EP2115741A1 - Advanced encoding / decoding of audio digital signals - Google Patents

Abstract

The method involves determining a frequency masking threshold from a masking curve calculation block (606) for applying to a sub band in order to apply a perceptual weighting to the sub band in the transformed field. The masking threshold is normalized for permitting spectral continuity between the two sub bands. The number of bits to be allocated to each sub band is determined from a spectral envelope based on the normalized masking curve calculation applied to the sub-band. Independent claims are also included for the following: (1) a method for decoding a signal (2) a computer program comprising a set of instructions to perform a method for coding a signal (3) a computer program comprising a set of instructions to perform a method for decoding a signal (4) a decoder comprising a memory.

Description

 Advanced coding / decoding of digital audio signals

The present invention relates to a sound data processing.

This processing is adapted in particular to the transmission and / or storage of digital signals such as audio-frequency signals (speech, music, or other).

Various techniques exist for digitally encoding an audio frequency signal. The most common techniques are: - waveform coding methods, such as PCM (for "Coded Pulse Modulation") and ADPCM (for "Pulse Modulation and Adaptive Differential Coding"), also known as "PCM" "and" ADPCM "in English, the methods of parametric coding by synthesis analysis as coding CELP (for" Code Excited Linear Prediction "), and - perceptual coding methods in subbands or by transform.

These techniques treat the input signal sequentially sample by sample (MIC or ADPCM) or sample blocks called "frames" (CELP and transform coding).

It is quickly recalled that a sound signal such as a speech signal can be predicted from its recent past (for example from 8 to 12 samples at 8 kHz) using parameters evaluated on short windows (10 to 20 ms). in this example). These short-term prediction parameters, representative of the transfer function of the vocal tract (for example to pronounce consonants), are obtained by LPC (for Linear Prediction Coding) analysis methods. A longer-term correlation is also used to determine the periodicities of voiced sounds (eg vowels) due to the vibration of the vocal cords. It is therefore a question of determining at least the fundamental frequency of the voiced signal which varies typically from 60 Hz (deep voice) to 600 Hz (high voice) according to the speakers. Then, by LTP analysis (for "Long Term Prediction"), the LTP parameters of a long-term predictor, and in particular the inverse of the fundamental frequency, often called the pitch period. We then define the number of samples in a pitch period by the ratio F _e / F ₀ (or its integer part), where: - F _e is the sampling rate, and - Fo is the fundamental frequency.

We therefore note that the LTP long-term prediction parameters, including the pitch period, represent the fundamental vibration of the speech signal (when it is voiced), while the LPC short-term prediction parameters represent the spectral envelope. of this signal.

In some coders, all of these LPC and LTP parameters, thus resulting from a speech coding, can be transmitted in blocks to a peer decoder, via one or more telecommunication networks, to then restore the initial speech signal.

In conventional speech coding, the encoder generates a fixed rate bit stream. This flow constraint simplifies the implementation and use of the encoder and decoder. Examples of such systems are the ITU-T G.711 64 kbit / s standard encoding, the ITU-T G.729 8 kbit / s standard encoding, or the 12.2 kbit / s GSM-EFR encoding.

In some applications (such as mobile telephony or VoIP for "Internet Protocol"), it is best to generate a variable rate bit stream. Flow values are taken in a predefined set. Such a coding technique, called "multi-rate" is therefore more flexible than a fixed rate coding technique.

Several multi-rate coding techniques can be distinguished: multi-mode coding controlled by the source and / or the channel, implemented in particular in 3GPP AMR-NB, 3GPP AMR-WB or 3GPP2 VMR-WB coders, the hierarchical coding (or "scalable" coding) that generates a so-called "hierarchical" bitstream because it comprises a core rate and one or more layer (s) enhancement (the G.722 standardized coding at 48, 56 and 64 kbit / s is typically bitrate scalable, while the ITU-T G.729.1 and MPEG-4 CELP codecs are scalable in both bit rate and bandwidth), the multi-description coding described in particular in: "A multiple description speech coder based on AMR-WB for mobile ad hoc networks", H. Dong, A. Gersho, JD Gibson, V. Cuperman, ICASSP, p. 277-280, vol. 1 (May 2004).

Hierarchical coding, having the capacity to provide varied bit rates, is described below by distributing the information relating to an audio signal to be coded in hierarchical subsets, so that this information can be used in order of importance. in terms of audio rendering quality. The criterion taken into account for determining the order is a criterion for optimizing (or rather reducing) the quality of the coded audio signal. Hierarchical coding is particularly suited to transmission over heterogeneous networks or having variable available rates over time, or to transmission to terminals with varying capacities.

The basic concept of hierarchical audio coding (or "scalable") can be described as follows.

The bit stream includes a base layer and one or more enhancement layers. The base layer is generated by a low-rate (fixed) codec, known as a "core coded", which guarantees the minimum quality of the coding. This layer must be received by the decoder to maintain an acceptable level of quality. Improvement layers are used to improve quality. However, they may not all be received by the decoder.

The main advantage of hierarchical coding is that it allows an adaptation of the bit rate simply by "truncation of the bit stream". The number of layers (i.e., the number of possible truncations of the bitstream) defines the granularity of the coding. We speaks of coding "strong granularity" if the bitstream comprises few layers (of the order of 2 to 4) and coding "fine granularity" allows for example a step of the order of 1 to 2 kbit / s .

More specifically, the scalable bandwidth and scalability encoding techniques are described below, with a CELP core-type coder, in a telephone band, and one or more broadband enhancement layer (s). An example of such systems is given in the ITU-T G.729.1 8-32 kbit / s fine grain standard. The G.729.1 coding / decoding algorithm is summarized below.

* Reminders on the G.729.1 encoder

The G.729.1 encoder is an extension of the ITU-T G.729 coder. It is a modified G.729 heart-shaped hierarchical encoder producing a bandwidth ranging from narrow band (50-4000 Hz) to wide band (50-7000 Hz) at a rate of 8 to 32 kbit / s for conversational services. This codec is compatible with existing VoIP devices (most of which are equipped according to G.729). Finally, it should be noted that G.729.1 was approved in May 2006.

The G.729.1 coder is shown schematically in FIG. 1. The broadband input signal s _wb , sampled at 16 kHz, is first broken down into two subbands by QMF (for "Quadrature Mirror Filter") filtering. The low band (0-4000 Hz) is obtained by LP low-pass filtering (block 100) and decimation (block 101), and the high band (4000-8000 Hz) by HP high-pass filtering (block 102) and decimation (block 103). The LP and HP filters are of length 64.

The low band is pretreated with a high-pass filter eliminating the components below 50 Hz (block 104), to obtain the signal s _LB , before CELP coding in narrow band (block 105) at 8 and 12 kbit / s. This high-pass filtering takes into account that the Useful band is defined as covering the range 50-7000 Hz. The narrow-band CELP coding is a cascaded CELP coding comprising as a first stage a modified G.729 coding without pre-processing filter and as a second stage an additional fixed CELP dictionary.

The high band is first pretreated (block 106) to compensate for the folding due to the high-pass filter (block 102) combined with the decimation (block 103). The high band is then filtered by a low pass filter (block 107) eliminating the components between 3000 and 4000 Hz from the high band (i.e. the components between 7000 and 8000 Hz in the original signal) to obtain the signal s _HB . A band extension (block 108) is then performed.

An important feature of the G.729.1 encoder according to Figure 1 is as follows. The error signal d _LB of the low band is calculated (block 109) from the output of the CELP coder (block 105) and a predictive coding by transform (for example of type

TDAC for "Time Domain Aliasing Cancellation" in the G.729.1 standard) is carried out at block 110. With reference to FIG. 1, it can be seen in particular that the TDAC encoding is applied to both the error signal on the band. bass and the filtered signal on the high band.

Additional parameters can be transmitted by the block 111 to a homologous decoder, this block 111 performing a so-called "FEC" treatment for "Frame Erasure Concealment", in order to reconstitute possible erased frames.

The different bit streams generated by the coding blocks 105, 108, 110 and 111 are finally multiplexed and structured into a hierarchical bit stream in the multiplexing block 112. The coding is performed by 20 ms sample blocks (or frames). 320 samples per frame. The G.729.1 codec therefore has a three-step coding architecture comprising:

- cascading CELP coding,

the parametric band extension by the module 108, of the TDBWE (for "Time Domain Bandwidth Extension") type, and a TDAC transform predictive coding, applied after an MDCT (for "Modified Discrete Cosine Transform") transformation; "Modified discrete cosine transformation").

* Reminders on the G.729.1 decoder

The homologous decoder according to the G.729.1 standard is illustrated in FIG. 2. The bits describing each 20 ms frame are demultiplexed in the block 200.

The bit stream of the 8 and 12 kbit / s layers is used by the CELP decoder (block 201) to generate the narrow-band synthesis (0-4000 Hz). The portion of the bit stream associated with the 14 kbit / s layer is decoded by the tape extension module (block 202). The portion of the bit stream associated with data rates greater than 14 kbit / s is decoded by the TDAC module (block 203). Pre-echo and post-echo processing is performed by blocks 204 and 207 as well as enrichment (block 205) and aftertreatment of the low band (block 206).

The extended band output signal s _wb , sampled at 16 kHz, is obtained via the QMF synthesis filter bank (blocks 209, 210, 211, 212 and 213) integrating the inverse folding (block 208).

The description of the transform coding layer is detailed below.

* Feedback on the TDAC Transform Encoder in the G.729.1 Encoder

The TDAC type transform coding in the G.729.1 encoder is illustrated in FIG. The filter W _LB (z) (block 300) is a perceptual weighting filter, with gain compensation, applied to the low band error signal d _LB. MDCT transforms are then calculated (block 301 and 302) to obtain: - the MDCT spectrum D ^ ₃ of the difference signal, perceptually filtered, and the MDCT spectrum S _HB of the original signal of the high band.

These MDCT transforms (blocks 301 and 302) apply to 20 ms of sampled signal at 8 kHz (160 coefficients). The spectrum Y (k) from the block 303 of fusion thus comprises 2 x 160, or 320 coefficients. It is defined as follows:

[F (O) 7 (1) -r (319)] = [D 1 (O) D 1 (I) -D 1 (159) S _HB (O) S _HB (1) - - - S _HB (\) 59)]

This spectrum is divided into eighteen sub-bands, a sub-band j being assigned a number of coefficients noted nb_coef (j). The subband splitting is specified in Table 1 below.

Thus, a sub-band j comprises the coefficients Y (k) with sb-bound (j) <k <sb-bound (j + 1).

Table 1: TDAC Encoding Boundary Limits and Size

The spectral envelope JlOg-ZmS (J) J _ _{o π} is calculated in block 304 according to the formula:

1 1 sb boundj j + I) -I log_rms (j) = -log ₂ Σ Y (kf + ε _rj , j = 0, ..., \ 7 nb_coef (j) _{t = a} • b_bound (j) where < ç _ = 2. The spectral envelope is coded at variable rate in block 305. This block 305 produces quantized values, integer, denoted rms _index (j) (with J = 0, ..., 17), obtained by simple scalar quantization: rms _ index (j) = round (2 • log_rms (J)) where the notation "round" designates the rounding to the nearest integer, and with the constraint:

This value rms _index (j) is transmitted to bit allocation block 306.

The coding of the spectral envelope, itself, is carried out again by the block 305, separately for the low band (rms _ index (j), with j = 0, ..., 9) and for the high band ( rms _index (j), with j = 10, ..., 17). In each band, two types of coding may be chosen according to a given criterion, and, more precisely, the rms values _index (j): may be coded by "differential Huffman" coding, - or may be coded by natural binary coding .

A bit (0 or 1) is transmitted to the decoder to indicate the encoding mode that has been chosen.

The number of bits allocated to each sub-band for its quantization is determined in block 306 from the quantized spectral envelope from block 305. The allocation of the bits performed minimizes the squared error while respecting the constraint of a number of integer bits allocated per subband and a maximum number of bits not to be exceeded. The spectral content of the subbands is then encoded by spherical vector quantization (block 307).

The different bit streams generated by the blocks 305 and 307 are then multiplexed and structured into a hierarchical bit stream at the multiplexing block 308. * Reminder on the decoder by transform in the G.729.1 decoder

The step of TDAC-type transform decoding in the G.729.1 decoder is illustrated in FIG. 4. In a symmetrical manner to the encoder (FIG. 3), the decoded spectral envelope (block 401) makes it possible to recover the allocation. bits (block 402). Envelope decoding (block 401) reconstructs the quantized values of the spectral envelope (rms _index (j), for j = 0, ..., 17), from the bit stream generated by block 305 (multiplexed) and deduce the decoded envelope: rms _q (j) = 2 ^{'A m} "- ^mdexϋ)

The spectral content of each of the subbands is found by inverse spherical vector quantization (block 403). The non-transmitted sub-bands, due to a lack of "budget" of bits, are extrapolated (block 404) from the MDCT transform of the signal at the output of the band extension block (block 202 of FIG. 2).

After upgrading this spectrum (block 405) according to the spectral envelope and post-processing (block 406), the spectrum MDCT is separated into two (block 407): with 160 first coefficients corresponding to the spectrum D ^ _B of a decoded difference signal in low band, perceptually filtered, and 160 subsequent coefficients corresponding to the spectrum S _HB of the original decoded signal in high band.

These two spectra are transformed into inverse MDCT transform time signals, denoted IMDCT (blocks 408 and 410), and the inverse perceptual weighting (filter denoted W ₁₈ (Z) ¹ ) is applied to the resulting signal of _B (block 409). of the inverse transform. The sub-bit allocation (block 306 of FIG. 3 or block 402 of FIG. 4) is described more particularly below.

Blocks 306 and 402 perform an identical operation from rms values _index (j), j = 0, ..., 17. It is therefore sufficient to describe only the operation of block 306.

The purpose of the bit allocation is to distribute between each of the sub-bands a certain bit budget (variable) denoted nbits VQ, with: nbits VQ = 351- nbits_rms, where nbits_rms is the number of bits used by the coding of the spectral envelope.

The result of the allocation is the integer number of bits, denoted nbit (j) (with j = 0, ..., 17), allocated to each of the sub-bands with as global constraint:

17 Σnbit (j) ≈ nbits _VQ

In the G.729.1 standard, the nbit (j) values (/=0,...,17) are further constrained by the fact that nbit (j) must be selected from a reduced set of values specified in Table 2 hereinafter.

Size of the Set of Allowed Values nbit (j) (in number of bits) subband j

nb_coef (j)

R _g - {0,7,10,12,13,14,15,16}

16 R ₁₆ - {0,9,14,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30, 31,32

Table 2: Possible values of number of bits allocated in TDAC subbands. The allocation in the G.729.1 standard is based on a "perceptual importance" per subband related to the energy of the subband, denoted ip (j) (/=0..17), defined as follows:

ψ (j) = -Iog ₂ (rms _q (jf x nb_coef (j)) + offset

where offset = -2.

Since the rms values _q (j) = 2 ^{Y2 rms} - ^{mdex (})} , this formula is simplified as:

- rms _index (j) for j = 0, ..., 16 ip (j) =

- {rms _index (j) - \) for j = 11

From the perceptual importance of each sub-band, the allocation nbit (j) is computed as follows: nbit (j) = arg rR m, in nb_ coef (j) x (ip (j) -A) - r

where λ is a parameter optimized by dichotomy.

The incidence of perceptual weighting (filtering of block 300) on the bit allocation (block 306) of the TDAC transform coder is described in more detail below.

In the G.729.1 standard, the TDAC coding uses the perceptual weighting W _LB (z) filter in the low band (block 300), as indicated above. In essence, perceptual weighting filtering allows you to format the coding noise. The principle of this filtering is to exploit the fact that it is possible to inject more noise in the frequency zones where the original signal has a high energy.

The most common perceptual weighting filters used in narrow-band CELP coding are of the form  (z / γ1) /  (z / γ2) where 0 <γ2 <γl <1 and  (z) represents a prediction spectrum linear (LPC). CELP coding synthesis analysis Thus, it amounts to minimizing the quadratic error in a perceptually weighted signal domain by this type of filter.

However, to ensure spectral continuity when the D ^ _B and S _HB spectra are contiguous (block 303 of Figure 3), the W _LB (z) filter is defined as:

A {z l γ2)

with γi = 0.96, γ ₂ = 0.6 and fac =

The fac factor makes it possible to ensure at the junction of the low and high bands (4 kHz) a gain of the filter at 1 to 4 kHz. It is important to note that in the G.729.1 TDAC coding, the coding is based on an energetic criterion only.

* Disadvantages of the prior art

In the G.729.1 standard, the TDAC encoder jointly processes: the difference signal, between the original low band and the CELP synthesis, perceptually filtered by a λ (z / γ1) / λ (z / γ2) compensated filter. gain (ensuring spectral continuity), and - the high band which contains the original high band signal.

The low band signal corresponds to the frequencies 50 Hz-4 kHz, while the high band signal corresponds to the frequencies 4-7 kHz.

The joint coding of these two signals is carried out in the MDCT domain according to the criterion of the quadratic error. Thus, the high band is coded according to energy criteria, which is suboptimal (in the "perceptual" sense of the term). More generally, the case of multi-band coding may be considered, a perceptual weighting filter being applied to the signal of at least one band in the time domain, and the set of subbands being coded together. by transform coding. If we want to apply the perceptual weighting in the frequency domain, then there is the problem of continuity and homogeneity of the spectra between subbands.

The present invention improves the situation.

It proposes for this purpose a method of coding a signal in several sub-bands, in which at least one first and one second sub-band, adjacent, are encoded by transform.

Within the meaning of the invention, to apply a perceptual weighting, in the transformed domain, to at least the second subband, the method comprises: a determination of at least one frequency masking threshold to be applied on the second subband and normalizing said masking threshold to provide spectral continuity between said first and second subbands.

The present invention therefore proposes to calculate a frequency perceptual weighting, using a masking threshold, on only a part of the frequency band (at least on the "second subband" mentioned above) and to ensure spectral continuity with at least another frequency band (at least the aforementioned "first sub-band") by normalizing the masking threshold on the spectrum covering these two frequency bands.

In a first embodiment of the invention, in which a number of bits to be allocated to each sub-band is determined from a spectral envelope, the allocation of the bits for the second sub-band at least is determined furthermore according to a standardized masking curve calculation, applied at least to the second sub-band.

Thus, instead of providing, in this first embodiment, a bit allocation based on purely energetic criteria, the application of the invention makes it possible to assign the bits to the sub-bands which require the most bits according to a perceptual criterion. For the purposes of this first embodiment, perceptual frequency weighting is then applied by masking a part of the audio band, so as to improve the audio quality by optimizing in particular the distribution of bits between subbands according to criteria. perceptual.

In a second embodiment of the invention, the transformed signal in the second subband is weighted by a factor proportional to the square root of the normalized masking threshold for the second subband.

In this second mode, the normalized masking threshold is not used for the allocation of the bits to the subbands as in the first application mode above, but it can advantageously be used to directly weight the signal of the second sub-band at least in the transformed domain.

The present invention is advantageously, but not exclusively, applied to a TDAC-type transform coding in a global encoder according to the G.729.1 standard, the first subband being included in a low frequency band, whereas the second subband is included in a low frequency band, while the second subband is included in a low frequency band. -band is included in a high frequency band that can extend up to 7000 Hz, or even more (typically up to 14 kHz) in band extension. The application of the invention may then consist in providing a perceptual weighting for the high band while ensuring spectral continuity with the low band. It will be recalled that in this type of global coder with a hierarchical structure, transform coding intervenes in a higher layer of a global hierarchical coder. Advantageously: the first subband then comprises a signal resulting from a core coding of the hierarchical coder, and the second subband comprises an original signal.

As in the G.729.1 coder, the signal from the core coding can be perceptually weighted and the implementation of the invention is advantageous in the sense that the entire spectral band can finally be perceptually weighted.

As in the G.729.1 coder, the signal from the core coding may be a signal representative of a difference between an original signal and a synthesis of this original signal (called "difference signal" or "error signal"). . It will be seen, with reference to FIG. 12 described below, that the fact of having the original signal is not necessarily a necessity for the implementation of the invention, advantageously.

The present invention also relates to a decoding method, homologous to the coding method defined above, in which at least one first and one second subband, adjacent, are decoded by transform. To apply a perceptual weighting, in the transformed domain, to at least the second subband, the decoding method then comprises: a determination of at least one frequency masking threshold to be applied on the second subband, starting from a decoded spectral envelope, and a normalization of said masking threshold to provide spectral continuity between said first and second subbands.

A first decoding application mode, homologous to the first application mode of the coding defined above, aims at the allocation of bits to the decoding and a number of bits to be allocated to each subband is determined from a decoding spectral envelope. According to one implementation of the invention, the bit allocation for the at least second subband is further determined according to a normalized masking curve calculation applied at least to the second subband.

A second method of applying decoding within the meaning of the invention consists in weighting the transformed signal in the second subband by the square root of the normalized masking threshold. This embodiment will be described in detail with reference to FIG.

Moreover, other advantages and characteristics of the invention will emerge on examining the detailed description, given by way of example below, and the appended drawings in which, in addition to FIGS. 1 to 4 presented above: FIG. 5 illustrates an advantageous spreading function for masking in the sense of the invention; FIG. 6 illustrates, for comparison with FIG. 3, the structure of a TDAC encoding using a masking curve calculation 606 for the allocation of bits according to a first embodiment of the invention, FIG. 7 illustrates, for comparison with FIG. 4, the structure of a homologous TDAC decoding of FIG. 6, using a curve calculation 702 according to the first embodiment of the invention,

FIG. 8 illustrates a normalization of the masking curve, in a first embodiment where the sampling frequency is 16 kHz and the masking of the invention applied for the high band 4-7 kHz, FIG. 9A illustrates the structure of a modified TDAC encoding, with directly weighting of the signal in the high frequencies 4-7 kHz in a second embodiment of the invention, and coding of the standardized masking threshold, FIG. 9B illustrates the structure of a TDAC encoding in a variant of the second application mode illustrated in FIG. 9A, here with a coding of the spectral envelope; FIG. 10A illustrates the structure of a homologous TDAC decoding of FIG. 9A, according to the second embodiment application of the invention, FIG. 1B illustrates the structure of a homologous TDAC decoding of FIG. 9B, according to the second embodiment of the invention, with here a calculation of the decoding masking threshold, FIG. 11 illustrates the normalization of the curve. in a second embodiment of the invention where the sampling frequency is 32 kHz and the masking of the invention applied for the high band widened from 4 to 14 kHz, and FIG. 12 illustrates the spectral power, at the output of the CELP coding, of the difference signal D _LB (in solid lines) and of the original signal S _LB (in dashed lines).

An application of the invention is described below which proves to be advantageous but not limiting in an encoder / decoder according to the G.729.1 standard described above with reference to FIGS. 1 to 4, incorporating, according to the invention, a piece of information masking.

Previously, however, the notions of gain compensation perceptual filtering and frequency masking are presented below, to better understand the principle of the invention.

The invention provides an improvement to the perceptual weighting performed in the transform coder by exploiting the masking effect known as "simultaneous masking" or "frequency masking".

This property corresponds to the modification of the hearing threshold in the presence of a so-called "masking" sound. This phenomenon is observed typically when, for example, one tries to hold a discussion with ambient noise, for example in the street and that the noise of a vehicle comes to "hide" the voice of a speaker.

An example of using masking in an audio codec can be found in Mahieux et al. : "High-quality audio transform coding at 64 kbps", Y. Mahieux, JP Small, IEEE Transactions on Communications, Volume 42, no. 11, Pages: 3010 - 3019 (November 1994).

In this document, an approximate masking threshold is calculated for each spectrum line. This threshold is the one above which the line concerned is supposed to be audible. The masking threshold is calculated from the convolution of the signal spectrum with a spreading function B (v) modeling the masking effect of a sound (sinusoid or filtered white noise) by another sound (sinusoid or noise filtered white).

An example of such a spreading function is shown in FIG. 5. This function is defined in a frequency domain whose unit is Bark. The frequency scale is representative of the frequency sensitivity of the ear. A usual approximation of the conversion of a frequency / in Hertz, in "frequencies" noted υ (in Barks), is given by the following relation:

υ =

In this document, the calculation of the masking threshold is performed by subband rather than by line. The threshold thus obtained is used to perceptually weight each of the subbands. The allocation of the bits is thus performed, not by minimizing the square error but by minimizing the "mask-to-mask noise" ratio, the aim being to shape the coding noise so that it is inaudible ( below the masking threshold).

Of course, other masking models have been proposed. Typically, the spreading function may be a function of the level of the line and / or the frequency of the masking line. Detection of "peaks" can also be implemented.

It should be noted that to reduce the under-optimality of coding according to G.729.1, one might consider integrating a frequency masking bit allocation, similar to that presented in Mahieux et al. Nevertheless, the heterogeneous nature of the two low-band and high-band signals does not make it possible to directly apply the full-band masking technique of this document. On the one hand, the full-band masking threshold can not really be calculated in the MDCT domain because the low band signal is not homogeneous with an "original" signal. On the other hand, applying a masking threshold over the entire frequency band would again weight the low band signal which is already weighted by the λ (z / γ1) / λ (z / γ2) type filter, the additional weighting by a threshold then having no meaning for this low band signal.

An application of the invention described hereinafter makes it possible to improve the TDAC coding of the encoder according to the G.729.1 standard, in particular by applying a perceptual weighting of the high band (4 to 7 kHz) while ensuring the continuity spectral between low and high bands for a satisfactory and joint coding of these two bands.

In an encoder and / or a decoder according to the G.729.1 standard, improved by the implementation of the invention, only the TDAC coder and decoder are modified, in the example described below.

The input signal is sampled at 16 kHz, bandwidth 50 Hz to 7 kHz. In practice, the encoder always operates at the maximum rate of 32 kbit / s, while the decoder can receive the core (8 kbit / s), as well as one or more enhancement layers (12 to 32 kbit / s per step). 2 kbit / s), as in G.729.1. The coding and decoding have the same architecture as that shown in FIGS. 1 and 2. Here, only blocks 110 and 203 are modified as described in FIGS. 6 and 7.

In a first embodiment described hereinafter with reference to FIG. 6, the modified TDAC coder is identical to that of FIG. 3, except that the allocation of bits following the quadratic error (block 306) is now replaced by a masking curve calculation and a modified bit allocation (blocks 606 and 607), the invention forming part of the calculation of the masking curve (block 606) and its use in the allocation of bits (block 607).

Likewise, the modified TDAC decoder is shown in FIG. 7 in this first embodiment. This decoder is identical to that of FIG. 4, except that the allocation of bits following the quadratic error (block 402) is replaced by a masking curve calculation and a modified bit allocation (blocks 702 and 703). . Symmetrically with the modified TDAC encoder, the invention relates to blocks 702 and 703.

Blocks 606 and 702 perform an identical operation from rms values _index (j), j = 0, ..., 17. Similarly, blocks 607 and 703 perform an identical operation from log_ mask (j) values. and rms _ index (j), j = 0, ..., 17.

In what follows, only the operation of blocks 606 and 607 is described.

Block 606 computes a masking curve from the quantized spectral envelope rms _ q (j) where j = 0, ..., 17 is the number of the subband.

The masking threshold M (J) of the sub-band j is defined by the convolution of the energy envelope σ ² (J) -rms -q (j) ² × nb-coef (j), by a function spreading B (v). In the exemplary embodiment given here of the TDAC coding in the G.729.1 encoder, this masking is performed only on the high band of the signal, with:

17

M {j) = Σ σ ² {k) x B {v _] - v _k ) i = 10 where v _k is the central frequency of the sub-band k in Bark, the sign "x" denoting "multiplied by", with the spreading function described below. In more generic terms, the masking threshold M (J) for a sub-band j is therefore defined by a convolution between:

an expression of the spectral envelope, and

a spreading function involving a central frequency of the sub-band j.

An advantageous spreading function is that shown in FIG. 5. It is a triangular function whose first slope is + 27dB / Bark and -10dB / Bark for the second. This representation of the spreading function allows the iterative calculation of the following masking curve:

M ^" (10) j - 10

M (j) = M ⁺ (j) ₊ M (j) + σ ² (j) j = 1, .., 16

M ⁺ [Il] J = V or

M ⁺ (j) = σ ² (J-I) - A ₂ (J) ₊ M ⁺ (J-I) - A ₂ (J) y = 11,., 17 M (j) = σ ² (j) ₊ 1) - A ₁ (J) ₊ M- (J ₊ I) - A ₁ (J) j = 10, .., 16 and

^{~ 1 (} V

A ₂ (J) = IO seen

The values of A ₁ (J) and A ₂ (J) can be pre-calculated and stored.

A first embodiment of the invention is described below for the allocation of bits in a hierarchical coder such as the G.729.1 encoder.

The bit allocation criterion is based here on the signal-to-mask ratio given by The low band is already filtered perceptually, the application of the masking threshold is limited to the high band. In order to ensure the spectral continuity between the low and high band spectrum weighted by the masking threshold and to avoid biasing the bit allocation, the masking threshold is normalized by its value on the last subband of the low band.

The perceptual importance is redefined as follows:

where offset = -2 and normfac is a normalization factor calculated according to the relation:

normfac = log _; ^ a ² U) XB (V ₉ - V ₁ )

We note that the perceptual importance ip (j), j = 0, ..., 9, is identical to that defined in the G.729.1 standard. On the other hand, the definition of the term ip (j) _> J ⁼ 10, ..., 17, is changed.

The perceptual importance redefined above is now written:

- rms _index (j) for j = 0, ..., 9 ip (j) =

- [rms _ index (j) - log_ mask (j)] for j = 10, ..., 17

where log_ mask (j) = log ₂ [M (J)) - normfac.

It will be understood that the second line of the brace for the calculation of the perceptual importance is an expression of the implementation of the invention according to this first application to the allocation of bits in a transform coding as an upper layer a hierarchical coder. An illustration of the standardization of the masking threshold is given in FIG. 8, showing the connection of the high band on which the masking (4-7 kHz) is applied to the low band (0-4 kHz).

Blocks 607 and 703 then carry out the bit allocation calculations: nbit (j) = arg rR m, in nb_coef (j) x (ip (j) - λ _opt ) - r

where A _{0 1} is obtained by dichotomy as in the G.729.1 standard.

The only difference with respect to blocks 307 and 402 of the prior art is therefore the definition of the perceptual importance ip (j) for the subbands of the high band.

In a variant of this embodiment in which the masking threshold is normalized with respect to its value on the last subband of the low band, the standardization of the masking threshold can be rather carried out from the value of the band. masking threshold in the first subband of the high band, as follows:

normfac = log ₂ Σ σ ² (j) x β (v ₁₀ -v _; )

/ = Io

In another variant, the masking threshold can be calculated over the entire frequency band, with:

17

M (j) = Σσ ² (k) x B (v ₎ -v _k ) i = 0

The masking threshold is then applied only to the high band after normalization of the masking threshold by its value on the last subband of the low band:

normfac = log ₂ - v _} )

J = O or by its value on the first subband of the high band:

normfac = log ₂ B (V ₁₀ - V ₁ )

Of course, these relations giving the normalization factor normfac or the masking threshold M (j) can be generalized to any number of sub-bands (different, in total, from eighteen) in the high band (with a different number of eight), as in low band (with a different number of ten).

In general, we also note that we are looking for a continuity in energy between the high band and the low band, whereas we use the difference signal, weighted perceptually, in the low band of _B , and not not the original signal itself. In fact, as illustrated in FIG. 12, the CELP coding on the difference signal (curve in solid line) gives, at the end of the low band (after 2700 Hz, typically), a level of energy very close to the original signal itself. same (curve in dotted lines). As in the G.729.1 coding, only the perceptually weighted difference signal is available in the low band, this observation is used to determine the normalization factor of the high band mask.

In a second embodiment, the normalized masking threshold is not used to weight the energy in the definition of the perceptual importance, as in the first embodiment described above, but it serves to directly weight the high band signal before TDAC coding.

This second embodiment is illustrated in FIGS. 9A (for encoding) and 10A (for decoding). A variant of this second mode, which is the object of the present invention, in particular for the decoding performed, is illustrated in FIGS. 9B (for encoding) and 10B (for decoding). In Figs. 9A and 9B, the spectrum Y (k) from block 903 is divided into eighteen subbands and the spectral envelope is calculated (block 904) as previously described.

On the other hand, the masking threshold is calculated (block 905 of FIG. 9A and block 906b of FIG. 9B) from the unquantized spectral envelope.

In the embodiment of FIG. 9A, information representative of the weighting is directly coded by the masking threshold M (J), rather than encoding the spectral envelope. In practice, in this embodiment, the scaling factors sf (j) are coded only from j = 10 to j = 17.

Indeed, the scale factors are given by: 5 / ^" (j) - 1, for j = 0, • • •, 9, on the low band, and by the square root of the standardized masking threshold M (J ), for the high band, Sf (J) = JMj]), for J = 10, ..., 17.

Thus, it is not necessary to code the scale factors for j = 0, ..., 9 and the scaling factors are only coded for j = 10, ..., 17.

Referring again to FIG. 9A, the information corresponding to the scaling factors sf (j), for y = 10, • • •, 17, may be encoded (block 906) by a packet envelope coding technique. same type as that used in the G.729.1 encoder (block 305 of FIG. 3), for example by scalar quantization followed by differential Huffman coding for the high band portion.

The spectrum Y (k) is then divided (block 907) by the decoded scale factors, sf _q (j), J = 0, - - -, before the so-called "gain-shape" type coding (of the English "gain-shape coding"). This coding is performed by algebraic quantization according to the quadratic error, as described in the document Ragot et al: "Low-complexity multi-rate lattice vector quantization with application to wideband TCX speech coding at 32 kbit / s", S. Ragot, B. Bessette, and R. Lefebvre, Proceedings ICASSP - Montreal (Canada), Pages: 501-504 , vol. 1 (2004).

This gain-form type quantization method is implemented in particular in the 3GPP AMR-WB + standard.

The peer decoder is shown in Figure 10A. The scaling factors sf _q (j), j - 0, - - -, 17, are decoded in the block 1001. The block 1002 is then carried out as described in the document Ragot et al. supra.

The extrapolation of the missing subbands (block 1003 of FIG. 10A) follows the same principle as in the G.729.1 decoder (block 404 of FIG. 4). Thus, if a decoded subband contains only zeros, then the spectrum decoded by the band extension replaces that subband.

Block 1004 also performs a function similar to that of block 405 of FIG. 4. However, scale factors sf _q (j), j = 0, - - -, 17, are used in place of the envelope spectral decoded, rms _q (j), j - 0, • • •, 17.

This second embodiment may be particularly advantageous especially in an implementation according to the standard 3 GPP-AMR-WB + which is presented as the preferred context of the document Ragot et al. supra.

In a variant of this second embodiment, as shown in FIGS. 9B and 10B (the same references in FIGS. 9A and 9B, and 10A and 10B, designating the same elements), the coded information remains the envelope of FIG. energy (rather than the masking threshold itself as in Figs. 9A and 10A).

At the coding, the masking threshold is calculated and normalized (block 906b of FIG. 9B) from the coded spectral envelope (block 905b). At decoding, the masking threshold is calculated and normalized (block 1011b of FIG. 10B) from the decoded spectral envelope (block 1001b), the decoding of the envelope making it possible to perform a level adjustment (block 1010b of FIG. 10B) from the values quantified rms_q (j).

Thus, in the case of zero decoded subbands, it is advantageously possible in this variant to extrapolate and maintain a decoded signal level that is correct.

In general, in the first as in the second embodiment, it will be understood that a masking threshold is calculated for each sub-band, at least for the sub-bands of the high frequency band, this masking threshold being normalized to ensure spectral continuity between the subbands concerned.

It is also indicated that the calculation of frequency masking in the sense of the invention may or may not be carried out according to the signal to be encoded (in particular if it is tonal or not).

It has indeed been observed that the calculation of the masking threshold is particularly advantageous when the signal to be coded is not tonal, in the first mode, as in the second embodiment, described above.

If the signal is tonal, the application of the spreading function B (v) results in a masking threshold very close to a tone a little more spread out in frequencies. The allocation criterion minimizing the masked coding noise ratio then gives a bit of bit allocation. The same is true for the direct weighting of the high band signal according to the second embodiment. It is therefore preferred, for a tonal signal, to use a bit allocation according to energy criteria. Thus, preferably, the invention is applied only if the signal to be encoded is not tonal.

In generic terms, we thus obtain information (from block 305) according to which the signal to be encoded is tonal or non-tonal, and the perceptual weighting of the high band, with the determination of the masking threshold and normalization, are conducted only if the signal is non-tonal.

The implementation of this observation is now described in an encoder according to the G.729.1 standard. The bit relating to the mode of the coding of the spectral envelope (block 305 of FIG. 3 in particular) indicates a "differential Huffman" mode or a "natural direct binary" mode. This mode bit can be interpreted as a tone detection, since, in general, a tonal signal leads to envelope coding by the "natural direct binary" mode, while most non-tonal signals, having a spectral dynamic more limited, lead to envelope coding by the "Differential Huffman" mode.

Thus, an advantage of "signal tone detection" can be derived to implement the invention or not. More particularly, the invention is applied in the case where the spectral envelope has been coded in "differential Huffman" mode and the perceptual importance is defined in the sense of the invention, as follows:

- rms _ index (j) for j - 0..9 ip (j) =

- [rms _ index (j) - log_ mask (j)] for 7 = 10..17

On the other hand, if the envelope has been coded in "direct natural binary" mode, the perceptual importance remains as defined in the G.729.1 standard:

- rms _ index (j) for j - 0, ..., 16 ip (j) =

- {rms _ index (j) - 1) for j = 17

It is indicated that in the second embodiment, the module 904 of FIG. 9A can, by calculating the spectral envelope, determine whether the signal is tonal or not and thus Block 905 is bypassed if yes. Likewise, for the embodiment described in FIG. 9B, the module 904 can make it possible to determine whether the signal is tonal or not and thus bypass the block 907 in the affirmative.

A possible application of the invention to an extension of the G.729.1 encoder, in particular in super-wideband, is now described.

Figure 11 generalizes the normalization of the masking curve (described in Figure 8) in the case of super-wide band coding. The signals in this embodiment are sampled at a frequency of 32 kHz (instead of 16 kHz) for a useful band of 50 Hz - 14 kHz. The masking curve log ₂ [M (J)] is then defined at least for the sub-bands ranging from 7 to 14 kHz.

Indeed, the spectrum covering the band 50 Hz - 14 kHz is coded by subbands and the allocation of bits to each subband is made from the spectral envelope as in the G.729.1 encoder. In this case, a partial masking threshold can be calculated as previously described.

The standardization of the masking threshold, as illustrated in FIG. 11, is thus generalized to the case where the high band has more subbands or covers a wider frequency range than in the G.729.1 standard.

With reference to FIG. 11, on the low band between 50 Hz and 4 kHz, a first transform T1 is applied to the time weighted difference signal. A second transform T2 is applied to the signal on the first high band between 4 and 7 kHz and a third transform T3 is applied to the signal on the second high band between 7 and 14 kHz.

It will thus be understood that the invention is not limited to signals sampled at 16kHz. Its implementation is particularly advantageous also for signals sampled at higher frequencies, such as for the extension of the G.729.1 encoder to signals sampled not at 16 kHz but at 32 kHz, as described above. If the TDAC coding is generalized to such a frequency band (50 Hz - 14 kHz instead of 50 Hz - 7 kHz currently), the advantage provided by the invention will be really major.

Indeed, in the 4-14 kHz frequency range, the limits of the quadratic error criterion become really unacceptable and, for bit allocation to remain near optimal, a perceptual weighting exploiting the frequency masking in the sense of the invention is very advantageous.

Thus, the invention also aims to improve the TDAC coding, in particular by applying a perceptual weighting of the high-bandwidth (4-14 kHz) while ensuring the spectral continuity between bands, this criterion being important for a joint coding of the band. first low band and the second high and extended band up to 14 kHz.

An embodiment has been described above in which the low band was always weighted perceptually. This embodiment is not necessary for the implementation of the invention. In one variant, the hierarchical coder is implemented with a heart coder in a first frequency band, and the error signal associated with this heart coder is directly transformed, without perceptual weighting in this first frequency band, to be coded. together with the transformed signal of a second frequency band. For example, the original signal can be sampled at 16 kHz and decomposed into two frequency bands (from 0 to

4000 Hz and 4000 to 8000 Hz) by a suitable filter bank, type QMF. The encoder can typically be, in such an embodiment, an encoder according to the standard

G.711 (with PCM compression). The transform coding is then performed on: the signal difference between the original signal and the synthesis G.711 in the first frequency band (0-4000 Hz), and the original signal, perceptually weighted in the frequency domain according to the invention, in a second frequency band (4000-8000 Hz).

Thus, in this embodiment, the perceptual weighting in the low band is not necessary for the application of the invention.

In another variant, the original signal is sampled at 32 kHz and decomposed into two frequency bands (0 to 8000 Hz and 8000 to 16000 Hz) by an appropriate filter bank, QMF type. The encoder can be here an encoder according to the G.722 standard (ADPCM compression in two sub-bands), and the transform coding is performed on: the signal difference between the original signal and the synthesis signal G.122 in the first frequency band (0-8000 Hz), and the original signal, which is still weighted perceptually according to the invention in a frequency domain restricted to the second frequency band (8000-16000 Hz).

Finally, it is pointed out that the present invention also relates to a first computer program, stored in a memory of an encoder of a telecommunication terminal and / or stored on a memory medium intended to cooperate with a reader of said encoder. This first program then comprises instructions for implementing the coding method defined above, when these instructions are executed by an encoder processor.

The present invention also relates to an encoder comprising at least one memory storing this first computer program.

It will be understood that FIGS. 6, 9A and 9B may constitute flowcharts of this first computer program, or further illustrate the structure of such an encoder, according to distinct embodiments and variants. The present invention also relates to a second computer program, stored in a memory of a decoder of a telecommunication terminal and / or stored on a storage medium intended to cooperate with a reader of said decoder. This second program then comprises instructions for implementing the decoding method defined above, when these instructions are executed by a processor of the decoder.

The present invention also relates to a decoder comprising at least one memory storing this second computer program.

It will also be understood that FIGS. 7, 10A, 10B can constitute flowcharts of this second computer program, or further illustrate the structure of such a decoder, according to different embodiments and variants.

Claims

A method of encoding a multi-subband signal, wherein at least one adjacent first and second sub-bands are transform coded (601, 602, 901, 902), characterized in that, for applying a perceptual weighting in the transformed domain to at least the second subband, the method comprises: determining at least one frequency masking threshold (606; 905; 906b) to be applied to the second subband; and - normalizing said masking threshold to ensure spectral continuity between said first and second subbands.

The method of claim 1, wherein a number of bits to be allocated to each subband is determined from a spectral envelope, characterized in that the bit allocation (607) for the second subband is minus is further determined based on a normalized masking curve calculation applied at least to the second subband (606).

The method of claim 2, wherein the coding is performed on more than two subbands, the first subband being included in a first spectral band and the second subband being included in a second spectral band, characterized in that the number of bits per subband nbit (j) is given, for each subband of index j, as a function of a perceptual importance ip (j) calculated from a relation of the type:

- ip (j) = - rms _index (j), if j is a subband index in the first band, and

- ip (j) = - [rms _index (j) -log_mask (j)], if j is a subband index in the

second band, with log_ mask (j) - log ₂ [M (J)) - normfac, where: rms _index (j) are quantized values from the envelope encoding, for subband j,

M (j) is the masking threshold for said subband of index j, and normfac is a normalization factor determined to ensure spectral continuity between said first and second subbands.

4. Method according to claim 1, characterized in that the transformed signal in the second subband is weighted (905) by a factor proportional to the square root of the normalized masking threshold for the second subband.

The method of claim 4 wherein the coding is performed on more than two subbands, the first subband being included in a first spectral band and the second subband being included in a second spectral band, characterized in that what is coded (906) of the weighting values of JM (j), where M (j) is the normalized masking threshold for a subband of index j included in the second spectral band.

6. Method according to one of the preceding claims, characterized in that the transform coding intervenes in an upper layer (110) of a hierarchical coder, the first subband comprising a signal (d ™ _B ) issued from a core coding (105) of the hierarchical coder, and the second subband including an original signal (S _HB ) -

7. Method according to claim 6, characterized in that the signal (dJ _B ) from the core coding is perceptually weighted (600; 900).

8. Method according to one of claims 6 and 7, characterized in that the signal (d ™ _B ) from the core coding is a signal representative of a difference between an original signal and a synthesis of this original signal.

9. Method according to one of claims 6 to 8, characterized in that the transform coding is TDAC type in a global encoder according to the G.729.1 standard, and in that the first subband is included in a band. low frequency (Tl), while the second subband is included in a high frequency band.

10. The method of claim 9, characterized in that the high frequency band extends up to 7000 Hz (T2), at least (T3).

11. Method according to one of the preceding claims, in which a spectral envelope is calculated (604; 904), characterized in that the masking threshold, for a sub-band, is defined by a convolution between: the spectral envelope, and a spreading function involving a central frequency of said subband.

12. Method according to one of the preceding claims, wherein one obtains an information (305) according to which the signal to be encoded is tonal or non-tonal, characterized in that the perceptual weighting of the second subband, with the determination of the masking threshold and normalization, are conducted only if the signal is non-tonal.

A method of decoding a multi-subband signal, wherein at least one adjacent first and second sub-bands are decoded by transform (709, 711; 1007, 1009), characterized in that, for applying a perceptual weighting in the transformed domain to at least the second subband, the method comprises: determining at least one frequency masking threshold (702; 1001;

1011b) to be applied on the second subband, from a decoded spectral envelope, and - a normalization of said masking threshold to ensure spectral continuity between said first and second subbands.

The method of claim 13, wherein a number of bits to be allocated to each sub-band (703) is determined from a spectral envelope decoding (701), characterized in that the bit allocation ( 703) for at least the second subband is further determined based on a normalized masking curve calculation (702) applied at least to the second subband.

15. Method according to claim 13, characterized in that the transformed signal in the second subband is weighted (1004) by a factor proportional to the square root of the normalized masking threshold for the second subband.

16. Computer program, stored in a memory of an encoder of a telecommunications terminal and / or stored on a storage medium intended to cooperate with a reader of said encoder, characterized in that it comprises instructions for the implementation encoding method according to one of claims 1 to 12 when said instructions are executed by an encoder processor.

17. Encoder, characterized in that it comprises at least one memory storing a computer program according to claim 16.

18. Computer program, stored in a memory of a decoder of a telecommunications terminal and / or stored on a storage medium intended to cooperate with a reader of said decoder, characterized in that it comprises instructions for the implementation of the decoding method according to one of claims 13 to 15 when said instructions are executed by a processor of the decoder.

19. Decoder, characterized in that it comprises at least one memory storing a computer program according to claim 18.