EP2345029A1

EP2345029A1 - Critical sampling encoding with a predictive encoder

Info

Publication number: EP2345029A1
Application number: EP09755960A
Authority: EP
Inventors: Pierrick Philippe; David Virette
Original assignee: France Telecom SA
Current assignee: Orange SA
Priority date: 2008-10-08
Filing date: 2009-10-05
Publication date: 2011-07-20
Anticipated expiration: 2029-10-05
Also published as: FR2936898A1; EP2345029B1; US20110178809A1; WO2010040937A1; US8880411B2; ES2542067T3; CN102177544A; CN102177544B

Abstract

The invention relates to a method for encoding and decoding a digital audio signal, said method comprising the steps of: encoding a first sequence of samples of the digital signal according to a transform encoding; encoding a second sequence of samples of the digital signal according to a predictive encoding; wherein the second sequence starts before the end of the first sequence, a subsequence common to the first and second sequences being thus encoded both by predictive encoding and by transform encoding.

Description

Critical sampling coding with predictive coder

The present invention relates to the field of coding digital signals.

The invention advantageously applies to the coding of sounds with alternating speech and music.

To effectively code speech sounds, CELP (Code Excited Linear Prediction) techniques are recommended. To effectively encode musical sounds, transform coding techniques are preferred.

CELP coders are predictive coders. They aim to model the production of speech from various elements: a long-term prediction to model the vibration of vocal chords in voiced period, a stochastic excitation (white noise, algebraic excitation), and a short prediction -term to model vocal tract changes.

Transform coders use critical-sampling transforms to compact the signal in the transformed domain. A "critical-sampling transform" is a transform for which the number of coefficients in the transformed domain is equal to the number of coefficients of the digitized sound.

One solution for effectively coding a signal containing these two types of content is to select the best technique over time. This solution has been recommended by the 3GPP ("3rd Generation Partnership Project") standardization organization, and a technique called AMR WB + has been proposed. This technique is based on CELP technology of the AMR WB type and transform coding based on a Fourier transform overlay.

This solution suffers from insufficient quality on the music. This deficiency comes particularly from transform coding.

Indeed, the overlapping Fourier transform is not a critical sampling transformation, and therefore, it is suboptimal.

In addition, the windows used in this encoder are not optimal vis-à-vis the concentration of energy: the frequency forms of these windows are relatively fixed.

Critical sampling transformations are known. For example, the transforms used in MP3 and AAC type music encoders. These transforms are based on the formalism called TDAC ("Time Domain Aliasing Cancellation").

The use of TDAC makes it possible to obtain excellent quality on the music. Nevertheless, this has the disadvantage of introducing temporal folds that make it difficult to combine with CELP type technologies.

Indeed, during a transition from TDAC to CELP the temporal folding of the TDAC portion is not canceled by the signal from the CELP which does not integrate any folding.

An object of the present invention is to propose a technique for reconstructing an audio signal, with good quality, by alternating transform coding techniques (for example with critical sampling) and predictive coding techniques (for example of the CELP type ). For this purpose, the present invention proposes a method of encoding a digital signal, comprising the steps:

encoding a first sequence of samples of the digital signal according to transform coding;

coding a second sequence of samples of the digital signal according to a predictive coding;

and wherein the second sequence begins before the end of the first sequence, whereby a subsequence common to the first and second sequences is encoded by both predictive and transform coding.

Thus, during the decoding of the digital audio signal, the folding created by the coding in the subsequence of the first sequence can be suppressed by means of samples of this subsequence resulting from the decoding of the sub-sequence within the second sequence. In addition, the second sequence can be decoded because the samples of the past, useful for predictive decoding, do not have this folding.

Advantageously, the transform coding is a critical sampling transform coding.

For example, transform coding is a TDAC type transform coding.

For example, predictive coding is a CELP encoding.

In an advantageous embodiment, the transform coding of the first sequence comprises the application of an analysis window making it possible to deduce from a perfect reconstruction relation of the digital signal a synthesis window comprising at least three parts: a first nominal part,

a second terminal portion substantially zero,

a third substantially continuous intermediate portion between the first and second portions.

It is then expected that at least the parts of the analysis window making it possible to deduce respectively the second and third parts of the synthesis window are applied to the subsequence common to the two sequences.

"Substantially continuous" means that the third part makes it possible to have no discontinuity between the first and second parts. Indeed, this type of discontinuity reduces the quality of decoding by adding decoding noise.

The perfect reconstruction relation imposes a relation between the forms of the windows of analysis and synthesis. In addition, when switching between a transform coding and a predictive coding, it is possible to describe the analysis window or the synthesis window in an equivalent manner. Indeed, in this case, the relation of reconstruction makes appear a direct relation between the two forms.

With an analysis window (and therefore synthesis) thus chosen, it is possible to reduce the area in which the folding appears on the decoding of the first sequence.

With the window thus defined, it is possible to reduce the number of samples of the second sequence (predictive coding) to be transmitted for decoding.

In addition, the additional number of samples is related to the size of the intermediate part. For example, the middle part is a sinus arch. For example again, the intermediate part is a function derived from "Kaiser-Bessel". In addition, it can come from a window optimization calculation and not have an explicit expression.

For example, the summary window is an asymmetric window.

Thus, it is possible to adapt the profile of the synthesis window (thus the analysis window) to the coding of the following sequence or preceding the first sequence.

In an advantageous embodiment, the synthesis window further comprises a fourth continuous initial portion between a substantially zero value and a non-zero value of the first part.

Thus, it is possible to minimize the impact of the transition between transform coding and predictive coding on transform coding.

For example, the fourth part of the synthesis window is a smooth transition between an initial value and a value of the nominal part, and the third part is an abrupt transition between a value of the nominal part and a value of the substantially zero part. .

Thus, a better concentration of the signal energy in the frequency domain is obtained for a better coding efficiency of the transformed part.

It can be provided that the first and second sequences belong to the same frame of the digital signal. Thus, the encoding of the first sequence can be used as a transition encoding after the encoding of a frame by transform coding. This makes it possible to improve the coding efficiency by not disturbing this frame.

The present invention also provides a method for decoding a digital signal, comprising the steps of:

receiving a transformed vector encoding a first sequence of samples of the digital signal according to transform coding;

receiving a prediction vector coding a second sequence of samples of the digital signal according to a predictive coding;

wherein the second sequence begins before the end of the first sequence, a common subsequence to the first and second sequences thus being received encoded by both predictive and transform coding; and which further comprises the steps:

a) applying to the transformed vector an inverse transformation of the transform coding for decoding a subsequence of the first non-coded sequence by predictive coding;

b) decoding at least in the prediction vector the subsequence common to the first and second sequences at least by a predictive decoding based on at least one sample from step a);

c) decoding in the predictive vector by a predictive decoding a subsequence of the second non-coded sequence by transform coding, based on at least one sample from one of the steps a) and b). Thus, the folding present in the decoded subsequence can be eliminated by using decoded samples by predictive decoding.

In an advantageous embodiment, step b) comprises the sub-steps:

b) decoding in the predictive vector the common subsequence to the first and second sequences by a predictive decoding based on at least one sample from step a);

b2) applying to the transformed vector an inverse transformation of the transform coding for decoding the common subsequence to the first and second sequences; and

b3) decoding the common subsequence to the first and second sequences by combining at least one sample from step b1) with a corresponding sample from step b2).

For example, the combination is a linear combination. By thus combining the samples, a more robust decoding is obtained.

In another advantageous embodiment, step b) comprises the sub-steps:

b4) decoding in the predictive vector the common subsequence to the first and second sequences by predictive decoding based on at least one sample from step a); b5) creating from at least one sample from step b4) a sample containing a folding equivalent to a transform coding followed by a transform decoding;

b6) applying to the transformed vector an inverse transformation of the transform coding for decoding the common subsequence to the first and second sequences; and

b7) decoding the common subsequence to the first and second sequences by combining at least one sample from step b5) with a corresponding sample from step b6).

Thus, the folding created by step b5) corresponds exactly to the folding present in the decoded subsequence.

The creation of the folding can be done by applying a matrix representing direct and inverse transformation operations. Such a matrix may be equivalent to the application of transform coding immediately followed by transform decoding.

Of course, one can use the same predictive coding for all samples.

Similarly, the same transform coding / decoding can be used, with the same analysis and synthesis windows, whenever such coding / decoding is performed.

In one embodiment, step a) comprises the application of a synthesis window comprising at least three parts:

a first nominal part,

a second terminal portion substantially zero, a third intermediate portion continuous between the first and second zones,

and at least the second and third portions are applied to samples coding the subsequence common to both sequences.

The present invention provides a computer program comprising instructions for implementing the encoding method as described, when the program is executed by a processor.

In addition, the present invention aims a support readable by a computer on which is recorded such a computer program.

The present invention also provides a computer program comprising instructions for implementing the decoding method as described, when the program is executed by a processor.

The present invention provides a coding entity adapted to implement the coding method as described.

Such a coding entity of a digital audio signal may comprise:

a transform encoder for encoding a first sequence of samples of the digital audio signal according to transform coding;

a predictive encoder for encoding a second sequence of samples of the digital audio signal according to a predictive encoding; it is expected that the second sequence begins before the end of the first sequence, a subsequence common to the first and second sequences thus being encoded by both predictive coding and transform coding.

The present invention provides a decoding entity adapted to implement the decoding method as described.

It is possible to provide a digital signal decoding entity comprising receiving means:

a transformed vector encoding a first sequence of samples of the digital signal according to transform coding; and

a prediction vector encoding a second sequence of samples of the digital signal according to a predictive coding;

wherein the second sequence begins before the end of the first sequence, whereby a common subsequence to the first and second sequences is encoded by both predictive and transform coding; and which further comprises:

a first decoder for applying to the transformed vector an inverse transformation of the transform coding for decoding a subsequence of the first non-coded sequence by predictive coding;

a second decoder for decoding at least in the predictive vector the subsequence common to the first and second sequences by at least one predictive decoding based on at least one sample derived from the first transform decoder; and a third predictive decoder for decoding in the predictive vector by a predictive decoding a subsequence of the second non-coded sequence by transform coding, based on at least one sample derived from one of the first and second coders.

In an advantageous embodiment, the second decoder comprises:

first means for decoding in the predictive vector the common subsequence of the first and second sequences by a predictive decoding based on at least one sample derived from the first transform decoder;

second means for applying to the transformed vector an inverse transformation of the transform coding for decoding the common subsequence to the first and second sequences; and

third means for decoding the common subsequence to the first and second sequences by combining at least one sample from the first means with a corresponding sample from the second means.

In another advantageous embodiment, the second decoder comprises:

fourth means for creating, from at least one sample returned by the first means, a sample containing folding equivalent to transform coding followed by transform decoding;

fifth means for applying to the transformed vector an inverse transformation of the transform coding for decoding the common subsequence to the first and second sequences; and

sixth means for decoding the common subsequence to the first and second sequences by combining at least one sample from the fourth means with a corresponding sample from the fifth means.

Of course, all means performing the same type of coding or decoding (predictive or transform) can be combined in one unit.

Similarly, one can provide a single unit (coding or decoding) to perform sometimes coding, respectively decoding, predictive and transform.

Of course, the encoders / decoders described may comprise a signal processor, storage elements, as well as means of communication between these elements.

The present invention thus makes it possible to alternate transformation coding techniques, for example with TDAC-type critical sampling, and predictive coding techniques, for example of the CELP type over time in order to obtain a good quality of reconstruction.

To this end, the invention proposes particular temporal relations between the two types of coding: the time position of the CELP and transformed frames being temporally offset. In advantageous embodiments, the invention also proposes extending the duration of the frames, or sequences covered by the CELP coding, by an overlap, during a transition from transform to CELP. This duration can be variable in time if the transform requires a good frequency concentration.

The duration of use of the CELP coding can be variable from one frame to another, in order to quickly adapt the coding technique to changes in the nature of the sounds.

According to an advantage of the present invention, a frame of M samples can be subdivided into several subframes mixing portions encoded in CELP and others in the transformed domain.

The invention finds its application in sound coding systems, in particular in standardized speech coders, in particular to the ITU ("International Telecommunications Union") or to the ISO ("International Standard Organization") for coding generic sounds, including speech signals.

Other features and advantages of the invention will appear on examining the following detailed description, and the appended figures among which:

FIG. 1 illustrates two windows of synthesis of a transform coding,

FIG. 2 illustrates synthetic windows of one embodiment of the invention,

FIG. 3 illustrates data frames processed by synthesis windows, FIG. 4 illustrates sample vectors obtained by application of synthesis windows,

FIG. 5 illustrates the case of a TDAC coding followed by an AMR WB coding, followed by a TDAC coding according to one embodiment of the invention,

FIG. 6 illustrates the same case of coding with an advantageous asymmetric window,

FIG. 7 illustrates a general context of a problem solved by the invention,

FIG. 8 illustrates a general scheme for solving this problem by the present invention,

FIG. 9 illustrates the steps of an embodiment of a coding method according to the invention,

FIG. 10 illustrates the composition of a synthesis window according to one embodiment of the invention,

FIG. 11 illustrates the steps of an embodiment of a decoding method according to the present invention,

FIG. 12 illustrates an advantageous decoding used in the decoding method,

FIG. 13 illustrates a variant of this advantageous decoding,

FIG. 14 illustrates an encoder according to one embodiment of the invention, FIG. 15 illustrates a decoder according to one embodiment of the invention,

FIG. 16 illustrates a hardware device adapted to produce an encoder or a decoder according to an embodiment of the present invention.

In the following, we begin by describing a transformation

TDAC with perfect reconstruction, then we present a technique to make it compatible with critical sampling.

Finally, a CELP coding and a combination of this coding with TDAC coding are described.

TDAC and perfect reconstruction

A digitized sound signal is considered according to a sampling period - (F _e being the sampling frequency). For

a given frame of index t, the samples are denoted χ _{n + [M} for each moment n + tM.

The expression of the TDAC transform to the frame encoding is shown below:

2M-I

* ,, * = Σ -W MO 0 ≤ k <M, n = 0

M represents the size of the transform,

- X _{1 k} are the samples in the transformed domain for the frame t,

Pk ( ⁿ ) = K (") QA ( ^" ) ^cos - (2n + 1 + Λ /) (2 * + 1) is a basic function of the transform of which: the term Λ _u (/ i) is called a prototype filter or

"analysis weighting window" and covers 2M samples,

and whose term C _nk defines the modulation.

To restore the initial time samples, the following inverse transform, at decoding, is applied in order to reconstruct the samples O ≤ n <M which are then in a recovery zone of two consecutive transforms. The decoded samples are then given by:

M -I xn, + t / H + M = Σ [* r ₊ u />; ( ") + * ,. _* />;("+ ^)],

* = 0

where p ^s (n) = h _s (n) C _{n <k} defines the synthesis transform, the synthesis weighting window being denoted h _s (n) and also covering 2M samples.

The reconstruction equation giving the decoded samples can also be written in the following form:

M -I

^X _{n +} M ₊ M ^≈ Σ [x, ₊ uMn) Q. _n ⁺ * αM "+ ^Λ θc _{tin + M} ]

= h _s (n) X, _Λ C _{kjl +} M

This other presentation of the reconstruction equation amounts to considering that two inverse cosine transforms can be carried out successively on the samples in the transformed domain X _tk and X _{t + lk} , their result being then combined by a weighting and addition operation. It is the addition of two consecutive frames which makes it possible to suppress the so-called folded components of the transformation. Indeed if one writes the operations of transformation direct and inverse in matrix form for the frames t = 0 and t = 1 one has:

/ z _û0 (2M-1) ^X 2M-1

At the synthesis, we get

- "oo / z _{(0 (0) 0 0} Q 00 10 C ₁ CM, -I 0 00 - * bo 0 ₁₀ (I) 01 C Q 0 ₁ CM, -I 1 01

^ 02M-I 0 ₄₀ (2AF -1) 02M-1 Q ₁ I C 2M-1 Cj 2.MM --1M X-1, OM-I

-I

With

Q 00 C, 10CM ₁ -IO Q ₀ C ₀₁ Co 2M-I

Q 01 C ₁ CM ₁ -IIQ ₀ c I, IQ 2M-I

5 =

Q 02M-IQ 2M-IC ₁ 2M-I MI C "-MI 0 C" -Af-I 1 CM ₁ -I 2M-I

S = 'M JM' M 'M' M + JM Where I _M is the square matrix identity of size M,

Where J _M is the anti-identity square matrix of size M, which, at a series of increasing index values, returns the same series of values with the decreasing indices,

- 0 _M is a square matrix of size M containing only zeros.

So, he comes:

J

and by analogy using the t = 1 frame:

J "al, 2M - \ - n ^x 3M - \ - n}

Thus, if we add \ _{M + n} and Jc ₁ "term to term we obtain:

^x M + n ~ -Xo.M + π ^{+ x} \, n ~ "tO, M + n ⁺ Kθ, 2M - \ - n ^x 2M - \ - ny [} ^ι a \ n ^x M + n ~ ' ^ι a \, M - \ - n ^x 2M - \ - n J ^x M + n J ^{+ X} 2M -] - n J

If we want to ensure χ _{M + n} = χ _{M + π} and thus obtain the perfect reconstruction, we obtain the following necessary conditions on the analysis and synthesis filters: ⁼ ®

that is to say

with

D {n) = h _{a0 n + M) - h {M -l _al -n) + h _s (n) - h _a0 (2M -l -n).

It appears that to ensure perfect reconstruction, the forms of analysis and synthesis are constructed by temporal reversal and weighting. Consequently, if h _s contains zeros n, then h _a will contain the symmetrical portion about M / 2, that is to say the index M- l -n.

The synthesis is illustrated by an example in FIG. 1. In this example, two inverse transforms of size M h _s o and h _s i are passed.

To reconstruct the samples between M and 2M-1, the samples covered by the common part are added between h _s o and h _S ). The reconstruction will be perfect if the windows verify the perfect reconstruction conditions stated above.

The usual case of reconstruction is therefore when, in a decoder, two consecutive spectra are received, for example X _t and X _{t +} 1, resulting from direct transformations and the inverse transformations are applied to them to obtain Jc ₀ and Jc _{1, respectively} . The original signal will be perfectly reconstructed by adding the last M samples of the first set with the first M of the second.

It can also be considered that X _t alone has been transmitted. The perfect reconstruction can be obtained if we know how to construct the signal Jc _{1 n} . This will be possible if we know the samples X _M to X _2M i -

In this way it will be possible, by weighting by the windows h _s i and h _a i _, to construct the vector for suppressing folding from the vector X ₀ .

In the foregoing, it has been considered that the signals X _t and X _{M are available} at X ₂ M i -

If now we consider that the following frame is transmitted in the frequency domain (X _{t + 2} ), we do not suppress the folding located between X ₂ M to X _3M - _I - To do this, it would have been necessary to have received these samples at prior. Nevertheless, this trivial solution is suboptimal from the point of view of critical sampling.

In the following, we present a way to overcome this disadvantage.

Effective time coding

It is proposed to choose particular windows for transmitting the coded signal in time when desired without losing the critical sampling (ie the same number of samples transmitted and reconstructed). This is illustrated in Figure 2.

By construction, as illustrated in FIG. 2, one chooses:

hsO = 0 for n between M + (M + Mo) / 2 and 2M-1, and

hs l = 0 for n between 0 and (M-Mo) / 2,

with M ₀ a given integer value between 1 and M-I.

For example, the downward and upward portions of hsO and hs l around the M + M / 2 sample consist of sinus arches given by the equation: h _s i (n) = sin (pi * (0.5 + n - ((M-Mo) / 2)) / 2 / Mo) for n between (M-Mo) / 2 and (M + Mo) / 2

h _s o (n) will be taken as symmetric in this area of h _s i to obtain the perfect reconstruction.

h _s i can be defined also by a function of "Kaiser

Bessel "derivative used for example in AAC coders.

Thus defined, the forms of h _s0 and h _s i make it possible to ensure perfect reconstruction.

As illustrated in FIG. 3, a first frame T30 (windowed by h _s o) combined with the frame T3 1 (windowed by hs l) makes it possible to reconstruct the segment from M to 2M-1, the frames T3 1 and T33 enabling 'get the samples 2M to 3M- 1 etc.

In the case where the signal of the frame T31 is transmitted in frequency, the critical sampling is respected and the reconstruction is perfect insofar as the analysis and synthesis filters verify the necessary condition.

Since the sample X3 _M / 2 ₊ 11 (n <Mo / 2) is transmitted in the frame T3 1 then the sample X ^ _{M / 2} i -n can be generated based on the knowledge of χa _{M + M} i _{2 + n} ^{1 SSU} of the T30 frame. We will build on the relationship:

- * 0 M + n => \ 0 M + "M +"% + "+ Λo IM 1 n ^x 2M l- ,,] P ^{0 Ur} Ω = M / 2

We will then have: x H) W / 2 + n

3M / 2 1-π K MU MU / 2 + n ^x ÏM / 2 + n

"α0 W / 2-l n SO W / 2 + n

This can be reiterated to find the samples in the overlap zone, ie between the samples (M-Mo) / 2 and M / 2. Using predefined relationships:)

Since h _s o contains zeros between M + (M + M ₀ ) / 2 and 2M-1, h _a i contains zeros between 0 and (MM _o ) / 2.

Similarly, because h _s i contains only zeros between 0 and

(M-Mo) / 2, h _a o contains only zeros between M + (M + Mo) / 2 and 2M-1.

hsO = 0 for n = M + (M + Mo) / 2 ... 2M-1,

hs l = 0 for n = 0 ... (M-Mo) / 2,

ha l = 0 for n = 0 ... (M-Mo) / 2,

haO = 0 for n = M + (M + Mo) / 2 and 2M-1.

Therefore, as illustrated in FIG. 4, the vector * _{o, M + n} contains 3 zones:

- x _{oM + n} = 0 of n = (M + Mo) / 2 ... M-1,

- χ _{$ M + n} does not contain folded components between n = 0 and n = (M-Mo) / 2, and

the central zone around M + M / 2 for which there are folded components.

Likewise:

- Jt _{1 n} = 0 between n = 0 and n = (M-Mo) / 2,

- x _lπ does not contain folding components between (M + Mo) / 2 and M-I, and the central zone around M / 2 for which there are folded components.

Thanks to these properties, we can recover the segment XM - - X2M _i ensuring a perfect reconstruction

This perfect reconstruction can be obtained:

by transmission in the transformed domain of the vector Xi,

- by transmission in the time domain of the samples

X3 M / 2 - X5M / 2 1

From the foregoing, one can now perform critical sampling TDAC coding while avoiding problems related to aliasing. In the following, a CELP coding is described, allowing an advantageous combination with the previously described TDAC coding.

TDAC + CELP

It is recalled that one places oneself within the framework of a type operation presented in the AMR WB + specification. A converted type coding using TDAC is alternated with a time-type coding which consists of a CELP coder (for example according to AMR recommendation WB).

Without loss of generality, with reference to FIG. 5, we take the case of a coding of a T51 frame by TDAC (windowed by h _{5 1} ) followed by a T52 frame in AMR WB and then a T53 frame to new in TDAC (windowed by hr ₅₃ ).

In order to reconstruct the samples, the AMR WB coding is based on a prediction of the periodicity of the signal, termed long-term prediction. As such, it builds its samples as follows The signal r is constructed with respect to old samples taken upstream of T samples weighted by a gain, transmitted and updated periodically, and a so-called stochastic part W _n assigned a gain b, transmitted and updated also over time. T represents the pitch. The AMR encoder WB estimates the components a, b and T and the part W _n to be added according to the flow rate considered.

Thus, in order to effectively achieve long-term prediction, the CELP decoder uses past samples that should not have artifacts. However, since the frame T51 is coded in TDAC, there will be folding in the samples between M + (M-MQ) H and M + (M + Mo) / 2 as long as the frame T52 will not be restored with the folding to delete that of the frame T51.

In order to allow the return of the samples of the CELP-encoded frame T52 without folding, the area of coverage of the samples transmitted by this coding is widened to cover the initial transition zone completely.

The duration of the CELP is extended to the content of index M + (M-Mo) / 2 ... 5M / 2.

In this sense, there is no critical sampling for the coded part by the predictive coding.

On the other hand, the zone M ₀ is limited in duration in order to avoid transmitting too much additional information.

For example, M ₀ is around 1 to 2 ms for a frame of duration M corresponding to 20 ms. The number of samples is calculated based on the sampling frequency. We can also choose Mo / 2 as a duration proportional to a subframe of CELP, that is to say the usual duration of updating pitch / gain values and stochastic vector, where a size adapted to fast algorithms for the search of the stochastic vector and its transmission effectively. For example, we take a power of 2.

To reconstruct the samples of the area between M and 2M-1, the period between M and (M-Mo) / 2 is reconstructed using the inverse transform of a frame T50 (not shown) preceding the frame T51. Then the area between M + (M-Mo) / 2 and M-I is reconstructed with the CELP alone which is based for the long term on the samples returned by the transformed part.

A variant to obtain samples between M +

(M-Mo) / 2 and M + (M + Mo) / 2- 1 is to combine the samples

CELP with samples containing folding from the frame

15 T51. In this case, we can achieve a linear combination of samples from the CELP and the previously determined equation

The linear combination works according to the model below:

- ^ ^"x LMI2 - \ - n ^{= a} n ^X 3M 12-ln ⁺ \) ~ ^a n) ^x ~ iM-12 H ₁ from CELP from the transformed

With α ,, a set of positive or zero coefficients less than or equal to one.

The portion 2M, ... 3M-1 is decoded using the end of the CELP samples transmitted between the indices 2M to 5M / 2. Then, in

Based on this decoded result, the samples from the following transform are reconstructed in the overlap area, which contains folding in a similar manner to the overlap area between the frames T51 and T52. The difference with the other direction of transition lies in the fact that the CELP will not supply all the samples of the transition zone of the transform, but only a half (ie M'o / 2 = M / 8 in our example for a transition size of M'o = M / 4). However, only half of this transition zone is necessary to be able to cancel the time folding of the transform.

The window hs i may be asymmetrical. Thus, the overlap zone between the CELP and TDAC part, denoted M ₀ ', may be different from M ₀ .

Transmission of CELP

Several alternatives are described hereinafter for transmitting the CELP frame.

In one embodiment, the CELP frame covers a duration equal to the size M + Mo / 2 as presented in FIG. 4. According to the AMR standard WB, this frame is divided into sub segments, of size denoted Mc in FIG. , allowing a frequent update of the parameters making it possible to synthesize a signal CELP of quality. Thus the values of pitch, gain and the stochastic part are initially transmitted and updated optionally.

The length of the first sub-segment (Mc ') immediately following the transform may be different if it is desired to use an arbitrary length Mo' with a standard CELP encoder with Mc imposed by this standard.

The pitch can be estimated on the decoded part before the sample of index M + (M-Mo) / 2. Thus, it is possible to avoid transmitting the initial pitch, only the pitch gain that is estimated according to the common method presented in AMR recommendation WB is transmitted.

In a variant of this embodiment, the pitch gain is not transmitted. It is estimated on the decoded signal in the transformed part.

In an alternative embodiment, the pitch estimate can be performed by including the period M + (M-Mo) / 2 to M + (M + Mo) / 2 which contains folded components.

The stochastic part is transmitted in preamble, or ignored. And this, especially if it is considered negligible because of its low power, or if during the reconstruction is based on the version using the weighting a _n .

Indeed, a stochastic part is implicitly present in the signal coming from the folded components coming from the transformed part.

The portion of duration Mo / 2 covered by the CELP can therefore be a specialized part, in that it can benefit from the information resulting from the complete decoding of the part resulting from the previous transform.

Mo / 2 can be equal to Mc if one looks for a particular compatibility with an existing coder. For example, in the context of an embodiment including a CELP AMR type WB, one can choose Mo / 2 = Mc = 5 ms.

An alternative embodiment is shown in Figure 6. In this embodiment, the CELP encoding covers a shorter length than the base frame of length M. The covered portion by the samples M + (MM / 2) / 2 to 2M + M / 16 is encoded from a transform of a size shorter than the initial size (M / 2).

In FIG. 6, only the T63 frame is coded in CELP. The frames T61, T62 and T64 are represented in the transformed domain of the TDAC. The frames T61 and T64 are encoded with transformations of length M (windows h ₆ ι and h ₆₄ ), the frame T62 being encoded with a transform of size M / 2 (window h ₆₂ ).

This coding is effective because the window h ₆ i is relatively soft, which allows to obtain a better concentration of energy in the frequency domain. The h ₆₂ window, on the other hand, has a steeper transition in the vicinity of the 2M sample, but this steep window does not penalize the quality of the coding too much because temporally the duration affected is short. The T63 is coded in CELP as shown above, here Mo = M / 8.

Thus a frame of length M can be subdivided into subparts coded in CELP or TDAC of variable size.

Once the samples are restored in the time domain, LPC synthesis filters can optionally be applied to restore the sound signal if necessary.

In a particular embodiment, the transform is operated in a weighted domain, that is to say that the transform is performed on the filtered signal by a weighting filter of type W (z) = A (z IY _\ ) H _of _ _emph (z) with A (z) the linear prediction filter

(LPC) and gamma a flattening factor of this filter, the filter Hd _e - emph (z) is a high frequency de-emphasis filter. The CELP coder operates it, that is to say that the excitation signal r _n will be well calculated in the residual domain of a linear prediction filter A (z). Special attention will be paid for the signal synthesized by the first inverse transform, and which is therefore in a perceptually weighted domain, is returned to the field of CELP excitation, so that the long-term portion of the CELP excitation can be calculated.

In the following an embodiment of the coding method is described.

Referring to Figure 7, we illustrate the problem of switching between a type of coding transformed with a predictive type coding.

We consider a signal x to code and then decode. Samples from 0 to 3M-1 are considered to be transform-coded, whereas samples from 3M to 4M-1 must be coded by predictive coding, as indicated by the double arrows T and P.

According to the prior art, the samples are coded from 0 to 2M-1 by transform coding according to a transformed vector X _o .

Decoding of this transformed vector gives the samples of

0 to 2M- 1 of a decoded signal *. This decoding reveals REP 1 refolding, especially in the samples from M to 2M-1.

Moreover, the samples of M to 3M-1 are coded by transform coding according to a transformed vector Xf.

Decoding of this transformed vector gives the samples from M to 3M-1 of the decoded signal *. This decoding shows the same folding with a sign opposite REP 1 in the samples from M to 2M-1 as during the decoding of x £. It also shows REP2 folding in samples from 2M to 3M-1 in *. Thus, it is possible by combination of the samples of M to 2M-1 respectively from the decoding of X _Q and x [delete (DELPR_REP) folding REPL.

The samples of 3M to 4M-1 of x are then coded by predictive coding according to the prediction vector Xζ.

To be decoded, this vector requires knowledge of previous samples. That is to say the samples from 2M to 3M-1. These samples are available at the decoding of X, ^r , nevertheless they are unusable because of the presence of REP2 folding.

Thus, Xf can not be decoded.

In addition, the removal of REP2 folding requires knowledge of x samples from 2M to 3M-1 to recreate the folding and delete it by combination. However, these samples are not available for decoding.

Thus, the decoding of Xf is not finished.

To solve these difficulties, the prior art proposes to communicate to the decoder the samples it needs in addition to the vectors from the transform and the prediction part.

Nevertheless, this solution is not optimal from the point of view of flow.

The present invention proposes the solution illustrated in FIG.

In this figure, we find the signal x, the transformed vector X, ^r , and the prediction vector X%. However, according to the present invention, the prediction vector Xξ encodes a number M of samples comprising part of the samples coded by xj.

This arrangement makes it possible to reconstruct the signal x at decoding.

Indeed, the samples preceding the REP folding creates the decoding of X, ^r are used for the decoding of the first samples that the decoding of Xξ will make it possible to obtain. That is, those he has in common with x, ^τ .

Thus, samples of x are retrieved to recreate REP folding. For example, the samples of x corresponding to REP are subjected to coding followed by a decoding identical to those undergone by the samples from M to 3M-1.

This folding thus created is combined with that present in the samples resulting from the decoding of x [, and xj can thus be completely decoded.

Then, the completely decoded M to 3M-1 samples can be used to decode x £.

In the following, with reference to Figure 9, there is described a coding method incorporating the principles described above.

In step S90, samples of a signal to be coded are received.

Then, in step S91, two sample sequences are delimited, so that the second sequence begins before the end of the first sequence. A first sequence SEQ 1 and a second sequence SEQ 2 are thus obtained. Each of these sequences is then coded according to a transform coding in step S93 for SEQ 1, and according to a predictive coding in step S94 for SEQ2.

With reference to FIG. 10, an embodiment is described in which the transform coding is done by applying an analysis window, making it possible to determine a synthesis window, by means of a perfect reconstruction relation, adapted to the present coding.

The windows of analysis and synthesis being linked by the perfect reconstruction relation, it is equivalent to describe one or the other.

In FIG. 10, the synthesis window H is described. This window comprises four particular parts.

INIT corresponds to the initial part of the filter, one chooses this part according to the coding of the preceding samples. For example, here, H makes it possible to reconstitute a part of SEQ 1 (samples 0 to M-I). If the samples preceding SEQ 1 are coded by transform, we advantageously choose INIT as a smooth transition. This makes it possible not to disturb these previous samples.

NOMI corresponds to a nominal part. Advantageously, this portion takes a substantially constant value.

NL corresponds to a substantially zero portion of the window. The duration of NL (or the number of coefficients of NL) can advantageously be chosen as a function of the duration (or number of coefficients) of NOMI.

Finally, the INTER part is a continuous part between NOMI and NL. This part can have a shape adapted to the transition between the transform coding of SEQ1 and the predictive encoding of SEQ2. For example, it's a relatively abrupt transition.

Thus, INIT and NOMI are applied to the sub-sequence S-

SEQ1 of SEQ1 which does not include a sample of S-SEQ, the subsequence common to SEQ1 and SEQ2. INTER is applied to S-SEQ.

And NL is applied to S-SEQ2, the subsequence of SEQ2 that does not have an S-SEQ sample.

Referring to Figure 11, an advantageous decoding method is described for decoding a digital signal according to the principles described above.

In steps S101 and S111, a transformed vector containing samples S-SEQ1 * encoding S-SEQ1, and a prediction vector comprising S-SEQ * samples encoding S-SEQ and S-SEQ2 * samples encoding S-, are respectively received. SEQ2.

In step S 112, an inverse transform is applied to the S-SEQl * samples. For example, it is an H-type window. For example, it is also possible to provide a step S11 comprising additional decoding operations to obtain S-SEQ1.

In step S14, S-SEQ1 is decoded by step S13 and S-SEQ *. At least by predictive decoding, at step S14 S-SEQ is decoded.

Finally, in step S 115, S-SEQ decoded in step S14 and S-SEQ2 * is received and then S-SEQ2 is decoded by predictive decoding. If necessary, it is also possible to use S-SEQ1 decoded in step S113.

An embodiment of step S 114 is described with reference to FIG. In this embodiment, both transform decoding and predictive decoding are involved.

In step S 120, S-SEQ 1 (from S14) and S-SEQ * are received, and then S-SEQ is decoded by predictive decoding. We obtain S-SEQ '.

In step S 121, an inverse transform (for example that already applied to S-SEQ 1 * to obtain S-SEQ 1) is applied to S-SEQ 1 *. We obtain S-SEQ ".

Finally, in step S 122, a linear combination of the samples S-SEQ 'and S-SEQ "is performed to obtain S-SEQ.

With reference to FIG. 13, another embodiment of step S14 is described.

In this embodiment, the decoding of the opposite sign generated by the decoding of S-SEQ * by transform (S-SEQ ") is recreated from S-SEQ * decoded by predictive decoding.

Thus, in this embodiment, at step S 130 S-

SEQ 1 and S-SEQ * and then S-SEQ is decoded. We obtain S-SEQ '.

Then, in step S 13 1, the same folding as S-SEQ "in S-SEQ 'is created, for which purpose the matrix S described above is applied.

S-SEQ "corresponds to the decoding of S-SEQ * by transform in step S 132.

Finally, S-SEQ '' and S-SEQ '' are combined in step S 133 to obtain S-SEQ.

With reference to FIG. 14, a coding entity COD adapted to implement the coding method described above is described. This coding entity comprises a processing unit 140 adapted to receive a digital signal GIS and to determine two sequences of samples: a first sequence comprising an S-SEQ subsequence common to both sequences, and a sub-sequence S-SEQ1, and a second sequence which begins before the end of the first sequence and which contains S-SEQ and an S-SEQ2 subsequence.

The coding entity also comprises a transform coder 141, and a predictive coder 142. These coders are adapted to implement the steps of the coding method described above, and respectively deliver a transformed vector V_T encoding the first sequence and a prediction vector V_P encoding the second sequence.

Communication means (not shown) may be provided for exchanging signals between the encoders.

With reference to FIG. 15, a decoding entity is described for implementing the decoding method described above.

This DECOD decoding entity comprises reception units 150 and 151 for respectively receiving a transformed vector V_T comprising samples S-SEQ 1 * encoding S-

SEQ 1, and a prediction vector V_P comprising samples

S-SEQ * encoding S-SEQ and S-SEQ2 * samples encoding S-SEQ2.

The unit 150 provides S-SEQ1 * to an inverse transform application unit 152. For example, it may further be provided that the unit 152 provides a result to a transform decode unit 153 for performing additional decoding operations. and provide S-SEQ l. Once decoded by the unit 153, the decoding unit 154 receives S-SEQ decoded by the unit 153, and S-SEQ * provided by the unit 151. The unit 154 decodes, at least by predictive decoding S-SEQ, and provided S-SEQ.

Finally, DECOD includes a predictive decoding unit 155 for receiving S-SEQ provided by the unit 154, and S-SEQ2 * provided by the unit 151, then decoding S-SEQ2 by predictive decoding and providing S-SEQ2. If necessary, the unit 153 also provides S-SEQ 1 previously decoded by the unit 153.

A computer program for comprising instructions for implementing the coding method described above could be established according to a general algorithm described in FIG. 9.

This computer program could be executed in a processor of a coding entity as described above, to encode a signal with at least the same advantages as those provided by the coding method.

In the same way, a computer program for comprising instructions for implementing the decoding method described above could be established according to a general algorithm described in FIG.

This computer program could be executed in a processor of a decoding entity as described above, for decoding a signal with at least the same advantages as those provided by the decoding method.

Referring to Figure 16, there is described a hardware device adapted to realize an encoder or a decoder according to an embodiment of the present invention. This device DISP comprises an input E to receive a digital signal SIG. The device also comprises a processor PROC of digital signals adapted to carry out coding / decoding operations in particular on a signal coming from the input E. This processor is connected to one or more memory units MEM adapted to store information necessary for driving. of the device for coding / decoding. For example, these memory units include instructions for implementing the coding / decoding method described above. These memory units may also include calculation parameters or other information. The processor is also adapted to store results in these memory units. Finally, the device comprises an output S connected to the processor to provide an output signal SIG *.

Of course, one can advantageously combine one or more characteristics described above.

Claims

A method of encoding a digital signal, comprising the steps of:

encoding (S93) a first sequence (SEQ. 1) of samples of the digital signal according to a transform coding;

coding (S94) a second sequence (SEQ2) of samples of the digital signal according to a predictive coding;

the method being characterized in that the second sequence

(SEQ2) begins before the end of the first sequence (SEQ 1), a subsequence (S-SEQ) common to the first and second sequences thus being encoded by both predictive coding and transform coding.

2. Method according to claim 1, in which the transform coding of the first sequence comprises the application of an analysis window (H) making it possible to deduce from a perfect reconstruction relation of the digital signal a synthesis window comprising at least three parts:

a first nominal part (NOMI),

a second terminal part (NL) substantially zero,

a third intermediate portion (INTER) continues between the first and second parts,

characterized in that at least portions of the analysis window for respectively deducing said second and third portions of the synthesis window are applied to the common subsequence of the two sequences.

3. Method according to claim 1, characterized in that the transform coding is critical sampling.

4. Method according to claim 2, characterized in that the synthesis window further comprises a fourth soft transition part between an initial value and a value of the nominal part, and in that the third part is an abrupt transition between a value of the nominal part and a value of the substantially zero part.

5. Method according to claim 1, characterized in that the first and second sequences belong to the same frame of the digital signal.

6. A method of decoding a digital signal, comprising the steps:

receiving (S110) a transformed vector encoding a first sequence of samples of the digital signal according to transform coding;

receiving (S101) a prediction vector encoding a second sequence of samples of the digital signal according to a predictive coding;

the method being characterized in that the second sequence begins before the end of the first sequence, a common subsequence to the first and second sequences being thus received coded both by predictive coding and by transform coding; and in that it further comprises the steps of:

a) applying (S 1 12) to the transformed vector an inverse transformation of the transform coding for decoding a subsequence of the first non-coded sequence by predictive coding;

b) decoding (S14) at least in the prediction vector the subsequence common to the first and second sequences at least by a predictive decoding based on at least one sample from step a);

c) decoding (S 1 15) in the predictive vector by a predictive decoding a subsequence of the second non-encoded sequence by transform coding, based on at least one sample from one of the steps a) and b ).

7. Method according to claim 6, characterized in that step b) comprises the sub-steps:

b l) decoding (S 120) in the predictive vector the common subsequence to the first and second sequences by a predictive decoding based on at least one sample from step a);

b2) applying (S 121) to the transformed vector an inverse transformation of the transform coding for decoding the common subsequence to the first and second sequences; and

b3) decoding (S 122) the common subsequence to the first and second sequences by combining at least one sample from step b1) with a corresponding sample from step b2).

8. Method according to claim 6, characterized in that step b) comprises the sub-steps:

b4) decoding (S 130) in the predictive vector the common subsequence to the first and second sequences by a predictive decoding based on at least one sample from step a);

b5) creating (S 131) from at least one sample from step b4) a sample containing a folding equivalent to a transform coding followed by a transform decoding

b6) applying (S 132) to the transformed vector an inverse transform of the transform coding for decoding the common subsequence to the first and second sequences; and

b7) decoding (S 133) the common subsequence to the first and second sequences by combining at least one sample from step b5) with a corresponding sample from step b6).

9. Method according to claim 6, characterized in that step a) comprises the application of a synthesis window comprising at least three parts:

a first nominal part,

and in that at least the second and third parts of the synthesis window are applied to samples coding the subsequence common to both sequences.

Computer program comprising instructions for implementing the method according to claim 1 when the program is executed by a processor.

A computer program comprising instructions for implementing the method according to claim 6 when the program is executed by a processor.

12. Coding entity (COD) of a digital signal (GIS), comprising:

a transform coder (141) for encoding a first sequence of samples of the digital signal according to transform coding;

a predictive coder (142) for coding a second sequence of samples of the digital signal according to a predictive coding;

the coding entity being characterized in that the second sequence begins before the end of the first sequence, a common subsequence (S-SEQ) in the first and second sequences thus being encoded by both predictive coding and transform coding.

13. Decoding entity (DECOD) of a digital signal, comprising receiving means (150, 151):

a transformed vector (V_T) encoding a first sequence of samples of the digital signal according to transform coding; and

a prediction vector (V_P) encoding a second sequence of samples of the digital signal according to a predictive coding;

the decoding entity being characterized in that the second sequence begins before the end of the first sequence, a common subsequence to the first and second sequences thus being encoded by both predictive coding and transform coding; and in that it further comprises:

a first decoder (152, 153) for applying to the transformed vector an inverse transformation of the transform coding for decoding a subsequence of the first non-coded sequence by predictive coding;

a second decoder (154) for decoding at least in the predictive vector the subsequence common to the first and second sequences by at least one predictive decoding based on at least one sample derived from the first transform decoder; and

a third decoder (155) predictive for decoding in the predictive vector by a predictive decoding a subsequence of the second non - coded sequence by transform coding, based on at least one sample from one of the first and second decoders.

14. Decoding entity according to claim 13, characterized in that the second decoder comprises:

first means for decoding in the predictive vector the subsequence common to the first and second sequences by predictive decoding based on at least one sample output by the first transform decoder;

15. Decoding entity according to claim 13, characterized in that the second decoder comprises:

first means for decoding in the predictive vector the subsequence common to the first and second sequences by predictive decoding based on at least one sample output by the first transform decoder; fourth means for creating a folding from at least one sample derived from the first means equivalent to a transform coding followed by a transform decoding;