermediate stationarity parameter of the ripening value ΔS is furthermore subject to a condition of superiority of the value of the intermediate stationarity parameter STAT * (n) with respect to a fourth determined positive value, denoted STATd in a comparison test denoted 1121f. Of course, the fourth positive value determined is representative of a chosen stationarity parameter value.

On a negative response to the aforementioned test 1121f, the value of the stationary parameter STAT (n) is assigned the value of the intermediate stationary parameter STAT * (n) in step 1121g.

On positive response to the aforementioned test 1121f, the value of the stationary parameter STAT (n) is assigned the value of the intermediate stationary parameter STAT * increased by the algebraic value of the refinement value ΔS, negative, the value of the parameter of intermediate stationarity thus being reduced to establish the stationarity parameter value STAT (n) in step 1121h.

At the end of steps 1121g, 1121h and 1121i, there is thus at step 1122 of FIG. 2b the stationarity parameter STAT (n).

Regarding the function f_r(Gpf, Gpb), we indicate that it can consist of. a nonlinear function of the relative value of the LPC filtering gains<i>"before"</i> and <i>"back"</i> where the relative value of the LPC filtering prediction gains <i>"before"</i> and <i>"back"</i> may itself consist of either the ratio or the difference in LPC filtering prediction gains "<i>before</i>"and"<i>back</i>". Other types of functions, such as linear functions, can be used.

A first example of a nonlinear function f_r(Gpf, Gpb) is shown in Figure 2d.

In the exemplary embodiment of FIG. 2d, value pairs of the LPC filtering prediction gain <i>"back"</i> Gpb plotted and LPC filtering gain "<i>before</i>"Gpf are used to assign refinement values ​​ΔS positive, ΔS&gt; 0 or negative ΔS &lt;0 for a value of the ratio ρ = Gpb / Gpf corresponding to a higher slope respectively lower than that of the line ΔS = 0.

In FIG. 2e, the case is shown where the relative value of the LPC filtering prediction gains <i>"before"</i> and <i>"back"</i> no longer corresponds to the ratio of gains ρ but to the difference of the aforementioned gains.

In this case, the function of the relative value of the LPC filtering prediction gains "<i>before</i>"and <i>"back"</i> f_r(Gpf, Gpb) can also be a nonlinear function allowing to attribute to the refinement value ΔS for values ​​of this difference corresponding to pairs of value Gpb, Gpf corresponding to straight lines whose abscissa at the origin is lower respectively higher, in algebraic value, to the abscissa at the origin of the line ΔS = 0.

In the case of FIG. 2e, the straight lines delimiting the zones as a function of the sign of the refinement value ΔS are parallel to each other.

In accordance with another particular aspect of the method which is the subject of the invention, it is further indicated that the stationarity index of the current block B should not be adapted_not during silence frames, when for example the audio frequency signal is constituted by a speech signal comprising silences. In such a case, step 1111 of step 111 shown in FIG. 2b can be preceded by a step 1111a consisting, for each successive current block, in determining the average energy of the digital audio frequency signal and in comparing in this same step ,. on an inferiority comparison criterion, this average energy at a determined threshold value representative of a frame of silence. On figure 2b, this threshold value is noted ENER_SIL. On positive response to the aforementioned test, the value of the stationarity parameter of the current block STAT (n) is assigned the value of the stationarity parameter of the previous block STAT (n-1) in the allocation step 1111b represented in the figure. 2b. The steps 1111a and 1111b are, in the above-mentioned figure, shown in dotted lines, since they are reserved for example for coding a speech signal.

A more detailed description of the implementation of the decision function D_not allowing obtaining the decision values ​​d_not(n) will now be given in connection with FIG. 2f. This description is given in a preferred embodiment in which this decision function, which can be compared to that described in the article previously mentioned by the description, published by S.PROUST, C.LAMBLIN and D.MASSALOUX, is however adapted temporally, in accordance with the object of the present invention in order to obtain the values ​​of choice of analysis d_not(n) successive.

From a step 120, for the current block B_not, we first calculate a distance, denoted by d_LPC, between the LPC filter of the current block and that of the previous block B_n-1. This distance calculation is carried out for example using the LSP frequency parameters as mentioned previously in the description relating to the method described in the aforementioned article.

We notice :<ul id="ul0011" list-style="dash" compact="compact"><li>S_PRED (n) and S_TRANS, S_STAT and G₁ the threshold values ​​involved in the criterion based on the LPC filters' prediction gains <i>"back"</i> and <i>"before"</i> ;</li><li>S_LSP_L and S_LSP_H the values ​​of the thresholds involved in the criterion based on the distances between LSP frequency vectors representing two filters. LPC "<i>before</i>"relating to two consecutive blocks B_n-1 and B_not ;</li><li>Gpf LPC filter prediction gain <i>"before";</i></li><li>Gpb the prediction gain of the filter <i>"back" ;</i> and</li><li>Gpi the filter prediction gain "<i>before</i>"interpolated according to the method exposed in the published article, mentioned previously in the description.</li></ul>

The criterion for establishing the decision function, in relation to FIG. 2f, is established as follows:<ul id="ul0012" list-style="dash" compact="compact"><li>if the consecutive LPC filters are very stationary, i.e. for d_LPC &lt;S_LSP_L, then, no LPC filtering switching "<i>back</i>"LPC filtering"<i>before</i>"is only performed if you are in LPC filtering mode"<i>back</i>", provided that the LPC filter's prediction gain <i>"back"</i> is greater than the prediction gain of the LPC filter <i>"before"</i> decreased by an S_STAT value. It is indicated that the value S_STAT is chosen so as to favor the choice of an LPC filter<i>"back"</i> in the presence of a large stationarity of the spectrum measured using the distance d_LPC ;</li><li>if the consecutive LPC filters have a significant transition, i.e. for d_LPC &gt; S_LSP_H and if Gpf&gt; Gpb-S_TRANS, then, the filtering mode chosen is the "before" LPC filtering, ie d_not(n) = 0, symbolic value <i>"fwd"</i>, otherwise, d_not(n) is taken equal to 1, symbolic value "<i>bwd</i>". We indicate that the value of S_TRANS is chosen so as to strongly favor the choice of the LPC filter"<i>before</i>"in the presence of a spectrum transition measured using the distance d_LPC ;</li><li>otherwise, in all other cases, if Gpb&gt; Gpf-S_PRED and Gpi&gt; Gpf-S_PRED, then the LPC filter used is the LPC filter <i>"back"</i> interpolated, provided that the gain of the latter and that of the LPC filter <i>"back"</i> pure exceed the threshold value G₁ previously mentioned. If the condition on the aforementioned prediction gain values ​​is not fulfilled, then we choose LPC filtering<i>"before".</i></li></ul>

To increase the number of LPC filters "<i>before</i>"transmitted and thereby increase the robustness of the transmission error coding system, the LPC filtering mode"<i>before</i>"can be chosen advantageously as soon as the energy of the signal to be coded E_not, i.e. the energy of block B_not corresponding, becomes lower than the value of the energy of a frame of silence ENER_SIL, this value of energy corresponding to the minimum audible level.

All the conditions allowing the establishment of the decision function D_not and obtaining the analysis choice values ​​d_not(n) corresponding, is illustrated in FIG. 2f with temporal adaptation of the decision function D_not.

The value of the stationary parameter STAT (n) can for example be located on a scale of 0, corresponding to the value STAT_m very little stationary, at 100, corresponding to the STAT value_M very stationary.

Depending on the value of the stationary parameter STAT (n), the decision function D_not is modified by adapting the value of the thresholds.

The more the stationarity of the signal increases, the more the LPC filtering mode <i>"back"</i> is favored: the thresholds S_PRED, S_LSP_L and S_LSP_H are increased.

By way of nonlimiting example, the modification functions are indicated for each current LPC boc B_not of the above threshold values:<ul id="ul0013" list-style="dash" compact="compact"><li>S_PRED (n) = f_{S_PRED}(STAT (n)) with f_{S_PRED} increasing function of the value of STAT (n);</li><li>S_LSP_L (n) = f_{S_LPC_L}(STAT (n)) with f_{S_LPC_L} increasing function;</li><li>S_LSP_L (n) = f_{S_LPC_H}(STAT (n)) with f_{S_LPC_H} increasing function.</li></ul>

In the adaptation of the aforementioned threshold values, it is indicated that the increasing functions mentioned are for example staircase functions as regards the functions f_{S_LPC_L} and F_{S_LPC_H}. The function f_{S_PRED} is an affine function of the stationarity parameter variable, of the form:<maths id="math0004" num=""><math display="block"><mrow><mtext>S_PRED (n) = α.STAT (n) + β</mtext></mrow></math><img file="EP0906613B1_D0004.tif" /></maths> where α and β are two real values ​​between 0 and 1 and where the value of S_PRED (n) is bounded in the interval [S_PRED_m, S_PRED_M], S_PRED_m and S_PRED_M represent two experimentally determined values.

In order to further limit the risk of filter tilting, it is then possible to choose, when the stationarity parameter STAT (n) is less than a threshold value S_FWD given, to impose the LPC filtering mode "<i>before</i>".

On the other hand, the threshold values ​​S_TRANS, S_STAT and G₁ keep a fixed value, these values ​​can for example be equal to -1 dB, 5 dB and 0 dB respectively.

Establishing the decision function D_not and obtaining the analysis choice values ​​d_not(n) are illustrated as follows in FIG. 2f: following the above-mentioned step 120, carrying out a test step 121 relating to the energy of the current LPC block B_not, by a comparison of inferiority to the value of energy of silence ENER_SIL or of the value of the stationarity parameter STAT (n), compared by a comparison of inferiority to the value S_FWD previously cited in the description. On a positive response to the aforementioned test 121, the value of choice of analysis d_not(n) is taken equal to 0, i.e. symbolic value <i>"fwd"</i> in step 122.

On a negative response to the aforementioned test 121, a new test is carried out relative to the value of choice of analysis d_n-1(n-1) at logical value 1, that is to say at symbolic value "<i>bwd</i>".

On positive response to the aforementioned test 123, a new test is carried out on the filtering distance LPC d_LPC mentioned above, in a step 124, with respect to the threshold value S_LSP_H (n) by comparison of superiority to this threshold value.

On a positive response to the above test 124, a new test 126a is carried out, consisting in comparing the prediction gain of the LPC filtering "<i>before</i>", Gpf, to the LPC filtering prediction gain <i>"back",</i> Gpb, reduced by the threshold value S_TRANS.

On a positive response to the above test 126a, at the value of choice of analysis d_not(n), is assigned the logical value 0, symbolic value "<i>fwd</i>", and on negative response to the above test 126a, is assigned to the same value of choice of analysis the logical value 1, symbolic value"<i>bwd</i>". The corresponding steps are noted 128 and 129.

On negative response to the test 124 previously mentioned, a new test 125 is carried out. Test 125 consists in making a comparison of the distance of the LPC filtering, d_LPC, by comparison of inferiority to the threshold value S_LSP_L (n).

On positive response to test 125, a new test 126b is performed by comparison of the superiority of the LPC filtering prediction gain <i>"back"</i> to the LPC filtering prediction gain <i>"before"</i> decreased by the value S_STAT previously mentioned.

On positive response to test 126b, at the value of choice of analysis d_not(n) is assigned in step 129 the logical value 1, that is to say the symbolic value "<i>bwd</i>".

On negative response to test 126b, to the value of choice of analysis d_not(n) is assigned the logical value 0, i.e. the symbolic value <i>"fwd",</i> step 128.

On the contrary, on a negative response to test 125, a new test is carried out, in a step 127, this test consisting in verifying the conditions for comparing the LPC filtering gain <i>"back"</i> Gpb to LPC filtering prediction gain <i>"before"</i> decreased by the threshold value S_PRED (n), by comparing the superiority of the intermediate LPC filtering prediction gain Gpi with the value of LPC filtering prediction gain "<i>before</i>"minus the aforementioned threshold value S_PRED (n) and the comparison of the superiority of the filtering prediction gain <i>"back"</i> Gpb at threshold value G₁, as well as comparing the value of the intermediate filtering prediction gain Gpi with the threshold value G₁.

It is indicated that the negative response to test 123 previously mentioned in the description also leads to the carrying out of the aforementioned test 127.

On a positive response to test 127 previously mentioned, at the value of choice of analysis d_not(n) is assigned the logical value 1, i.e. the symbolic value <i>"bwd"</i> at step 129, while at the negative response to the above test 127, at the analysis choice value d_not(n) is instead assigned the logical value 0, i.e. the symbolic value <i>"fwd"</i> in step 128.

We thus have, thanks to the implementation of the decision function D_not, of the analysis choice value d_not(n) obtained with the aforementioned logical values ​​1 or 0, these logical values ​​however being linked to a priority value or lack of priority to return to the filtering mode <i>"back"</i> or <i>"before"</i> depending on the value of the stationarity parameter.

A more detailed description of a device for coding a digital audio signal by double analysis on the LPC analysis selection criterion "<i>before</i>" respectively <i>"back"</i> in a transmitted coded signal, in accordance with the object of the present invention, will now be given in connection with FIG. 3.

In a practical manner, it is indicated that the digital signal to be coded is subdivided into frames constituted by successive blocks of samples, each block comprising a given number N of samples for example.

In FIG. 3, the mode of constitution of the digital audio frequency signal to be coded in successive blocks of samples B_not has not been shown since this operating mode is perfectly known from the state of the art and can be carried out from a simple buffer memory, for example read periodically at the frame frequency and at the block frequency.

As also shown in FIG. 3 above, the coding device which is the subject of the invention comprises an LPC analysis filter "<i>before</i>", labeled 1A, and an LPC analysis filter <i>"back",</i> bearing the reference 1B, in order to allow the delivery of a transmitted coded signal consisting of LPC filtering parameters accompanied by an analysis decision indication, as well as Pr parameters_not relating to the harmonic analysis and the CELP excitation signal.

In general, it is indicated that the analysis decision indication corresponds to the analysis choice value d_not(n) as mentioned previously in the description. As regards the LPC filtering parameters, it is indicated that these correspond to specific parameters, in accordance with the mode of implementation of the coding method which is the subject of the present invention, as will be described below in the description.

In FIG. 3, there has also been shown, in the coding device according to the invention, the existence of an adaptive filter as a function of the value of the stationarity parameter, this adaptive filter bearing the reference 1E. This adaptive filter 1E naturally receives the original digital signal, noted s_{n (t)}, i.e. the current block B_not. The 1E filter uses the LPC filtering parameters to calculate the residual signal which will then be coded by the 1F module. These LPC parameters, as well as the filter decision indication constitute a part of the coded signal which is transmitted to the decoder.

In addition, as shown in FIG. 3, the coding device which is the subject of the present invention comprises a coding means, bearing the reference 1F, of a residual coding signal not transmitted, the residual coding signal, designated by res_{n (t)} is directly available at the output of the adaptive filter 1E, this signal being thus delivered at the input with the digital audio frequency signal to the coding module of the coding residue signal not transmitted, to generate a synthesis residue signal, res_syn_not(t).

A reverse filtering module, bearing the reference 1G, receives the synthesis residue signal and makes it possible to deliver a synthesis signal referenced s_syn_{n (t)}.

A 1H storage module receives the aforementioned synthesis signal s_syn_{n (t)} to deliver the aforementioned synthesis signal for the block prior to the current block B_not, the synthesis signal thus obtained being designated by s_syn_n-1(t). This synthesis signal is delivered to the LPC analysis filter<i>"back"</i> bearing the reference 1B in FIG. 3 above.

The coding device, object of the present invention, as shown in FIG. 3, makes it possible to carry out coding of the digital audio signal on the aforementioned digital audio signal from the LPC filter "<i>before</i>"for non-stationary areas and on the above synthesis signal s_syn_n-1(t) from the LPC filter <i>"back"</i> 1B for stationary zones, as will be described below.

As will be observed in FIG. 3 above, the device which is the subject of the invention comprises for this purpose, for each current LPC block B_not, a module 1C for calculating the degree of stationarity of the digital audio-frequency signal according to a stationarity parameter whose value is between a maximum stationarity value and a minimum stationarity value. Of course, the stationarity parameter is the STAT (n) parameter previously described in the description in accordance with the coding method which is the subject of the present invention. The maximum and minimum stationarity values ​​are also defined above.

As shown in addition in FIG. 3, the coding device which is the subject of the invention comprises a module, denoted 1D₁, establishing from the above-mentioned stationarity parameter STAT (n) a decision function and an LPC analysis choice value, the decision function being denoted D_not as mentioned previously in the description, and the LPC analysis choice value being of course and corresponding to the LPC analysis choice value denoted d_not(n) previously described in the description. Recall that the value of choice of analysis d_not(n) can take the values ​​0 or 1, logical values, which correspond to the symbolic value of choice of analysis <i>"fwd"</i> and <i>"bwd"</i> for LPC analysis <i>"before"</i> and "rear" respectively.

It is understood in particular that with regard to the means of establishing the decision function D_not, this corresponds to a software embodiment for example, as described above in connection with FIG. 2f. In addition, the coding device according to the invention as shown in FIG. 3 comprises an LPC filtering analysis discrimination module, denoted 1D₂, this module receiving the value of choice of analysis d_not(n) and making it possible to deliver, for the current LPC block B_not the value of the LPC filter parameters <i>"back"</i> respectively <i>"before"</i> as a function of the aforementioned analysis choice value. We understand of course that the LPC filtering analysis parameters "<i>back</i>"as well as the LPC analysis filtering parameters"<i>before</i>"are of course available in digital form at the filters bearing the reference 1B and 1A respectively in FIG. 3. These parameters are designated respectively Af_not(z) for LPC filter analysis parameters <i>"before"</i> regarding the LPC analysis filter "<i>before</i>", bearing the reference 1A, and by Ab_not(z) for LPC analysis parameters <i>"back"</i> regarding the LPC analysis filter <i>"back"</i> bearing the reference 1B. These parameters are delivered to the 1D module₁ and to the 1D module₂ respectively.

Regarding the physical implementation of the 1D discrimination module₂, it is indicated that this can for example, in a nonlimiting embodiment, consist of two distinct memory zones allowing the memorization of the filtering parameters Af_not(z) and Ab_not(z) respectively, the analysis choice value d_not(n) as a function of its current logic value, 0 or 1, allowing the addressing in reading of the values ​​of filtering parameters memorized by the 1D module₂ for example and the transmission of these filtering parameters by the latter.

Finally, as shown in FIG. 3, it is indicated that the coding device in accordance with the object of the present invention, for producing the adaptive filter as a function of the stationarity value carrying the reference 1E, can be produced by a filter element whose transfer function, denoted A (z), is established from the values ​​of filter parameters delivered by the discrimination module 1D₂ previously mentioned.

It is thus understood that the adaptive filtering module 1E can be produced by a filter with adjustable coefficients, to the value of the coefficients of the latter being assigned the values ​​of filtering parameters delivered by the discrimination module 1D₂ previously mentioned. The filtering carried out by the module 1E is thus of the adaptive type as a function of the degree of stationarity of the digital audio frequency signal to be coded. The 1E module thus delivers, from the original digital audio signal s_{n (t)}, the LPC filtering residual signal designated by res_not(t) the 1F residue coding module, which then makes it possible to deliver the LPC synthesis residue signal designated by res_syn_not(t).

Finally, the module 1G is a filtering module whose transfer function is the inverse of the transfer function of the module 1E obtained from the memorized parameters of the latter. It receives the LPC synthesis residue signal res_syn_not(t) delivered by the coding module of the coding residue delivered by the module 1F. It is thus understood that the coding of the digital audio signal s_not(t) is performed at module 1E level thanks to LPC analysis <i>"before",</i> respectively <i>"back"</i> performed by LPC analysis filters <i>"before"</i> 1A and LPC analysis <i>"back"</i> 1B, the coded signal s_c_not(t) consisting in the transmission of the LPC filtering parameters <i>"before"</i> when the analysis choice value d_not(n) has symbolic value <i>"fwd"</i> as well as the indication of the choice of analysis, that is to say of the value of the choice of analysis previously cited. This operating mode makes it possible to carry out the coding of the digital audio-frequency signal and to favor the maintenance in one of the LPC filtering modes.<i>"before",</i> respectively <i>"back",</i> according to the degree of stationarity of the digital signal and to further limit the number of switches from one to the other of the filtering modes considered.

A device for decoding a digital audio signal coded in double analysis on LPC analysis selection criterion <i>"before",</i> respectively <i>"back",</i> in a coded signal transmitted in accordance with the coding method which is the subject of the present invention, and thanks to the implementation of a coding device as shown in FIG. 3 for example, will now be described in conjunction with FIG. 4.

Generally, it is indicated that the coded signal transmitted s_c_not(t) consists for each LPC analysis block of the above-mentioned analysis choice value and, in the case where the analysis choice value corresponds for the LPC analysis block considered to an LPC analysis "<i>before</i>", in the LPC filter settings <i>"before"</i> as well as the LPC filtering residue coding parameters, Pr parameters_not, i.e. res signal_not(t) into a synthesis residue signal res_syn (t) by the residue coding module 1F.

As shown in FIG. 4, it is indicated that the decoding device comprises at least one synthesis module, referenced 2A, of the filtering residue signal receiving the coding parameters of the LPC residue delivered by the module 1F. The module 2A decodes the coding parameters supplied by the module 1F and consequently delivers a synthesis residue signal, which is referenced in FIG. 4 res_syn_not(t).

The decoding device as shown in FIG. 4 also includes a module, bearing the reference 2B, of adaptive reverse filtering as a function of the degree of stationarity, receiving the previously mentioned synthesis residue signal, delivered by the module 2A, and allowing d 'generate a synthesis signal s_syn_not(t) representative of the digital audio frequency signal, this signal in fact constituting the decoded signal.

It will of course be understood that the inverse filtering module 2B implements the filtering parameters received by the decoder due to the transmission, ie the LPC analysis parameters "<i>before</i>"when these are transmitted and the analysis decision corresponds to an LPC analysis"<i>before</i>"or, on the contrary, the filter analysis parameters <i>"back"</i> as will be described below.

To this end, the decoding device which is the subject of the present invention obviously includes an LPC filtering module. <i>"back",</i> carrying the 2D reference, receiving the synthesis signal, that is to say the signal referenced s_syn_not(t) for the LPC block prior to the current LPC block, this synthesis signal thus being referenced s_syn_n-1(t) in FIG. 4. It is understood for this purpose that the synthesis signal relating to the current block B_not and referenced s_syn_not(t) can then be delivered to the LPC filter module <i>"back"</i> 2D by means of a storage module, bearing the reference 2E, which in fact makes it possible, by addressing with suitable reading, to shift the reading of the synthesis signal to that corresponding to the block preceding the current block B_not.

Finally, and to ensure the above-mentioned operating mode, the decoding device which is the subject of the present invention, as shown in FIG. 4, finally comprises a discriminator module bearing the reference 2C, making it possible to discriminate the LPC analysis <i>"before",</i> respectively <i>"back".</i> The module 2C receives, on the one hand, for discrimination control, the value of choice of analysis received, that is to say the value d_not(n), and, on the other hand, the LPC filtering parameters "<i>before</i>", i.e. the parameters Af_not(z) transmitted, as well as the LPC filtering parameters <i>"back"</i> Ab_not(z) obtained by means of the 2D module. The module 2C thus makes it possible to deliver, as a function of the value of choice of analysis, that is to say of the value d_not(n), i.e. the LPC filtering parameters <i>"before"</i> Af_not(z), i.e. the LPC filtering parameters <i>"back"</i> Ab_not(z) to the adaptive reverse filter module 2B according to the degree of stationarity.

As regards the hardware embodiments of the modules 2C and 2B, it is indicated that these can simply consist of modules substantially identical to the modules 1D₂ and 1E or, more particularly, 1G in FIG. 3.

With regard to the actual production of a coding device in accordance with the object of the present invention, allowing the implementation of the method as described previously in the description, two specific embodiments have been produced.

*

CELP coder in telephone band, according to broadband extension of the ITU-T standard at 8 kb / s

:

The actual encoder consisted of a telephone band encoder from 300 to 3400 Hz, at a rate of 12 kb / s of the CELP type. The frames were formed over a duration of 10 ms for an excitation provided by algebraic dictionary according to the so-called ACELP technique previously mentioned in the description.

LPC analysis "<i>before</i>"was a 10-order analysis and the LPC analysis <i>"back"</i> 30 order analysis every 80 samples.

A separation for coding the residue into two sub-blocks of 40 samples was carried out. Each block B_not included 80 samples.

Adaptation of the stationary parameter STAT (n)

The above stationarity parameter varies between two extreme values ​​0 and 100, the STAT values_m and STAT_M mentioned above.

The adaptation functions previously described in the description, and in particular the functions f_at(N_BWD) and f_b(N_BWD) were such that:<maths id="math0005" num=""><img file="EP0906613B1_D0005.tif" /></maths><maths id="math0006" num=""><img file="EP0906613B1_D0006.tif" /></maths> In these relationships, x = DGfb.

Regarding the function f_r allowing to establish the refinement value ΔS previously mentioned in the description, this is a step function of the variable x, with x = Gpb - Gpf and ΔS = f_r(x) and having the value:<maths id="math0007" num=""><img file="EP0906613B1_D0007.tif" /></maths>

The refining of STAT (n) is also subject to the following conditions mentioned above in relation to FIG. 2c: If ΔS&gt; 0:       If STAT * (n) &lt;STAT_i STAT (n) = STAT * (n) + ΔS       Otherwise STAT (n) = STAT * (n) If not :       STAT (n) = STAT * (n)       with STAT_i = 40,6.

The other test conditions referenced 1121d, 1121c and 1121f in FIG. 2c were not used in this embodiment.

Adaptation of decision thresholds

Regarding decision thresholds: S_PRED is adapted as follows: S_PRED (n) = 0.03.STAT (n) + 1.0 S_PRED ∈ [S_PRED_m, S_PRED_M], S_PRED_m = 1.03 and S_PRED_M = 4; The threshold S_LSP_L is adapted using the following staircase function:<maths id="math0008" num=""><img file="EP0906613B1_D0008.tif" /></maths> The value of the threshold S_STAT used in case of stationarity of the LPC filters measured using the threshold S_LSP_L was fixed at 4.0 dB. The threshold S_LSP_H was not used in this embodiment. The value of the threshold G₁ was attached to OdB. Regarding the energy value characterizing an ENER_SIL silence frame, this value was fixed at 40 dB measured on the 80 samples s (i) of the current block B_not :<maths id="math0009" num=""><img file="EP0906613B1_D0009.tif" /></maths>

Regarding the value of the threshold S_FWD mentioned above and intended to further limit the risk of failover by imposing the LPC filtering mode "<i>before</i>"when the value STAT (n) is lower than this threshold, this value S_FWD was set at 40.6.

*

A second embodiment of a CELP type encoder in a band extended to two 16/24/32 kb / s sub-bands was carried out under the conditions below

:

<ul id="ul0014" list-style="dash" compact="compact"><li>wideband encoder from 0 to 7000 Hz in two sub-bands. A main band was coded with the CELP technique, frame of 120 samples, excitation created by algebraic dictionaries, and transmission of certain energy and spectrum characteristics of a host band between 6000 Hz and 7000 Hz.</li><li>LPC analysis "<i>before</i>"with 14 coefficients and LPC analysis"<i>back</i>"at 50 coefficients every 120 samples. In LPC analysis mode"<i>before</i>", separation into two LPC sub-blocks of 60 samples, the filter used for the first sub-block being interpolated from the current filter and the previous filter.</li></ul>

Calculation of the stationary parameter STAT (n)

In this embodiment, the above-mentioned stationarity parameter varies between the two extreme values ​​0 and 120, the STAT values_m and STAT_M mentioned above.

As regards the adaptation of the value of the stationary parameter STAT (n), the values ​​of the functions f_at(N_BWD) and f_vs(N_BWD) are such that:<maths id="math0010" num=""><img file="EP0906613B1_D0010.tif" /></maths><maths id="math0011" num=""><img file="EP0906613B1_D0011.tif" /></maths>

Regarding the function f_r used to establish the refinement value ΔS previously mentioned in the description, this is a step function of the variable x, with x = Gpb / Gpf and ΔS = f_r(x) and having the value:<maths id="math0012" num=""><img file="EP0906613B1_D0012.tif" /></maths>

The refining of STAT (n) is also subject to the following conditions mentioned above in relation to FIG. 2c: If ΔS&gt; 0:   If Gpb&gt; If       If STAT * (n) &lt;STAT_i STAT (n) = STAT * (n) + ΔS       Otherwise STAT (n) = STAT * (n)    Otherwise STAT (n) = STAT * (n) If not :       If STAT * (n) &lt;STAT_i STAT (n) = STAT * (n) + ΔS       Otherwise STAT (n) = STAT * (n) with STAT_i = 80, S_i = 0dB.

The other test conditions referenced 1121h and 1121d in FIG. 2c were not used in this embodiment.

Adaptation of decision thresholds

Regarding decision thresholds: S_PRED is adapted as follows: S_PRED (n) = 0.03.STAT (n) - 0.5 bounded in the interval [S_PRED_m, S_PRED_M], with S_PRED_m = 0.5 and S_PRED_M = 2.5. The threshold S_LSP_L is adapted using the following staircase function:<maths id="math0013" num=""><img file="EP0906613B1_D0013.tif" /></maths> The threshold S_LSP_H is adapted using the following staircase function:<maths id="math0014" num=""><img file="EP0906613B1_D0014.tif" /></maths> The value of the threshold S_TRANS used in the event of transition of the LPC filters measured using the threshold S_LSP_H was fixed at 0 dB. The value of the threshold S_STAT used in case of stationarity of the LPC filters measured using the threshold S_LSP_L was fixed at 2.5 dB. The value of the threshold G₁ has been set to 0dB. Regarding the energy value characterizing an ENER_SIL silence frame, this value was fixed at 50 dB measured on the 120 samples s (i) of the current block B_not :<maths id="math0015" num=""><img file="EP0906613B1_D0015.tif" /></maths>

Regarding the value of the threshold S_FWD mentioned above and intended to further limit the risk of failover by imposing the LPC filtering mode "<i>before</i>"when the value STAT (n) is lower than this threshold, this value S_FWD was set at 60.