EP2115741B1 - 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

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).

• les méthodes de codage de forme d'onde, telles que le codage MIC (pour "Modulation par Impulsions Codées") et MICDA (pour "Modulation par Impulsion et Codage Différentiel Adaptatif"), dits aussi "PCM" et "ADPCM" en anglais,
• les méthodes de codage paramétrique par analyse par synthèse comme le codage CELP (pour "Code Excited Linear Prediction"), et
• les méthodes de codage perceptuel en sous-bandes ou par transformée.

Various techniques exist for digitally encoding an audio frequency signal. The most common techniques are:

• waveform coding methods, such as MIC (for "Coded Pulse Modulation") and ADPCM (for "Pulse Modulation and Adaptive Differential Coding") coding, also known as "PCM" and "ADPCM" in English ,
• parametric coding methods using synthetic analysis such as Code Excited Linear Prediction (CELP), and
• perceptual encoding 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).

• F_e est la cadence d'échanlillonnage, et
• F₀ est la fréquence fondamentale.

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
• F ₀ 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 the decoder. Examples of such systems are the ITU-T G.711 64 kbit / s standard coding, the 8 kbit / s ITU-T G.729 coding or the 12.2 kbit / s GSM-EFR coding.

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.

• le codage multi-modes contrôlé par la source et/ou le canal, mis en oeuvre notamment dans les codeurs 3GPP AMR-NB, 3GPP AMR-WB, ou 3GPP2 VMR-WB,
• le codage hiérarchique (ou codage "scalable") qui génère un flux binaire dit « hiérarchique » car il comprend un débit coeur et une ou plusieurs couche(s) d'amélioration (le codage normalisé selon G.722 à 48, 56 et 64 kbit/s étant typiquement scalable en débit, tandis que les codecs UIT-T G.729.1 et MPEG-4 CELP sont scalables à la fois en débit et en largeur de bande),
• le codage à descriptions multiples, décrit notamment dans :
  • " A multiple description speech coder based on AMR-WB for mobile ad hoc networks", H. Dong, A. Gersho, J.D. Gibson, V. Cuperman, ICASSP, p. 277-280, vol.1 (Mai 2004 ).

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) 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 bitrate and bitrate. bandwidth),
• 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, pp. 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-bandwidth codec (fixed), called a "codec heart", guaranteeing the minimum coding quality. 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 .

Specifically described below are scalable scalability and bandwidth encoding techniques, with a CELP heart-coder, 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-coded core encoder producing a bandwidth ranging from narrowband (50-4000 Hz) to wideband (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 schematized on the figure 1 . The broadband input signal s _wb , sampled at 16 kHz, is first decomposed 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 (that is, 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 the figure 1 is the following. 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 the TDAC type for "Time Domain Aliasing Cancellation" in the G.729.1) is carried out at block 110. With reference to figure 1 it is seen in particular that the TDAC encoding is applied to both the error signal on the low band and the signal filtered 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 bitstreams 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.

• le codage CELP en cascade,
• l'extension de bande paramétrique par le module 108, de type TDBWE (pour « Time Domain Bandwidth Extension »), et
• un codage prédictif par transformée TDAC, appliqué après une transformation de type MDCT (pour « Modified Discrete Cosine Transform » ou « transformation en cosinus discrète modifiée »).

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 (Modified Discrete Cosine Transform) transformation.

* Reminders on the G.729.1 decoder

The G.729.1 homologous decoder is illustrated on the figure 2 . The bits describing each frame of 20 ms are demultiplexed in 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 broadband output signal ŝ _wb , sampled at 16 kHz, is obtained via the QMF synthesis filter bank (blocks 209, 210, 211, 212 and 213) incorporating 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 on the figure 3 .

• le spectre MDCT $D_{\mathit{LB}}^w$
  
  du signal de différence, filtré perceptuellement, et
• le spectre MDCT S_HB du signal original de la bande haute.

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 (blocks 301 and 302) to obtain:

• the MDCT spectrum $D_{\mathit{LB}}^w$
  
  of the difference signal, filtered perceptually, and
• the MDCT spectrum S _HB of the original signal to 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: $$\left[\mathrm{<figure-callout\; id="0"\; label="Y"\; filenames="imgf0005.png,imgf0008.png"\; state="\{\{state\}\}">Y</figure-callout>}\left(0\right)\mathrm{<figure-callout\; id="1"\; label="Y"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">Y</figure-callout>}\left(1\right)\cdots Y\left(319\right)\right]=\left[{\mathrm{<figure-callout\; id="0"\; label="D\; LB\; w"\; filenames="imgf0005.png,imgf0008.png"\; state="\{\{state\}\}">D\; </figure-callout>}}_{\mathit{LB}w}^{}\left(0\right){\mathrm{<figure-callout\; id="1"\; label="D\; LB\; w"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">D\; </figure-callout>}}_{\mathit{LB}w}^{}\left(1\right)\cdots D_{\mathit{LB}}^w\left(159\right){\mathrm{<figure-callout\; id="0"\; label="S\; HB"\; filenames="imgf0005.png,imgf0008.png"\; state="\{\{state\}\}">S\; </figure-callout>}}_{\mathit{HB}}\left(0\right){\mathrm{<figure-callout\; id="1"\; label="S\; HB"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">S\; </figure-callout>}}_{\mathit{HB}}\left(1\right)\cdots S_{\mathit{HB}}\left(159\right)\right]$$

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 subband j comprises the coefficients Y ( k ) with sb_bound ( j ) ≤ k <sb_bound ( j + 1). Table 1: TDAC Encoding Boundary Limits and Size J sb_bound ( j ) nb_coef ( j ) 0 0 16 1 16 16 2 32 16 3 48 16 4 64 16 5 80 16 6 96 16 7 112 16 8 128 16 9 144 16 10 160 16 11 176 16 12 192 16 13 208 16 14 224 16 15 240 16 16 256 16 17 272 8 18 280 -

où ε_rms = 2^-24.The spectral envelope {log_ rms (j)} _{j = 0,} ..., 17 is calculated in block 304 using the formula: $$\mathrm{log\_}\mathit{rms}\left(j\right)=\frac{1}{2}{\log }_2\left[\frac{1}{\mathit{nb\_coef}\left(j\right)}{\displaystyle \underset{\mathit{k}\mathit{=}\mathit{sb\_bound}\left(j\right)}{\overset{\mathit{sb\_bound}\left(j+1\right)-1}{\Sigma }}}Y{\left(k\right)}^2+{\epsilon }_{\mathit{rms}}\right],j=0,...,17$$

where ε _rms = 2 ^-24 .

où la notation « round » désigne l'arrondi à l'entier le plus proche, et avec la contrainte : $$-11\le \mathit{r}\mathit{m}\mathit{s}\mathrm{\_}\mathit{i}\mathit{n}\mathit{d}\mathit{e}\mathit{x}\left(\mathit{j}\right)\le +20$$

Cette valeur quantifiée rms _index(j) est transmise au bloc d'allocation de bits 306.The spectral envelope is coded at a variable rate in block 305. This block 305 produces quantized, integer values, denoted rms_index ( j ) (with j = 0, ..., 17), obtained by simple scalar quantization: $$\mathit{r}\mathit{m}\mathit{s}\mathit{\_}\mathit{i}\mathit{not}\mathit{d}\mathit{e}\mathit{x}\left(\mathit{j}\right)=\mathit{r}\mathit{o}\mathit{u}\mathit{not}\mathit{d}\left(2\cdot \mathrm{l}\mathrm{o}\mathrm{boy\; Wut}\mathrm{\_}\mathit{r}\mathit{m}\mathit{s}\left(\mathit{j}\right)\right)$$

where the notation "round" designates the rounding to the nearest integer, and with the constraint: $$-11\le \mathit{r}\mathit{m}\mathit{s}\mathrm{\_}\mathit{i}\mathit{not}\mathit{d}\mathit{e}\mathit{x}\left(\mathit{j}\right)\le +20$$

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

• peuvent être codées par codage dit « de Huffman différentiel »,
• ou peuvent être codées par codage binaire naturel.

Un bit (0 ou 1) est transmis au décodeur pour indiquer le mode de codage qui a été choisi.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 can be chosen according to a given criterion, and, more precisely, the values rms_index ( j ):

• can be coded by "differential Huffman" coding,
• or can be encoded 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 TDAC type transform decoding step in the G.729.1 decoder is illustrated on the figure 4 .

Symmetrically to the encoder ( figure 3 ), the decoded spectral envelope (block 401) makes it possible to find the allocation of the bits (block 402). The 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 the block 305 (multiplexed) and deduce the decoded envelope: $$\mathit{r}\mathit{m}\mathit{s}\mathrm{\_}\mathit{q}\left(\mathit{j}\right)=2^{\mathrm{½}\mathit{r}\mathit{m}\mathit{s}\mathrm{\_}\mathit{i}\mathit{not}\mathit{d}\mathit{e}\mathit{x}\left(\mathit{j}\right)}$$

The spectral content of each of the subbands is found by inverse spherical vector quantization (block 403). The sub-bands not transmitted, 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. figure 2 ).

• avec 160 premiers coefficients correspondant au spectre ${\hat{D}}_{\mathit{LB}}^w$
  
  du signal de différence décodé en bande basse, filtré perceptuellement,
• et 160 coefficients suivants correspondant au spectre Ŝ_HB du signal original décodé en bande haute.

After upgrading this spectrum (block 405) as a function of the spectral envelope and after-treatment (block 406), the MDCT spectrum is separated into two (block 407):

• with 160 first coefficients corresponding to the spectrum ${\hat{D}}_{\mathit{LB}}^w$
  
  the decoded difference signal in a low band, filtered perceptually,
• and 160 subsequent coefficients corresponding to the spectrum Ŝ _HB of the original decoded high band signal.

(bloc 409) résultant de la transformée inverse.These two spectra are transformed into time signals by inverse MDCT transform, denoted IMDCT (blocks 408 and 410), and the inverse perceptual weighting (filter denoted W _LB ( z ) ^-1 ) is applied to the signal. ${\hat{d}}_{\mathit{LB}}^w$

(block 409) resulting from the inverse transform.

The subbands bit allocation (block 306 of FIG. figure 3 or block 402 of the figure 4 ).

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

où nbits_rms est le nombre de bits utilisés par le codage de l'enveloppe spectrale.The purpose of the binary allocation is to distribute between each of the sub-bands a certain bit budget (variable) denoted nbits_VQ, with: $$\mathit{not}\mathit{b}\mathit{i}\mathit{t}\mathit{s}\mathrm{\_}\mathit{V}\mathit{Q}=351-\mathit{not}\mathit{b}\mathit{i}\mathit{t}\mathit{s}\mathrm{\_}\mathit{r}\mathit{m}\mathit{s},$$

where nbits_rms is the number of bits used by the coding of the spectral envelope.

The result of the allocation is the integer bits nbit noted (j) (j = 0, ..., 17), allocated to each sub-band with such global constraint: $${\displaystyle \underset{j=0}{\overset{17}{\Sigma }}}\mathit{nbit}\left(j\right)\approx \mathit{nbits\_}\mathit{VQ}$$

In the G.729.1 standard, the nbit ( j ) values ( j = 0, ..., 17) are further constrained by the fact that nbit ( j ) must be selected from a reduced set of values specified in Table 2. below. Table 2: Possible values of number of bits allocated in TDAC subbands. Size of the sub-band j nb_coef ( j ) Set of allowed values nbit ( j ) (in number of bits) 8 R ₈ = {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}

où offset = -2.The allocation in G.729.1 is based on a "perceptual importance" per subband related to the energy of the subband, denoted ip ( j ) ( j = 0..17), defined as follows: $$\mathit{ip}\left(j\right)=\frac{1}{2}{\log }_2\left(\mathit{rms\_q}{\left(j\right)}^2\times \mathit{nb\_coef}\left(j\right)\right)+\mathit{offset}$$

where offset = -2.

A partir de l'importance perceptuelle de chaque sous-bande, l'allocation nbit(j) est calculée comme suit : $$\mathit{nbit}\left(j\right)=\arg \underset{r\in {\mathrm{R}}_{\mathit{nb\_coef}\left(j\right)}}{\min }\left|\mathit{nb\_coef}\left(j\right)\times \left(\mathit{ip}\left(j\right)-{\lambda }_{\mathit{opt}}\right)-r\right|$$

où λ _opt est un paramètre optimisé par dichotomie.Since the rms values _ q ( j ) = 2 ^{½ rms _ index ( j )} , this formula is simplified as: $$\mathit{ip}\left(j\right)\mathrm{=}\{\begin{array}{ll}\frac{\mathrm{1}}{\mathrm{2}}\mathit{rms\_index}\left(j\right)& \mathrm{for}j=0,...,16\\ \frac{\mathrm{1}}{\mathrm{2}}\left(\mathit{rms\_index}\left(j\right)-1\right)& \mathrm{for}j=17\end{array}$$

From the perceptual importance of each sub-band, the allocation nbit ( j ) is calculated as follows: $$\mathit{nbit}\left(j\right)=\arg \underset{r\in {\mathrm{R}}_{\mathit{nb\_coef}\left(j\right)}}{\min }\left|\mathit{nb\_coef}\left(j\right)\times \left(\mathit{ip}\left(j\right)-{\lambda }_{\mathit{Opt}}\right)-r\right|$$

where λ _opt 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 ≤ γ1 <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.

et S_HB sont accolés (bloc 303 de la figure 3), le filtre W_LB (z) est défini sous la forme : $$W_{\mathit{LB}}\left(z\right)=\mathit{fac}\frac{\hat{A}\left(z/{\gamma }_1\right)}{\hat{A}\left(z/\gamma 2\right)}$$

avec γ₁ = 0,96, γ₂ = 0,6 et $$\mathit{fac}=\left|\frac{{\displaystyle \sum _{i=0}^p}{\left(-{\gamma }_2\right)}^i{\hat{a}}_i}{{\displaystyle \sum _{i=0}^p}{\left(-{\gamma }_1\right)}^i{\hat{a}}_i}\right|$$

However, to ensure spectral continuity when the spectra $D_{\mathit{LB}}^w$

and S _HB are contiguous (block 303 of the figure 3 ), the filter W _LB ( z ) is defined as: $$W_{\mathit{LB}}\left(z\right)=\mathit{uni}\frac{\widehat{\mathrm{AT}}\left(z/{\mathrm{<figure-callout\; id="1"\; label="\gamma "\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">\gamma </figure-callout>}}_1\right)}{\widehat{\mathrm{AT}}\left(z/\gamma 2\right)}$$

with γ ₁ = 0.96, γ ₂ = 0.6 and $$\mathit{uni}=\left|\frac{{\displaystyle \underset{i=0}{\overset{p}{\Sigma }}}{\left(-{\gamma }_2\right)}^i{\widehat{\mathrm{at}}}_i}{{\displaystyle \underset{i=0}{\overset{p}{\Sigma }}}{\left(-{\gamma }_1\right)}^i{\widehat{\mathrm{at}}}_i}\right|$$

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

• le signal différence, entre la bande basse originale et la synthèse CELP, filtré perceptuellement par un filtre du type Â(z/γ1)/Â(z/γ2) compensé en gain (assurant une continuité spectrale), et
• la bande haute qui contient le signal bande haute original.

Le signal bande basse correspond aux fréquences 50 Hz-4 kHz, tandis que le signal bande haute correspond aux fréquences 4-7 kHz.In the G.729.1 standard, the TDAC encoder deals jointly with:

• the difference signal, between the original low band and the CELP synthesis, perceptually filtered by a gain-compensated λ (z / γ1) / λ (z / γ2) filter (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.

• une détermination d'au moins un seuil de masquage fréquentiel à appliquer sur la deuxième sous-bande, et
• une normalisation dudit seuil de masquage pour assurer une continuité spectrale entre lesdites première et deuxième sous-bandes.

Within the meaning of the invention, in order 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 transmit the bits to the sub-bands that 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.

• la première sous-bande comporte alors un signal issu d'un codage de coeur du codeur hiérarchique,
• et la deuxième sous-bande comporte un signal original.

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 has 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"). . We will see indeed, with reference to the figure 12 described below, that the fact of having the original signal is not necessarily a necessity for the implementation of the invention, advantageously.

• une détermination d'au moins un seuil de masquage fréquentiel à appliquer sur la deuxième sous-bande, à partir d'une enveloppe spectrale décodée, et
• une normalisation dudit seuil de masquage pour assurer une continuité spectrale entre lesdites première et deuxième sous-bandes.

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, from a decoded spectral envelope, and
• normalizing 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 the figure 10B .

• la figure 5 illustre une fonction d'étalement avantageuse pour le masquage au sens de l'invention,
• la figure 6 illustre, à titre de comparaison avec la figure 3, la structure d'un encodage TDAC utilisant un calcul de courbe de masquage 606 pour l'allocation de bits selon un premier mode d'application de l'invention,
• la figure 7 illustre, à titre de comparaison avec la figure 4, la structure d'un décodage TDAC homologue de la figure 6, utilisant un calcul de courbe de masquage 702 selon le premier mode d'application de l'invention,
• la figure 8 illustre une normalisation de la courbe de masquage, dans une première forme de réalisation où la fréquence d'échantillonnage est de 16 kHz et le masquage de l'invention appliqué pour la bande haute 4-7 kHz,
• la figure 9A illustre la structure d'un encodage TDAC modifiée, avec pondération directement du signal dans les hautes fréquences 4-7 kHz dans un deuxième mode d'application de l'invention, et codage du seuil de masquage normalisé,
• la figure 9B illustre la structure d'un encodage TDAC dans une variante du deuxième mode d'application illustré sur la figure 9A, avec ici un codage de l'enveloppe spectrale,
• la figure 10A illustre la structure d'un décodage TDAC homologue de la figure 9A, selon le deuxième mode d'application de l'invention,
• la figure 10B illustre la structure d'un décodage TDAC homologue de la figure 9B, selon le deuxième mode d'application de l'invention, avec ici un calcul du seuil de masquage au décodage,
• la figure 11 illustre la normalisation de la courbe de masquage en bande super-élargie dans une deuxième forme de réalisation de l'invention où la fréquence d'échantillonnage est de 32 kHz et le masquage de l'invention appliqué pour la bande haute élargie de 4 à 14 kHz, et
• la figure 12 illustre la puissance spectrale, en sortie de codage CELP, du signal de différence D_LB (en trait plein) et du signal original S_LB (en traits pointillés).

Moreover, other advantages and characteristics of the invention will appear on examining the detailed description, given by way of example below, and the appended drawings in which, in addition to the Figures 1 to 4 previously presented:

• the figure 5 illustrates an advantageous spreading function for masking within the meaning of the invention,
• the figure 6 illustrates, for comparison with the figure 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,
• the figure 7 illustrates, for comparison with the figure 4 , the structure of a TDAC decode homologous to the figure 6 using a masking curve calculation 702 according to the first embodiment of the invention,
• the figure 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,
• the Figure 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,
• the Figure 9B illustrates the structure of a TDAC encoding in a variant of the second mode of application illustrated on the Figure 9A , here with a coding of the spectral envelope,
• the figure 10A illustrates the structure of a TDAC decoding peer of the Figure 9A according to the second embodiment of the invention,
• the figure 10B illustrates the structure of a TDAC decoding peer of the Figure 9B according to the second embodiment of the invention, with here a calculation of the decoding masking threshold,
• the figure 11 illustrates the normalization of the super-wideband masking curve in a second embodiment of the invention where the sampling frequency is 32 kHz and the masking of the invention applied for the broadband widened from 4 to 14 kHz, and
• the figure 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 which is advantageous but not limiting in an encoder / decoder according to the G.729.1 standard described above with reference to FIGS. Figures 1 to 4 , incorporating, according to the invention, a masking information.

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.

• " High-quality audio transform coding at 64 kbps", Y. Mahieux, J.P. Petit, IEEE Transactions on Communications, Volume 42, no. 11, Pages: 3010 - 3019 (Novembre 1994 ).

An example of using masking in an audio codec can be found in Mahieux et al. :

• " High-quality audio transform coding at 64 kbps ", Mahieux Y., JP Petit, 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).

Dans ce document, le calcul du seuil de masquage est effectué par sous-bande plutôt que par raie. Le seuil ainsi obtenu est utilisé pour pondérer perceptuellement chacune des sous-bandes. L'allocation des bits est ainsi effectuée, non pas en minimisant l'erreur quadratique mais en minimisant le rapport « bruit de codage à masque », le but étant de mettre en forme le bruit de codage de tel sorte qu'il soit inaudible (en dessous du seuil de masquage).An example of such a spreading function is presented on the figure 5 . This function is defined in a frequency domain whose unit is the Bark. The frequency scale is representative of the frequency sensitivity of the ear. A usual approximation of the conversion of a frequency f in Hertz, in "frequencies" noted ν (in Barks), is given by the following relation: $$\upsilon =13\cdot \arctan \left(0.00076\cdot f\right)+3.5\cdot \arctan \left({\left(\frac{f}{7500}\right)}^2\right)$$

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. A 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 that 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. Coding and decoding have the same architecture as that presented to figures 1 and 2 . Here, only blocks 110 and 203 are modified as described in figures 6 and 7 .

In a first embodiment described below with reference to the figure 6 , the modified TDAC coder is identical to that of the figure 3 , except that the allocation of the bits following the squared 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).

Similarly, the modified TDAC decoder is presented on the figure 7 in this first embodiment. This decoder is identical to that of the figure 4 , except that the allocation of the bits following the squared 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 the values rms_index ( j ) , j = 0, ..., 17. Similarly, the blocks 607 and 703 perform an identical operation from the values log_mask ( j ) 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.

• où v_k est la fréquence centrale de la sous-bande k en Bark,
• le signe « × » désignant « multiplié par », avec la fonction d'étalement décrite ci-après.

The masking threshold M ( j ) of the subband j is defined by the convolution of the energy envelope σ ² ( j ) = rms_q ( j ) ² × nb_coef ( j ) , by a spreading function 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: $$M\left(j\right)={\displaystyle \underset{k=10}{\overset{17}{\Sigma }}}{\hat{\sigma }}^2\left(k\right)\times B\left({\nu }_j-{\nu }_k\right)$$

• where v _k is the central frequency of the sub-band k in Bark,
• the sign "×" denoting "multiplied by", with the spreading function described below.

• une expression de l'enveloppe spectrale, et
• une fonction d'étalement faisant intervenir une fréquence centrale de la sous-bande j.

In more generic terms, the mask threshold M (j) for a subband 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 .

où $$\begin{array}{cc}{\begin{array}{c}M\end{array}}^{+}\left(j\right)={\hat{\sigma }}^2\left(j-1\right)\cdot {\mathrm{\Delta }}_2\left(j\right)+{\begin{array}{c}M\end{array}}^{+}\left(j-1\right)\cdot {\mathrm{\Delta }}_2\left(j\right)& j=11,\mathrm{..},17\end{array}$$

$$\begin{array}{cc}{\begin{array}{c}M\end{array}}^{-}\left(j\right)={\hat{\sigma }}^2\left(j+1\right)\cdot {\mathrm{\Delta }}_1\left(j\right)+{\begin{array}{c}M\end{array}}^{-}\left(j+1\right)\cdot {\mathrm{\Delta }}_1\left(j\right)& j=10,\mathrm{..},16\end{array}$$

et $${\mathrm{\Delta }}_2\left(j\right)={10}^{\frac{-10}{10}\left({\upsilon }_j-{\upsilon }_{j-1}\right)}$$

$${\mathrm{\Delta }}_1\left(j\right)={10}^{\frac{27}{10}\left({\upsilon }_j-{\upsilon }_{j+1}\right)}$$

An advantageous spreading function is that presented to the figure 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: $$M\left(j\right)=\{\begin{array}{cc}{\begin{array}{c}M\end{array}}^{-}\left(10\right)& j=10\\ {\begin{array}{c}M\end{array}}^{+}\left(j\right)+{\begin{array}{c}M\end{array}}^{-}\left(j\right)+{\hat{\sigma }}^2\left(j\right)& j=1,\mathrm{..},16\\ {\begin{array}{c}M\end{array}}^{+}\left(17\right)& j=17\end{array}$$

or $$\begin{array}{cc}{\begin{array}{c}M\end{array}}^{+}\left(j\right)={\hat{\sigma }}^2\left(j-1\right)\cdot {\mathrm{\Delta }}_2\left(j\right)+{\begin{array}{c}M\end{array}}^{+}\left(j-1\right)\cdot {\mathrm{\Delta }}_2\left(j\right)& j=11,\mathrm{..},17\end{array}$$

$$\begin{array}{cc}{\begin{array}{c}M\end{array}}^{-}\left(j\right)={\hat{\sigma }}^2\left(j+1\right)\cdot {\mathrm{\Delta }}_1\left(j\right)+{\begin{array}{c}M\end{array}}^{-}\left(j+1\right)\cdot {\mathrm{\Delta }}_1\left(j\right)& j=10,\mathrm{..},16\end{array}$$

and $${\mathrm{\Delta }}_2\left(j\right)={10}^{\frac{-10}{10}\left({\upsilon }_j-{\upsilon }_{j-1}\right)}$$

$${\mathrm{\Delta }}_1\left(j\right)={10}^{\frac{27}{10}\left({\upsilon }_j-{\mathrm{<figure-callout\; id="1"\; label="\upsilon \; j\; +"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">\upsilon \; </figure-callout>}}_{j+}\right)}$$

The values of Δ ₁ ( j ) and Δ 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: $$\frac{1}{2}{\log }_2\left(\frac{{\hat{\sigma }}^2\left(j\right)}{M\left(j\right)}\right)$$

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.

où offset = -2 et normfac est un facteur de normalisation calculé suivant la relation : $$\mathit{normfac}={\log }_2\left[{\displaystyle \sum _{j=9}^{17}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_9-{\nu }_j\right)\right]$$

On note que l'importance perceptuelle ip(j), j=0,...,9, est identique à celle définie dans la norme G.729.1. Par contre, la définition du terme ip(j) , j=10,...,17, est changée.The perceptual importance is redefined as follows: $$\mathit{ip}\left(j\right)\mathrm{=}\{\begin{array}{ll}\frac{\mathrm{1}}{\mathrm{2}}{\log }_2\left({\hat{\sigma }}^2\left(j\right)\right)+\mathit{offset}& \mathit{for\; j}=\mathrm{0..9}\\ \frac{\mathrm{1}}{\mathrm{2}}\left[{\log }_2\left(\frac{{\hat{\sigma }}^2\left(j\right)}{M\left(j\right)}\right)+\mathit{normfac}\right]+\mathit{offset}& \mathit{for\; j}=\mathrm{10..17}\end{array}$$

where offset = -2 and normfac is a normalization factor calculated according to the relation: $$\mathit{normfac}={\log }_2\left[{\displaystyle \underset{j=9}{\overset{17}{\Sigma }}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_9-{\nu }_j\right)\right]$$

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.

où log_ mask(j) = log₂(M (j)) - normfac.
On comprendra que la deuxième ligne de l'accolade pour le calcul de l'importance perceptuelle est une expression de la mise en oeuvre de l'invention selon cette première application à l'allocation de bits dans un codage par transformée en tant que couche supérieure d'un codeur hiérarchique.The perceptual importance redefined above is now written: $$\mathit{ip}\left(j\right)\mathrm{=}\{\begin{array}{ll}\frac{\mathrm{1}}{\mathrm{2}}\mathit{rms\_index}\left(j\right)& \mathit{for\; j}=0,...,9\\ \frac{\mathrm{1}}{\mathrm{2}}\left[\mathit{rms\_index}\left(j\right)-\mathrm{log\_}\mathit{mask}\left(j\right)\right]& \mathit{for\; j}=10,...,17\end{array}$$

where log_ mask ( j ) = log ₂ ( M ( j )) - normfac .
It will be understood that the second line of the brace for calculating 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 figure 8 , showing the connection of the high band on which the masking (4-7 kHz) is applied to the low band (0-4 kHz).

où λ _opt est obtenu par dichotomie comme dans la norme G.729.1.Blocks 607 and 703 then perform the bit allocation calculations: $$\mathit{nbit}\left(j\right)=\arg \underset{r\in {\mathrm{R}}_{\mathit{nb\_coef}\left(j\right)}}{\min }\left|\mathit{nb\_coef}\left(j\right)\times \left(\mathit{ip}\left(j\right)-{\lambda }_{\mathit{Opt}}\right)-r\right|$$

where λ _opt 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: $$\mathit{normfac}={\log }_2\left[{\displaystyle \underset{j=10}{\overset{17}{\Sigma }}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_{10}-{\nu }_j\right)\right]$$

Le seuil de masquage est ensuite appliqué uniquement à la bande haute après normalisation du seuil de masquage par sa valeur sur la dernière sous-bande de la bande basse: $$\mathit{normfac}={\log }_2\left[{\displaystyle \sum _{j=0}^{17}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_9-{\nu }_j\right)\right],$$

ou encore par sa valeur sur la première sous-bande de la bande haute : $$\mathit{normfac}={\log }_2\left[{\displaystyle \sum _{j=0}^{17}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_{10}-{\nu }_j\right)\right]$$

In another variant, the masking threshold can be calculated over the entire frequency band, with: $$M\left(j\right)={\displaystyle \underset{k=0}{\overset{17}{\Sigma }}}{\hat{\sigma }}^2\left(k\right)\times B\left({\nu }_j-{\nu }_k\right)$$

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: $$\mathit{normfac}={\log }_2\left[{\displaystyle \underset{j=0}{\overset{17}{\Sigma }}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_9-{\nu }_j\right)\right],$$

or by its value on the first subband of the high band: $$\mathit{normfac}={\log }_2\left[{\displaystyle \underset{j=0}{\overset{17}{\Sigma }}}{\hat{\sigma }}^2\left(j\right)\times B\left({\nu }_{10}-{\nu }_j\right)\right]$$

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 high band (with a different number of eight), as in low band (with a different number of ten).

, et non pas le signal original lui-même. En réalité, comme illustré sur la figure 12, le codage CELP sur le signal différence (courbe en trait plein) donne, en fin de bande basse (après 2700 Hz, typiquement), un niveau d'énergie très proche du signal original lui-même (courbe en traits pointillés). Comme dans le codage G.729.1, seul le signal différence pondéré perceptuellement est disponible en bande basse, on utilise cette observation pour déterminer le facteur de formalisation du masque en bande haute.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 low band. $d_{\mathit{LB}}^W,$

, not the original signal itself. In reality, as illustrated on the figure 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 (curve in dashed 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 formalization 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 on the Figures 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 on the Figures 9B (for encoding) and 10B (for decoding).

In the Figures 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 the Figure 9A and block 906b of the Figure 9B ) from the unquantized spectral envelope.

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

• sf(j) = 1, pour j = 0,...,9 , sur la bande basse,
• et par la racine carrée du seuil de masquage normalisé M(j), pour la bande haute, soit $\mathit{sf}\left(j\right)=\sqrt{M\left(j\right)},$
  
  , pour j = 10, ..., 17.

Indeed, scale factors are given by:

• sf ( j ) = 1, for j = 0, ..., 9, on the low band,
• and by the square root of the normalized masking threshold M ( j ), for the high band, either $\mathit{sf}\left(j\right)=\sqrt{M\left(j\right)},$
  
  For j = 10, ..., 17.

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

With reference to the Figure 9A furthermore, the information corresponding to the scaling factors sf ( j ), for j = 10, ···, 17, may be encoded (block 906) by an envelope coding technique of the same type as that used in FIG. 'G.729.1 encoder (block 305 of the figure 3 ), for example by scalar quantization followed by differential Huffman coding for the high band portion.

• " Low-complexity multi-rate lattice vector quantization with application to wideband TCX speech coding at 32 kbit/s", S. Ragot, B. Bessette, et R. Lefebvre, Proceedings ICASSP - Montreal (Canada), Pages: 501-504, vol.1 (2004 ).

Cette méthode quantification de type gain-forme est mise en oeuvre en particulier dans la norme 3GPP AMR-WB+.The spectrum Y ( k ) is then divided (block 907) by the decoded scale factors, sf _ q ( j ), j = 0, ···, 17 before the so-called "gain-shape" type coding. 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 kbps, 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 the figure 10A . The scale factors sf _ q (j), j = 0, ···, 17, are decoded in block 1001. The block 1002 is then performed as described in Ragot et al. supra.

The extrapolation of the missing subbands (block 1003 of the figure 10A ) follows the same principle as in the G.729.1 decoder (block 404 of the figure 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 the figure 4 . However, the scaling factors sf_q ( j ) , j = 0, ···, 17 are used in place of the decoded spectral envelope, rms_q ( j ) , j = 0, ···, 17.

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

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

At the coding, the masking threshold is calculated and normalized (block 906b of the Figure 9B ) from the coded spectral envelope (block 905b). At decoding, the masking threshold is calculated and standardized (block 1011b of the figure 10B ) from the decoded spectral envelope (block 1001b), the decoding of the envelope making it possible to perform a level adjustment (block 1010b of the figure 10B ) from the quantized values 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 the figure 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 the "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: $$\mathit{ip}\left(j\right)\mathrm{=}\{\begin{array}{ll}\frac{\mathrm{1}}{\mathrm{2}}\mathit{rms\_index}\left(j\right)& \mathit{for\; j}=\mathrm{0..9}\\ \frac{\mathrm{1}}{\mathrm{2}}\left[\mathit{rms\_index}\left(j\right)-\mathrm{log\_}\mathit{mask}\left(j\right)\right]& \mathit{for\; j}=\mathrm{10..17}\end{array}$$

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: $$\mathit{ip}\left(j\right)\mathrm{=}\{\begin{array}{ll}\frac{\mathrm{1}}{\mathrm{2}}\mathit{rms\_index}\left(j\right)& \mathit{for\; j}=0,...,16\\ \frac{\mathrm{1}}{\mathrm{2}}\left(\mathit{rms\_index}\left(j\right)-1\right)& \mathit{for\; j}=17\end{array}$$

It is indicated that in the second embodiment, the module 904 of the Figure 9A can, by calculating the spectral envelope, determine whether the signal is tonal or not and so Block 905 is bypassed if yes. Similarly, for the embodiment described in Figure 9B , the module 904 can make it possible to determine whether the signal is tonal or not and so 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.

The figure 11 generalize 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 on the figure 11 , so also generalizes to the case where the high band has more subbands or covers a wider frequency area than in G.729.1.

With reference to the figure 11 on the low band between 50 Hz and 4 kHz, a first T1 transform 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.

• le signal différence entre le signal original et la synthèse G.711 dans la première bande de fréquences (0-4000 Hz), et
• le signal original, pondéré perceptuellement dans le domaine fréquentiel selon l'invention, dans une deuxième bande de fréquences (4000-8000 Hz).

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 a variant, the hierarchical coder is implemented with a core coder in a first frequency band, and the error signal associated with this core 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. By way of example, the original signal can be sampled at 16 kHz and decomposed into two frequency bands (from 0 to 4000 Hz and from 4000 to 8000 Hz) by an appropriate filter bank, of the QMF type. The encoder can typically be, in such an embodiment, an encoder according to the G.711 standard (with PCM compression). The transform coding is then performed on:

• the difference signal between the original signal and the G.711 synthesis 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.

• le signal différence entre le signal original et le signal de synthèse G.722 dans la première bande de fréquences (0-8000 Hz), et
• le signal original, lequel est encore pondéré perceptuellement selon l'invention dans un domaine fréquentiel restreint à la deuxième bande de fréquences (8000-16000 Hz).

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 difference signal between the original signal and the G.722 synthesis signal 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 figures 6 , 9A and 9B may constitute flowcharts of this first computer program, or further illustrate the structure of such an encoder, according to different 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 figures 7 , 10A , 10B may constitute flowcharts of this second computer program, or further illustrate the structure of such a decoder, according to different embodiments and variants.

Claims (19)

1. Method of encoding an audio signal in a plurality of subbands, wherein at least one first and one second adjacent subband are encoded by transform (601, 602; 901, 902),
   characterized in that, to apply a perceptual weighting, in the transformed domain, at least to the second subband, the method comprises:

   - a determination of at least one frequency masking threshold (606; 905; 906b) to be applied to the second subband, and

   - a normalization of said masking threshold to ensure a spectral continuity between said first and second subbands.
2. Method according to Claim 1, wherein a number of bits to be allocated to each subband is determined from a spectral envelope,
   characterized in that the allocation of the bits (607) for at least the second subband is also determined according to a normalized masking curve calculation, applied to at least the second subband (606).
3. Method according to Claim 2, wherein the encoding is done 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, according to a perceptual importance ip(j) calculated on the basis of a relation of the following type:

   - ip(j)=½rms_index(j), if j is a subband index within the first band, and

   - ip(j)=½[rms_index(j)-log_mask(j)], if j is a subband index within the second band with log_mask(j)=log2(M(j))-normfac in which:

   - rms_index(j) are quantized values deriving from the encoding of the envelope, for the subband j,

   - M(j) is the masking threshold for said subband of index j, and

   - normfac is a determined normalization factor to ensure the 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.
5. Method according to Claim 4, wherein encoding is done over 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 weighting values of $\sqrt{M\left(j\right)}$

   

   are encoded (906), M(j) being 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 encoding by transform takes place in an upper layer (110) of a hierarchical encoder,

   - the first subband including a signal $\left(d_{\mathit{LB}}^W\right)$

   

   deriving from a core encoding (105) of the hierarchical encoder,

   - and the second subband including an original signal (S_HB).
7. Method according to Claim 6, characterized in that the signal $\left(d_{\mathit{LB}}^W\right)$

   

   deriving from the core encoding is perceptually weighted (600; 900).
8. Method according to one of Claims 6 and 7, characterized in that the signal $\left(d_{\mathit{LB}}^W\right)$

   

   deriving from the core encoding 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 encoding by transform is of TDAC type in a global encoder according to the standard G.729.1, and in that the first subband is included in a low frequency band (T1), whereas the second subband is included in a high frequency band.
10. Method according to Claim 9, characterized in that the high frequency band extends to 7000 Hz (T2), at least (T3).
11. Method according to one of the preceding claims, wherein a spectral envelope is calculated (604; 904), characterized in that the masking threshold, for a subband, is defined by a convolution between:

    - an expression of the spectral envelope, and

    - a spreading function involving a central frequency of said subband.
12. Method according to one of the preceding claims, wherein information (305) is obtained 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 the normalization, are conducted only if the signal is non-tonal.
13. Method of decoding an audio signal in a plurality of subbands, wherein at least one first and one second adjacent subband are decoded by transform (709, 711; 1007, 1009),
    characterized in that, to apply a perceptual weighting, in the transformed domain, at least to the second subband, the method comprises:

    - a determination of at least one frequency masking threshold (702; 1001; 1011b) to be applied to the second subband, based on a decoded spectral envelope, and

    - a normalization of said masking threshold to ensure a spectral continuity between said first and second subbands.
14. Method according to Claim 13, wherein a number of bits to be allocated to each subband (703) is determined from a spectral envelope decoding (701),
    characterized in that the allocation of the bits (703) for at least the second subband is furthermore determined according to 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 telecommunication terminal and/or stored on a memory medium intended to cooperate with a reader of said encoder,
    characterized in that it comprises instructions for implementing the encoding method according to one of Claims 1 to 12 when said instructions are executed by a processor of the encoder.
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 telecommunication terminal and/or stored on a memory medium intended to cooperate with a reader of said decoder,
    characterized in that it comprises instructions for implementing 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.

Images