FR2980620A1 - Method for processing decoded audio frequency signal, e.g. coded voice signal including music, involves performing spectral attenuation of residue, and combining residue and attenuated signal from spectrum of tonal components - Google Patents

Abstract

The method involves identification of tonal components (E303a) within an audio signal represented in a frequency domain. A spectrum of tonal components identified from a sinusoidal model is calculated (E303b). The residue of difference between a decoded audio signal and a signal from the spectrum of tonal components is obtained (E304). Spectral attenuation of the residue is performed (E305), and the residue and the attenuated signal are combined (E306) from the spectrum of tonal components. Independent claims are also included for the following: (1) a decoder (2) a computer program comprising instructions to perform a method for processing a decoded audio frequency signal.

Description

The present invention relates to the field of the processing of audio-frequency signals and in particular of speech or music signals which have been coded and decoded by speech coders and decoders. The audio processing device according to the invention is particularly suitable for transmitting and / or storing audio-frequency signals. The invention relates more particularly to the post-processing of the decoded signals to improve the quality of the decoded musical signals.

Different techniques exist for converting into digital form and compressing an audio frequency signal (speech, music, etc.). The most common techniques in telecommunication services are waveform coding methods, such as MIC (for "Coded Pulse Modulation") and ADPCM (for "Pulse Modulation and Adaptive Differential Coding") coding. also "PCM" or "ADPCM" in English, the methods of parametric coding by synthesis analysis such as coding CELP (for "Code Excited Linear Prediction" in English), and perceptual coding methods in subbands or by transform. These techniques process the input signal sequentially sample by sample (MIC or ADPCM) or so-called frame sample blocks (CELP, transform coding). Of particular interest here is the case of CELP coding as an example of a speech codec, however, the invention applies to the general case of speech coders (eg: MIC, ADPCM, CELP, etc.). The CELP coding - of which its variant called ACELP (for "Algebraic CELP") used for example in the 3GPP AMR and AMR-WB standards - is a predictive coding based on the source-filter model of speech production. The filter generally corresponds to an all-pole transfer function filter 1 / A (z) obtained by linear prediction LPC (for "Linear Predictive Coding"). The signal is synthesized using the quantized version, 30 1 / Å (z), of the filter 1 / A (z). The source - that is, the excitation of the predictive linear filter 1 / A (z) - is generally represented as the combination of adaptive excitation (obtained by long-term prediction modeling the vibration of the vocal cords). , and a fixed excitation (or innovation) coded efficiently in the form of pulse dictionaries (ACELP), noise dictionaries, etc. The search for "optimal" excitation is performed by minimizing a quadratic error criterion in the domain of the signal weighted by a transfer function filter I4 (4 derived from the linear prediction filter A (z), from the form IMA-A (z / MIA (z / y 2).

It may be noted that in the 3GPP AMR-WB codec, which is described in the article by B. Bessette et al., Entitled "The Adaptive Multilate Wideband Speech Coder (AMR-WB)", IEEE Transactions on Speech, Audio and Language Processing, Nov. 2002, the ACELP encoding is actually applied in frames of 20 ms not to the direct signal sampled at 16 kHz but to a pre-emphasized signal sampled at 12.8 kHz in a reduced audio band (0-6400 Hz ); pre-emphasis (or pre-emphasis) is performed by the transfer function filter 1-az-1 with a = 0.68. The perceptual weighting filter used for coding noise shaping is of the form W (z) = A (z 1 y) / (1-az-1) with 7 = 0.92. CELP coding is based on a temporal and predictive coding approach based on a signal model (LPC); this type of coding is very effective on (clean) speech, but it gives at low speed a quality which is often mediocre for cases of signals moving away from the assumptions of the speech production model. Thus, for musical signals having a structure composed of tonal components that are badly encoded by a CELP model (eg harmonic, multipitch, or inharmonic signals, such as organ, piano, etc.), it is well known that a transform coding is much more suitable. Nevertheless, CELP speech coders are historically deployed in fixed and mobile telephony applications because they provide better low-rate quality for the speech signals that are the most important signals for telephony applications. However, music signals or mixed content (mix of speech and music) still represent an important signal class in some use cases such as waiting music, return tone, and so on. It is therefore relevant and important to seek to improve the quality of speech coders already deployed in services for music and mixed content. Figure 1 shows the concrete example of a musical signal (an organ sound) sampled at 16 kHz, prefiltered by a P.341 mask (50-7000 Hz) and analyzed by short-term FFT on a support of 512 samples (32 ms). The spectrum of the signal ('sig') on the considered frame shows a number of tonal components (ct). After encoding the signal by the 12.65 kbit / s AMR-WB encoder, the corresponding noise spectrum ('err') appears as relatively strong between the tonal components. This characteristic of the coding noise is explained by the fact that the CELP coding is temporal and forms the noise in principle according to the frequency response of the filter W (z) 1, unlike a transform coding operating in a frequency domain , CELP coding can not "dig" between harmonics (or tonal components).

An example of a technique improving CELP decoding for musical signals is presented in the article by T. Vaillancourt et al. entitled "Inter-tone noise reduction in a low bit rate CELP decoder", Proc. ICASSP 2009. This technique is used in the ITU-T G.718 "narrow band" mode. The principle of this technique consists in carrying out a post-processing in the frequency domain (by short-term FFT) of the decoded CELP signal and reducing the CELP coding noise between the partials (tonal components) by a spectral attenuation function and a gain (or energy) adjustment. This post-processing is similar to a conventional noise reduction by short-term spectral attenuation. It comprises the following steps illustrated in FIG. 2: a pre-emphasis (block 201) and a short-term discrete Fourier transform (FFT) (block 202) are applied in 20 ms frames to the decoded signal sampled at 16 kHz - FFT has time support of 30 ms - A decoded CELP signal classification (block 203) to evaluate the stationarity level in the current frame and adjust the post-processing frequency area and the reduction level maximum noise. An attenuation of the coding noise (block 204): at each time index frame t, the level of coding noise AT, (; (i) is estimated by critical subbands of index i (CB for "Critical Bands "in English) A signal to noise ratio SNR (f) is then estimated by frequency line, which is defined as the ratio of the energy of the decoded signal to the frequency line f and the noise energy. AT, (;,), (i) in the critical band including the frequency line f A spectral attenuation function g, (f) is finally calculated line by line, as a function of the noise signal, and the gain gs ( f) is then smoothed - the attenuation correction (block 205) by a correction gain gc, (f) to compensate for the energy loss due to the spectral attenuation of the block 204 as well as the attenuation of the highs frequencies by the CELP model - time synthesis (block 206) by addition-overlap (block 207) and deceleration (block 208). aillancourt et al is of reasonable complexity and causes an additional delay (10 ms) compatible with conversational applications. However, it has several drawbacks: - Cutting into "critical bands" divides the spectrum into arbitrary adjacent regions; subband coding noise estimation is simplistic and imprecise. However, the quality of noise reduction processing depends greatly on the reliability of the estimation of the noise level and therefore the signal-to-noise ratio. This technique can produce artifacts such as musical noise. - The noise reduction between the tonal components (block 204) is performed by applying an attenuation factor (gain) to each line of the amplitude spectrum mixing signal and coding noise. The spectral attenuation can thus affect both the signal, and more particularly its harmonic structure (including the lobes associated with each tonal component), and the coding noise, while it would be preferable not to disturb the signal and the signal. only attenuate the coding noise.

There is therefore a need, for the post-processing of the musical signals, to preserve the signal and more particularly its short-term harmonic structure while effectively reducing the coding noise. This technique must be applied in the case where it is not necessary to transmit (from the encoder to the decoder) additional information for the processing - here we consider the case of a blind post-processing at the decoder. The present invention improves the situation. To this end, it proposes a method for processing a decoded audio-frequency signal such that it comprises the following steps: identification of tonal components in the audio-frequency signal represented in the frequency domain; calculating a spectrum of tonal components identified from a sinusoidal model; obtaining a residue by the difference between the decoded audio-frequency signal and a signal coming from the spectrum of the tonal components; spectral attenuation of the residue; and combining the attenuated residue and the signal from the spectrum of the tonal components.

Thus, the sinusoidal model used to obtain a spectrum of tonal components is adapted to represent the short-term harmonic structure (or tonal structure) of musical signals. Performing the spectral attenuation on the residue obtained by the difference between the decoded signal and that coming from the spectrum of the tonal components makes it possible not to affect the harmonic structure (or tonal structure) of the signal during the processing.

The various particular embodiments mentioned below can be added independently or in combination with each other, to the steps of the allocation method defined above. In a first embodiment, the steps of obtaining the residue, attenuation and combination are performed in the frequency domain.

Thus, the residue is obtained by the difference between the spectrum of the decoded signal and the spectrum of the tonal components. In a second embodiment the steps of obtaining the residue, attenuation and combination are performed in the time domain.

The residue is here obtained by the difference between the decoded signal and the time signal from the spectrum of tonal components. The time-frequency inverse transform is then performed here after obtaining the spectrum of the tonal components. According to a particular embodiment, the tonal components are identified by an interpolation method on a predetermined number of frequency lines. Thus, the estimation of certain parameters (frequency, amplitude) of the tonal components is performed by an interpolation method from a predetermined number of frequency lines.

Interpolation makes it possible to identify the tonal components with a precision exceeding the inter-line resolution of the time-frequency transform. In a particular embodiment, the identification of the tonal components, more particularly the estimation of the parameters (amplitude, phase), uses a spectral pattern pre-calculated and stored in memory.

This spectral pattern makes it possible to perform a closed-loop sinusoidal analysis, that is to say, to take into account the pattern synthesis analysis. This method is advantageous for ensuring a total correspondence between analysis and synthesis in the sinusoidal model. Advantageously, the calculation of the spectrum of tonal components uses a spectral pattern pre-calculated and stored in memory.

The use of a spectral pattern makes it possible here to obtain the sinusoidal spectrum associated with a frequency peak with a precision exceeding the inter-line resolution of the time-frequency transform. In one possible embodiment, the spectral attenuation is performed according to a data of classification of the audiofrequency signal.

Since post-processing is intended to improve the quality of musical signals, it is essential to apply it on signals classified as music. Thus, in a possible embodiment, when the signal is not detected as a musical signal, the attenuation gain is different. It can for example be equal to 1 for a signal that is not detected as musical, thus canceling the attenuation, in order to preserve the signals such as speech or transients. In the case of an attenuation in the time domain, the spectral attenuation is carried out by filtering the residue in the time domain. The present invention also provides a device for processing a decoded audio-frequency signal, such as it comprises: a module for identifying tonal components in the audio frequency signal represented in the frequency domain; a module for calculating a spectrum of tonal components identified from a sinusoidal model; a module for obtaining a residual by the difference between the decoded audio-frequency signal and a signal coming from the spectrum of the tonal components; a spectral attenuation module of the residue; and a combination module of the attenuated residue and the signal coming from the spectrum of the tonal components. This device has the same advantages as the method described above, which it implements. The invention also relates to an audio frequency signal decoder comprising a processing device as described above.

It relates to a computer program comprising code instructions for implementing the steps of the processing method as described, when these instructions are executed by a processor. Finally, the invention relates to a storage medium, readable by a processor, integrated or not to the processing device, optionally removable, storing a computer program implementing a method of treatment as described above. Other features and advantages of the invention will emerge more clearly on reading the following description, given solely by way of nonlimiting example, and with reference to the appended drawings, in which: FIG. 1 previously described illustrates a example of a short-term spectrum of a musical signal and the associated CELP coding noise; FIG. 2 represents a flowchart of the noise reduction method of the state of the art, as described above; FIG. 3 illustrates in the form of a block diagram, both the processing device and the processing method that it implements according to a first embodiment of the invention; FIGS. 4a and 4b show examples of short-term frequency analysis of a sinusoidal signal used in one embodiment of the invention; FIG. 5 illustrates the principle of parabolic interpolation used in a particular embodiment of the invention; FIGS. 6a, 6b and 6c illustrate examples of time windows used for the implementation of the time-frequency transform or inverse transform steps, in one embodiment of the invention; FIGS. 7a and 7b illustrate an exemplary spectrum of tonal components in an embodiment of the invention and a frequency pattern calculated and stored in one embodiment of the invention; FIG. 8 illustrates in flowchart form the steps implemented for calculating a spectrum of tonal components according to one embodiment of the invention; FIG. 9 illustrates in flowchart form the detail of a step of calculating the spectrum of tonal components according to one embodiment of the invention; FIGS. 10a and 10b illustrate examples of attenuation factors used in the treatment method according to the invention; FIG. 11 illustrates the steps of windowing and synchronizing the input signal and the output signal of the processing device, in one embodiment of the invention; FIG. 12 illustrates in the form of a block diagram, both the processing device and the processing method that it implements according to a second embodiment of the invention; and FIG. 13 illustrates an example of a hardware representation of a processing device according to the invention. FIG. 3 according to a first embodiment of the invention represents a processing device 320 for improving the quality of a decoded audio-frequency signal. This processing device may be integrated into a decoder, then comprising the decoding module 301, or may be independent of the decoder and thus receive as input a decoded audio signal (x (n)) - the advantage of integrating the module of decoding 320 processing to the decoder is that it can benefit from internal parameters to the decoder, such as decoded parameters. Figure 3 also serves to illustrate the main steps implemented by the processing method according to this first embodiment of the invention. An audio frequency signal, for example, coming from a 3GPP AMR-WB type standard speech encoder is therefore first decoded by a decoding module 301. In this particular embodiment, without loss of generality, it is considered here the audiofrequency signals are sampled at F = 16 kHz and divided into frames of L = 320 samples (20 ms). In an alternative embodiment, the processing can also be applied to the decoded signal filtered by a perceptual filter I / 1 / (z) or by a linear prediction filter. In this case, one skilled in the art will understand that the inverse filter will have to be used after post-treatment 320.

In the exemplary embodiment illustrated, this decoding block 301 corresponds to the AMR-WB decoder as described in the article by B. Bessette et al., Entitled "The Adaptive Mutlirate Wideband Speech Coder (AMR-WB)", IEEE Transactions on Speech, Audio and Language Processing, Nov. 2002.

It performs ACELP decoding at the sampling rate of 12.8 kHz, a band extension to generate the frequencies of 6.4-7 kHz and a combination of the low band (0-6400 Hz) and the high band (6400-7000 Hz). The block 301 operates in a 20 ms frame (L = 320 samples at 16 kHz) and synthesizes a signal denoted x (n), the samples of which are conventionally numbered by the indices n = 0, ..., L -1, with L = 320, in the current frame - according to this convention the previous decoded frame thus corresponds to the indices n = -N, ..., - 1. It will be recalled here that the decoder AMR-WB processes a bit stream every 20 ms, and that this bit stream contains the binary representation of the following decoded parameters: o LPC coefficients coded in the form of ISF (for "Immittance Spectral Frequencies") ) for an LPC order of 16 - for each 20 ms frame o Pitch delay - for each 5 ms sub-frame o Binary flag (indicating "pitch flag") indicating whether the pitch contribution is filtered low-pass or not - for each 5 ms subframe o Fixed dictionary indices for each subframe - for each 5 ms subframe o Adaptive and fixed dictionary dies - for each 5 ms subframe We do not list here the decoded parameters only at the rate of 23.85 kbit / s on the high band (6400-7000 Hz) because the post-processing device is rather dedicated to the improvement of the signal decoded by the CELP model in the low band ( 0-6400 Hz). In an alternative embodiment, the post-processing 320 is applied to the low band decoded signal by the ACELP decoder of AMR-WB. In this case, the post-processing 320 operates with frames of 20 ms (256 samples) at the sampling frequency of 12.8 kHz. The high band (6.4-7 kHz) of the AMR-WB decoder is then time shifted by a predetermined number of samples, in order to align the synthesized high band (6400-7000 Hz) with the low band (0- 6400 Hz) decoded and post-processed. This variant of the embodiment, where the decoded signal is sampled at 12.8 kHz sampling, could also include the case where the decoded signal is pre-filtered before applying the post-processing 320, the pre-filter being realized for example by a perceptual weighting filter of the form F (z) = λ (z / y) / (1-az 1)), where λ (z) is the LPC filter decoded in each CELP subframe. This pre-filtering would make it possible to apply the post-processing 320 in the domain of the "weighted signal" corresponding to the domain of the CELP coding synthesis analysis. This variant involves applying a post-filtering by 1 / F (z) on the result x post (n) of the post-processing.

Block 302 performs a short-term frequency analysis (E302) of the signal x (n), n = - (N-L), ..., L -1, including the samples of the current decoded frame (n = 0, ..., L-1) and a part of the previous decoded frame (n = - (N - L), ..., -1).

The frequency analysis of the module 302 is implemented, in the embodiment described, by windowing and fast Fourier transform (FFT): o An analysis window is applied to the signal to obtain the windowed signal xw (m) = wa (m) .x (m- (N-L)), m = 0, ..., N -1 where the analysis window wa (m) is defined as a Hamming window of length N = 512: wa (m) = 0.54 - 0.46 cos 22r m, m = 0, ..., N -1 N 1 This window is illustrated in Figure 6a. Of course, other analysis windows may be provided, for example asymmetrical windows that would better condition the synthesis window subsequently used in 307, but have spectral properties less interesting analysis. For reasons of computational efficiency justified further, it is recommended in practice to integrate a "phase recentering" into the time / frequency transform (and then its inverse, for the synthesis). This consists in making a circular permutation of the windowed signal of N / 2 samples to the left (for analysis), ie the half-length of the window, before performing the Fourier Transform. Alternatively, this amounts to applying a correction of the phase spectrum resulting from the Fourier Transform by adding a linear phase irk / V / K. By one or the other means, this amounts to changing the definition of the transform into: K-1 2z1c NX (k) = K 2 k = 0, ..., K -1 m = 0 In a variant of embodiment, the windowed signal xi, v (m), m = 0, ..., N -1, is transformed by FFT of length K = N = 512 to obtain the discrete spectrum X (k): K-1 2z1c X (k) = xi, (m) .e K, k = 0, ..., K -1 m = 0 In this case, one skilled in the art will readily understand how to adapt the main embodiment to take into account the phase shift involved in the circular permutation of N / 2 samples.

In the two embodiments described for the time / frequency transform (with or without phase recentering), the FFT is calculated according to the method described in the article by H.V. Sorensen et al., Entitled "Real-valued fast Fourier transform algorithm," IEEE Trans. on Signal Processing, Vol.35, No.6, pp 849-863, 1987.

In a further variant of the embodiment, other time-frequency transforms may be used, for example the "Modulated Complex Lapped Transform" (MCLT) combining modified cosine transforms and modified discrete sine quadrature. For the sinusoidal model to apply it is important to use a complex transform to estimate the amplitude and phase of the tonal components. Time-frequency analysis techniques such as parsimonious time-frequency representations (sparse in English) also apply to implement the sinusoidal model. In the preferred embodiment, the length of FFT is chosen to be identical to that of the analysis window, which is perfectly justified in the case where K = N = 512 is a power of 2. However, the choice of K = N is not restrictive within the meaning of the invention, and the invention also applies to the case where the length K of FFT is made greater than that N of the windowed signal by means of a zero padding and possibly time shift of the signal - this alternative is in general interesting in the case where the length N of the windowed signal is not a power of 2. The signal xi, (m) being real, the spectrum X (k) is with Hermitian symmetry and only the positive frequencies k = 0,, KI 2 are necessary. The spectrum obtained X (k), k = 0, ..., K 2 is set in polar coordinates by separating amplitude X (k) and phase / X (k). As will be seen below, in a certain embodiment, it is not necessary to calculate this phase term, saving a certain amount of operation. We now describe a standard method of frequency analysis of signals composed of pure sinusoids.

The discrete sinusoid of amplitude A0, of normalized frequency coo and of phase 00, v (n) = Aocos (coon + 00), has for Fourier transform (in discrete time): V (co) = - (5 (co + coo For a short-term frequency analysis, the signals are in general weighted by a factor of 0. +/- 9 ° + .5 (co-coo) e-8), coe [-7-1-, 71-] 2. window w (n) of integer length N> 0 such that w (n) 0 for n = 0, ..., N -1 and w (n) = 0 otherwise. In this case the windowed sinusoid vw (n) = w (n) v (n) a for Fourier transform: V. (co) = 2 (W (co + w0) e8 ° + W (co-co) -18 °) where it is important to note that the spectrum W (co) is periodic with a period of 271-.

Thus the windowed sinusoid has the spectrum of the result of the convolution of the periodized transform of the window W (co) by two Diracs centered on - ± coo. This result easily extends to the case where the signal is a linear combination of sinusoids. By linearity, if 1-1 v (n) = cos (Wed ± ei), i = o / -1 A 17 '(co) E (1 / Vo + + W (o - ci) e-., = 0 2 When the signal is analyzed by discrete Fourier transform of length KN, the transform becomes VW co =, k = 0, ..., K -1, to lighten the notations, the discrete Fourier transform K of vi, (n ) is denoted by y, (k) instead of VK, so that: / -1 A V '(k) = E (W (k + ki) efe' + W (k - ki) e - fe '), k = 0, K -1 i = 0 2 with k = -Kw 2, r The spectrum W ,, (k) (also called 14 ", here) of length KN of the window n (m) is defined by the equation: K-1 2zIc WK (k) = wzp (m) e K, k = 0, K -1, m = 0 the window wzp (m) being obtained by a zero padding in English) around the window n (m) [w (m) m = 0, N -1 wzP (m) 0 me [0, N [20 Advantageously and coherently with the windowed signal transform xM, (m) , it is recommended to use the modified version of the time-frequency transform integrating the phase recenter described above, and that i is mathematically written: K-1 -i-2z1c (rn-1) K (k) = wzp (m) e. K 2 k = 0, ..., K -1 m = 0 The advantage is that having to deal with a symmetrical window, the "phase recentering" assures it a real spectrum (that is to say that the imaginary part of the spectrum is null, or that its phase is zero to n near), and that whatever the frequency oversampling chosen (which depends on the stuffing of zeros). then An example of short-term frequency analysis is shown in FIGS. 4a and 4b for the following conditions: o 1 = 2 sinusoids with A0 = 1 and Al = 0.5 con = 27r-4 co = 2z12.5 - 32 32 00 = 0 91 = 0 o Hamming window (normalized) of length N = 32: (n) w (l) Nwh am n = 0, ..., N -1 Ewham (n) n = 0 with n wh (n) = 0.54 - 0.46cos 27r, n = N -1 N -1} It can be noted that the normalization by the average of the window is only an example here, the window could not be normalized or a Another normalization factor could be used, which only translates into dB the spectral values discussed here. o FFT of length K = N = 32 (FIG. 4a) or K = 8/1 = 256 (FIG. 4b). The amplitude spectra (in dB) obtained by discrete Fourier transforms of respective lengths K = 32 and K = 256 are therefore represented in FIGS. 4a and 4b; only the part corresponding to the positive frequencies is shown. There are two peaks (of form -117, (k + ki)) around k0 and in FIG. 24a the second sinusoid has a fractional k1 position, whereas in FIG. 4b the two sinusoids are localized on whole lines ( "tins" in English); it is checked at the same time that the transform of length K = 256 is an interpolated version of that of length K = 32 (indicated by the points 'o'). The processing method according to the invention makes it possible to determine, from a spectrum of limited resolution amplitude (as in FIG. 4a), the number and the parameters of sinusoids present in the short term in an audio signal and spectrally extract these sinusoidal components by complex subtraction of the discrete spectrum of the form A 2 (1/11, (k + kl) e18 '+1/11, (k -1 (,) eA).

The example of Figures 4a and 4b shows that the accuracy of a discrete Fourier transform of length K is generally insufficient to directly obtain an estimate of the frequency of the tonal components. It is important to note that in the domain of the discrete time Fourier transform, the amplitude of the "peak" associated with the i-th sinusoid is W (0) at the position ± ki. In the domain of the fast Fourier transform of length K, the peak at the position ± ki is not necessarily directly observable because of the inter-line precision; assuming that the position corresponds to a Fourier line, the amplitude of the "peak" associated with the sinusoid becomes: a = 2141, (0) Subsequently, the term "amplitude" of a sinusoid will essentially refer to amplitude in the Fourier domain, ai, and not the amplitude of the sinusoid in the time domain, A. Note that the value of ai naturally integrates the length of the window K, if the analysis window is normalized adequately, it is possible to ensure that WK (0) = 1.

In a particular embodiment, an interpolation method is used to improve the accuracy of this estimate. This is carried out in the module 303 of FIG. 3, at a step E603a for identifying the tonal components in the audio frequency signal represented in the frequency domain (X (k)).

The interpolation is useful for precisely estimating the parameters of the sinusoids (frequency, amplitude and phase) from an amplitude spectrum resulting from a discrete Fourier transform of relatively limited length N. Zero padding does not improve the resolution of the Fourier transform, but it allows the discrete time Fourier transform to be sampled on more lines and thus to obtain a better accuracy. In a preferred embodiment, the zero stuffing is not used for the signal analysis, but only for the pre-storage of the transform of the analysis window in order to facilitate the synthesis of sinusoids whose frequency is estimated with greater accuracy than the inter-line resolution of the discrete Fourier transform of analysis.

The estimation of the parameters (, which are the maximum amplitude and the fractional frequency of the sinusoid) is carried out according to the invention by parabolic interpolation of the amplitude spectrum, as described in the article by JO Smith and X. Serra , entitled "PARSHL: A Program for the Analysis / Synthesis of Sounds Based on a Sinusoidal Representation", Proceedings of the International Computer Music Conference, Urbana-Champaign, Illinois, 1987.

The principle of parabolic interpolation for spectral analysis is recalled in FIG. 5, where the example described in FIG. 4 is repeated. The parabolic interpolation is based on the fact that a parabola is completely determined by the values at 3 distinct points; thus, given the amplitudes (in dB) Y (/, -1), Y (/,), Y (/, +1) of 3 adjacent Fourier lines of indices 1, -1, 1, 1. +1, it is easy to determine the parameters agi, d, and, / 3 of the parabola of general equation: Y (k) = (k - (d,)) 2 + A where ce, is the curvature of the parabola, 1. + d, gives the position of the vertex of the parabola, and, / 3 is the value of the parabola at its vertex (because / = Y (1, + d,)).

From the general equation of the parabola and given the values Y (k -1), Y (k), Y (k +1), we obtain: IY (li -1) = ai (1 + di) 2 + fli Y (1i) = aidi2 +, 13 Y (li + 1) = a (1- di) 2 + fli Calculating Y (1, +1) + Y (1, -1) -2Y (1, ), from the 3 preceding equations one finds: 1 = 2- (Y (li + 1) + Y (li -1) -2Y (0) Then, by calculating Y (1, -1) -Y (1, + 1), we find: d. = (1, -1) - Y (1, +1) 4a. Z Finally, the last parameter is calculated as: Y (1,) = aidi2 + A j3, = Y (1,) - d i2 By combining the expressions of di and, one can also write: = Y (1i) - (aid) of which gives: Y (0 Y (1, -1) -Y (1, + 1) d 4 For the example of Figure 5, the parabolic interpolation on the 2 "peaks" centered at / 0 = 4 and / 1 = 12 gives the following estimated parameters: lao = -7,1964 ai = -8 , 2003 do = -0.0001 and d1 = -0.4926 130 = -5.9779 / 31 = -11.6735 In this example, the parabolic interpolation gives an estimate of the frequency of the i-th sinusoid under the form h + this e stimulation is accurate in this example. In addition to the estimated amplitude in dB,) 6 = 101og10 (ai) where at = 2 (the analysis window being normalized such that WK (0) = 1, for this example), is close to the exact values of - 6 dB and -12 dB. Thus, in the embodiment illustrated in FIG. 3, the spectrum of the signal x (n) is analyzed in block 303 in order to extract the tonal components that are to be preserved by the post-processing. The analysis comprises the steps E603a for identifying the tonal components in the audio frequency signal represented in the frequency domain and E603b for calculating a spectrum of the identified tonal components, from a sinusoidal model as described above. Note that the identification of sinusoidal components is achieved by combining the detection of sinusoids and the estimation of their parameters (frequency, amplitude, phase); the detection is used to determine if a peak corresponds to a "valid" (or "selected") sinusoid. The set of sinusoids thus detected and estimated constitutes the sinusoidal model of the signal in the current frame. This sinusoidal model is represented in an equivalent way by the complex spectrum combining the different sinusoids identified. This model is used to extract the tonal components of the signal. The extraction is carried out here by subtracting the spectrum from the sinusoidal model to the spectrum of the current signal, to obtain a residue that will be post-processed. Step E603a includes a detection of the frequency peaks in the amplitude spectrum of the signal. A peak is detected at a line of index k in the amplitude spectrum IX (k) 1 if the following criterion is satisfied: X (k) 1> IX (k -1) 1 and 1X (k) 11X (k + 1) 1, k -1, ..., K / 2 -1 The indices k locating the different peaks thus found are gathered in a vector - - -, / i), where / is the number of peaks detected. In a variant of the embodiment, it will be possible to ensure that I does not exceed a predetermined number of peaks, for example 100, in order to reduce the complexity of the operations which depend on the number of peaks. According to this notation, 1, corresponds to the number of Fourier line locating the peak of index In order to extract the tonal components in order of importance, the detected peaks are sorted in amplitude, so that they verify: X (4 ) 1X (/ 2) 1 ... 1X (//) 1 The detected peaks of position 1 ,, 1 = 1, ..., 1 are used to identify and extract tonal components in the X spectrum ( k). A peak of index 1, is supposed to be associated with a sinusoid whose amplitude ai and the frequency k, are estimated by parabolic interpolation of the spectrum of amplitude X (k) from the 3 contiguous Fourier lines of indices li -1, 1, and 1, + 1. The curvature of the parabola is calculated as described previously according to the equation: 1 ai = - (X) -2.X, m (1,) + X dB (1, ± 1)) where X dB (k) 201ogio ( X (k)). The frequency of the i-th sinusoid corresponds to the position of the vertex of the parabola: X (1 - 1) - X dB (1 + 1) k, = 1, + 4a, which can also be written: = 1 + 1 X dB (1 -1) - X dB (1 ± 1) 2 X dB (li -1) - 2.XdB (0+ X dB (1 + 1) The amplitude of the i-th sinusoid corresponds to the value at the top of the dish: a, = 10120 with X dB (1 i - X dB (1 + 1) (k, -1,) = X dB (1 i) 4 A sinusoid associated with the index peak i is selected, in other words considered as "valid", if it satisfies the following criteria: i <_20 OR 0,75 <= <1,5 (with a'f - 7,3904) a Thus, one chooses in the embodiment described here, to systematically select the sinusoids associated with the first 20 peaks detected because the AMRWB coding preserves relatively well the most important tonal components (in terms of amplitude) For the following peaks a sinusoid is considered valid if the curvature has , estimated by interpolation does not deviate too much from a curvature of The choice of selecting by default the first 20 peaks could of course be adjusted to select for example only the largest peak, without altering the invention. The aref -7.3904 value is derived from a parabolic interpolation on the amplitude spectrum of the wa (m) analysis window (of length N); this value must generally be adjusted according to the parameters of the frequency analysis implemented. If the sinusoid of estimated frequency ki and associated with the peak located on line h is "valid" (according to the criteria above), the phase of this sinusoid is estimated by interpolation. The estimation of phase 6 is a function of the phases of the 2 contiguous lines of index kr and k +: = Lki j and ki + = Si LX (kt +) - LX (kt-)> r, LX (ki +) is firstly modified as follows: LX (k) = {LX (ki +) + 2z if LX (ki-) 0 LX (ki) - 2z if LX (k) <0 Phase 6 is estimated by the following linear interpolation: LX (ki +) - LX (ki-) - In a variant of the embodiment, the estimation of the parameters can be carried out other than parabolic interpolation, in the logarithmic domain (log10) of the spectrum of amplitude and linear interpolation of the spectrum phase. For example, the estimation of the frequency can use one of the maximum detection methods described in the article by RB Fischer and DK Naidu, "A Comparison of Algorithms for Subpixel Peak Detection", which considers the context of a detection of peaks in an image with a resolution less than the pixel but easily adapts to the problem of frequency estimation. Several alternative methods of frequency interpolation can be drawn from the article by Fischer and Naidu, for example: - the Gaussian approximation which corresponds to a parabolic interpolation on the amplitude spectrum brought back to the natural logarithmic domain (In or log) - in other words, the amplitude spectrum in the log domain is approximated by a Gaussian, which gives 1 1nX (/, -1) -1nX (/, +1) 1 k = 1 + 1 1 2 lnIX ( 1, -1) 1- 2.1n1X (01+ 1n1X (/, +1) 1 - the parabolic interpolation on the direct signal (here the amplitude spectrum) with 20 25 ki = 1 1 IX (1, - 1) 1-1X (1, +1) 1 + 2 X (1, -1) 1- 2.1X (1,) 1 + 1X (1, +1) - the linear interpolation with 1 1X (/ +1 ) 1-1X (/, +1) 1 / + if 1X (11 + 1) 1> IX (11-1) 1 2 IX (/ i) I -IX (/, -1) 11X (/ 1, + 1) -1X (/ +1) li + otherwise 2 IX (01-1X (/, +1) 1 - the center of gravity with X (/, +1) 1-1X (/, +1) 1 k = 11 + (1, -1) 1 + 1X (11) 1 + 1X (1, +1) 1 It goes without saying that the frequency estimation variants above can also be adapted by replacing 1X (/,) 1 with a derived value, such as 1X (012. Similarly, given the estimated frequency, the estimate of the amplitude and phase described above can be advantageously replaced by that of a complex multiplicative factor (thus synthesizing the role of amplitude and phase) c, = a, .efi9 'using the following method. Since the detected sinus frequency is now known, it is therefore possible to define, as described below, the reference spectrum portion Y (a "pattern") which will be added to the sinusoidal modeling spectrum, after application of the complex multiplicative factor c. In the variant proposed here, the coefficient ci is calculated so as to minimize the reconstruction error of the portion of the spectrum concerned in the least squares sense. for "Least Square" in English) can be summed up in the well-known form of a ratio between a scalar product and a standard squared: ci = Wi * .X / 11Wi112 = Wi * .X / (Wi * .Wi) 20 where the notation *, applied here to a column vector of complex coefficients, corresponds to the conjugated transpose of the said vector, and the multiplication of vectors (vector line on the left, vector column on the right) is understood as a scalar product. each element of the pattern - defined specified further - includes 277 25 elements; each scalar product involves the products and sum of i pairs of complex terms, each complex product involving 2 multiplications and one addition.

It is important to note that the calculation of the coefficient c, is applicable only when the frequency of the sinusoid is at least at a distance / 7/2 of the zero frequency and the Nyquist frequency, in order to avoid the spectral folding. In the opposite case, one skilled in the art will be able to adapt the formula taking into account the frequency folding, or he may use the linear interpolation of the phase. By convention, in the following of the document as well as in the figures, we can denote by (a'k'0,) the parameters of the sinusoid i, in a way equivalent to the parameters (c'k,) even if it is the complex coefficient c, which is estimated and used, and in particular phase 0, does not need to be calculated explicitly.

In the case of a symmetric analysis window and a "recenter phase transform" as described above, the stored "pattern" is a real spectrum, so the number of multiplications and In a recommended variant of the invention, a version of the normalized pattern will have been stored in memory so that the standard of the pattern calculated on the frequency subsamples is always equal to 1. This standardization reduces the number of operations. The amplitude and phase estimation using the pattern is an advantageous alternative to the separate estimation of the amplitude and the phase by parabolic and linear interpolation, in that it avoids a certain number of expensive operations (divisions, arc-tangent phase calculations ...) This estimation by pattern is a form of analysis by synthesis, because the estimate of the parameters of the sinusoids rests on the motive used at the synthesis. that e the line of index k = 0 is also counted as a valid sinusoid with: 4 = 1X (0) 1, ki = 0, = 0 or ir depending on whether the value of X (0) is positive or negative, respectively. The block 303 then performs a step E303b for calculating the spectrum (S (k)) of the "valid" tonal or sinusoidal components thus identified. For this, it uses in an exemplary embodiment a precalculated pattern and stored in a step E300 by a module 300.

According to the embodiment described here of the treatment method of the invention, the spectrum of a sinusoid is calculated taking into account the analysis window used to perform the cost-term Fourier spectral analysis in block 302. .

As recalled above, the spectrum of a windowed sinusoid is obtained by convolution of the spectrum of the window considered with Diracs which is equivalent to translation operations (modulo 27c) and addition / recovery. Since the frequency k can take any value in [- K, 1 ±], including fractional values, the translation of the spectrum of the window by an arbitrary frequency shift is implemented at the same time. using interpolation of the spectrum of the window over a length M, where M> K. Advantageously, the device according to the invention uses a rapid method for calculating the short-term spectrum of sinusoids from a "pattern" P (k) frequency pre-calculated and stored in memory at 300. It is assumed here that the window The analysis of length N applied to each frame in block 302 is a normalized Hamming window, as described above but with N = 512, and we will write wa (m). any weighting window of length N.

In the preferred embodiment of the invention, a transform of length K = N = 512 and an interpolation factor M / K = 128, ie M = 65536, which gives us a spectrum W4k) which corresponds to Discrete (normalized) spectrum of the wa (m) window with interpolation of MIK = 128. The interpolation guarantees sufficient precision to represent (fractional) frequencies ki with an inter-line precision of 1/128 (compared to a discrete Fourier transform of length M. In fact for K = 512, the resolution of the FFT The spectrum WM (k) is represented with a resolution of 31.25 / 128 0.24 Hz. A part of the WM spectrum ( k) is illustrated in Figure 7a where the negative frequencies are included for purposes of illustration It appears that the attenuation of the secondary lobe is of the order of -43 dB which justifies using essentially the main lobe for to represent the spectrum of the window To reduce the number of values of WM (k) to store and the complexity of the complex subtraction (see the description of block 304), the "pattern" (normalized) is defined as: P (k ) = WM (k), k = 147m (0) 1 by retaining only the interval 0, 2 for the length of transform N. M e 2K This pattern matches only for positive frequencies negative frequencies are obtained by square symmetry.

In the embodiment described here, ri = 4 lines (for the transform length K) are used. The choice of ri = 4, given the interpolation factor M / K = 128, advantageously limits the storage of the "pattern 'P (k) to 2 + 1 = 257 values .The pattern P (k) K is illustrated in FIG. FIG. 7b represents its amplitude, its phase being zero in the privileged case of a symmetrical window and of a "recenter phase" transform, thus the "pattern" P (k), k = 0, ... , 256, is calculated prior to the steps implemented in the processing method and pre-stored in ROM memory in block 300. This spectral pattern can be stored in polar representation (amplitude, phase) as well as in Cartesian representation ( real part, imaginary part), according to the chosen calculation methods. It goes without saying that the invention covers the variant where this pattern P (k) is redundantly stored, also including the values corresponding to the negative frequencies; this variant, however, provides only a negligible advantage in terms of calculation complexity.

Step E303b of block 303 then performs the calculation of the combined spectrum. The combined spectrum of the sinusoids is first initialized to the null value: S (k) =, k = 0, ..., K / 2 It is subsequently considered that these values are collected in a vector of length / Q2 + 1: S = (S (0), S (1), ..., S (K / 2)) The vector S is updated iteratively by adding the discrete spectrum of each sinusoid (or tonal component) d index i. We approximate the discrete spectrum of a sinusoid of parameters (a 'k, 0) by frequency translation of ± k frequency lines (with the inter-line accuracy of K / M = 1/128) and addition-recovery of the "pattern" pre-computed P (k), k = 0, ..., 256, and multiplication by ci or ateA This calculation is illustrated in detail in Figure 8, for the iteration i, that is to say for the calculation of the discrete spectrum of the sinusoid of index i (> 0) in the case where this sinusoid is valid.

Thus, the block E801 indicates that the flowchart of FIG. 8 implements the construction (const) of the spectrum from the estimated parameters a ,, 0, of the sinusoid. A step E802 initializes the result of the summation, which is implemented as a loop (E803, E804, E807, e808). The terms of the sum are calculated in steps E805 and E806, the details of which are shown in Figure 9.

Thus, for the i-th sinusoid, the deviations kr and k7 between the frequency ki and the integer indices of the near lines for the transform of length K are converted by homothety into indices AT and 3; relating to the transform to length M, as follows: Δt = K (kt-LkiJ) = K (ki-k = 1 1; 7-j-ki) = `1147 (ki + -ki) where L.] represents round to the lower integer. By definition Ai + Ai11 / The spectrum of the i-th sinusoid is constructed by retaining only i = 4 lines of the pattern P (k) (for positive frequencies). The flowchart of FIG. 8 implements the summation according to the following equation: 2 S = S + spec (KE p P (Lp.f, `+ spec (kti + p, where the function spec (k, V) is now defined through the flowchart shown in Figure 9.

In order to complete the description made above of the so-called "LS" method of estimating the complex amplitude parameter c1 equivalent to the amplitude parameters aiet of phase 9 ', this equation is reformulated as follows: S = S + = S + i 2 PP ( [19 1A + 1 + = spec (ki + - p -1, P ([19. '+' + ATM + spec (KE p = 0 This factorization is exact under the assumption (generally verified, except in special cases) that the sinusoid is not too close to the null and Nyquist frequencies (of indices 0 and K / 2), so that the sampled pattern adds up only on the half-spectrum of synthesis, without spectral folding.

Block E901 indicates that the flowchart of FIG. 9 implements the partial updating of the spectrum (spec) in the interval [0, K / 2] for the sinusoid considered. This update takes as parameter the Fourier line and the spectral value Associated with this line. Step E902 initializes the result of the calculation. Then 4 cases are distinguished: o k <0 (step E903): the hermitian symmetry is put to advantage in step E904, which amounts to writing the complex conjugate of V to the position line -k; p = 0 ieP (Lp + Al) ok = 0 or K / 2 (step E905): the DC and Nyquist lines being by definition real and without a factor 1/2, the considered value is 2.Re (V) where Re (.) gives the real part (step E906); o k> KI2 (step E907): the periodicity of the discrete spectrum and the Hermitian symmetry for the negative frequencies is put to advantage in step E908, which amounts to writing the complex conjugate of V to the line of position K-k; o In other cases (step E909): the value V is directly added to the position k.

The calculation of S presented previously is repeated for each valid sinusoid of index 4, which amounts to calculating: S i = 0 2 P ([19. '+' + ATM + spec (ki + + ql [p4 + Al)) SpeC (ki + p = 0 where I is the number of valid sinusoids, or identified tonal components It is assumed here, without loss of generality (because it is enough of an appropriate indexation), that all the invalid sinusoids are not counted. Note that for the sinusoid of index 0, which corresponds to the DC line (DC component), is taken into account in the formula above.This operation makes it possible not to modify the value DC (DC component) by post-processing. .

In a variant of the embodiment, the peak detection process is repeated iteratively, in order to detect at the i-th iteration only the peak position corresponding to the maximum of the spectrum X (k) which is subtracted at each iteration. spectrum of sinusoids already extracted. This variant corresponds to a form of analysis known in the state of the art under the name of "matching pursuit".

The block 304 of FIG. 3 computes at step E304 a residual by the difference between the decoded audio-frequency signal and a signal coming from the spectrum of the tonal components. In the present case of FIG. 3, this residue is obtained in the frequency domain by the difference between the spectrum of the audiofrequency signal and the spectrum of the tonal components identified. This gives the residual spectrum after extraction (subtraction) of the synthesized sinusoids. The residue is obtained in the complex domain as follows: R (k) = X (k) - S (k), k = 0,, K 12 In the ideal case where the original signal is constituted as a combination of sinusoids which are preserved by coding-decoding, the residue then corresponds to the coding noise. However, in general, the resulting residue includes: o a part corresponding to the sinusoidal modeling residue of the original signal (including the non-tonal components of the signal as well as the modeling errors) o a part corresponding to the coding noise.

In block 305 of FIG. 3, a spectral attenuation step E305 is implemented. This operation generally uses all available information. It may for example be a function of the classification of the extracted tonal components, the residual signal and the decoded parameters. Block 305 spectrally attenuates the spectrum of the residue in the following form: Rat, (k) = g (k) .R (k), k = 0, ..., K 1 2 where g (k) is a Adaptive gain generally satisfying: 0 g (k) 1. In a variation of the preferred embodiment, the gain applied to the residue will be temporally smoothed to attenuate interframe variations and avoid certain artifacts.

In this case block 305 will perform the calculation R (k) = (k) .R (k), k = 0,..., K 1 2 where the smoothed gain g (k) is obtained for example by exponential smoothing. of g (k) according to the equation g (k) = (k) <(k) + (1- (k)) g (k) with a forgetting factor Ç (k) which can be either fixed (by example of value 0.75) is adaptive (for example of value Ç (k) = 1- g (k)). In the particular embodiment, the classification information obtained by the block 310 is used. This block 310 will be described later. Thus, the attenuation gain is for example forced to the value g (k) = 1, k = 0,, N / 2 when the current frame is classified as "other". When the frame is classified as "music", in the preferred embodiment, the gain g (k) is calculated as the gain of a frequency (constrained) Wiener filter in the form: RSB (k) g (k) = max, k = 0, ..., N 12 g min 1+ RSB (k) where gmin defines a maximum reduction level (with for example grnin = 0.25) and the signal-to-noise ratio RSB (k) is estimated by the equation RSB (k) = (k) 12 B (k) X (k) 12 corresponding to the energy of the line of index k and B (k) being an estimate of the level of the calculated coding noise from the residual signal R (k) according to the equation: B (t) (k) = 0.2.B (t-1) (k) + 0, 6 .E (t) (k) + 0, 2.E (t-1) (k) where tet t-1 correspond to the current and previous frames, respectively, and 1 stop (k) -1 E (t) (k) = IX (i) 12 stop (k) - start (k) ,, start (k) is the average energy in the Bark band (incomplete near k = 0 and k = K2) centered around the line k. The Bark band is defined by the lines in the [start (k), stop (k) -1] interval, where the start and end of stop (stop) arrays are defined below for the example K = 512: start = {0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 124, 125, 126, 127, 128, 129, 130, 131, 132, 132, 133, 134, 135, 136, 137, 138, 139, 139, 140, 141, 142, 143, 144, 145, 146, 147, 147, 148, 149, 150, 151, 152, 153, 154, 154, 155, 156, 157, 158, 159, 160, 161, 161, 162, 163, 164, 165, 166, 167, 168, 168, 169, 170, 171, 172, 173, 174, 174, 175, 176, 177, 178, 179, 180, 180, 181, 182, 183, 184, 185, 186, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 213, 214, 215, 216, 217, 218, 219, 220, 220, 221, 222, 223, 224, 225, 226, 226, 227, 228, 229, 230, 231, 232, 233} stop = {3, 4 , 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 23, 24, 25, 26, 27, 28, 29, 30 , 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56 , 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83 , 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 108, 109, 110, 111 , 112, 113, 114, 115, 116, 118, 119, 120, 121, 122, 123, 124, 125, 126, 128, 129, 130, 131, 132, 133, 134, 135, 136, 138, 139 , 140, 141, 142, 143, 144, 145, 147, 148, 149, 150, 151, 152, 153, 155, 156, 157, 158, 159, 160, 16 1, 162, 164, 165, 166, 167, 168, 169, 170, 172, 173, 174, 175, 176, 177, 178, 180, 181, 182, 183, 184, 185, 186, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 202, 203, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 238, 239, 240, 241, 242, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257, 257} It should be noted that other smoothing coefficients (of value 0.2, 0.6 and 0.2) can be adjusted differently. .

In this embodiment, the coding noise level 5 (k) is therefore estimated using "floating" critical bands in the sense that each Fourier line of index kse is associated with an integration interval [ start (k), stop (k) -1]. The tables given above correspond to an interval of 1/2 Bark on either side of each line of index k, in other words the interval [start (k), stop (k) -1] has a width (approximate) of 1 Bark. In a variant of the embodiment, the estimation of the coding noise may use intervals of different size, for example 2 Barks or 1/2 Bark. Moreover, the maximum attenuation level, gm, n, can be made adaptive according to the classification, as in the technique of Vaillancourt et al. Moreover, the Wiener gain can be replaced by another reduction function using the signal-to-noise ratio RSB (k) The signal-to-noise ratio RSB (k) is in essence relatively high in the vicinity of the extracted sinusoids - indeed in this vicinity the sinusoidal modeling residual is of lower level than that of the signal since a sinusoid has been cut off; on the other hand, for the lines of index k located in the inter-tone frequency zones, the residue has a level corresponding to that of the signal and the ratio RSB (k) will have a value close to 0 dB. In a variant of the embodiment, the value of the signal-to-noise ratio can be artificially lowered by applying to it a factor (for example 1/16) in order to force a stronger attenuation than the Wiener gain associated with a 0 dB SNR. .

In a variant of the embodiment, the gain g (k) is calculated from the coefficients of the decoded LPC filter λ (z) associated with the last subframe (5 ms) of the current frame. These coefficients are provided by the decoding block 301. That is, the transfer function filter: λ (z / yi) H (z) = A (z / r2) with yi = 0.6 and 72 = 0.95. It will be recalled that the coefficients of the filter λ (z) and therefore the transfer function H (z) are defined for the sampling frequency of 12.8 kHz. In the present description, the processed signal is sampled at 16 kHz. We then define K '= LKx12, 8/16 j = 410, which is the frequency index corresponding approximately to the frequency 12.8 kHz. From there, the gain g (k) is calculated as follows: g (k) g1H- (e, 2zkix-) k = 0, ..., K72 g (K72) k = K72 + 1, ..., K12 KY2 where go is a normalization factor such that g (k) = 1. It is obvious that a k = 0 transposition of the invention to a sampling domain identical to that where is defined  (z) would imply a value K '= K. In order to avoid a too strong attenuation, the value of the gain g (102) k = K72 + 1, ..., K / 2 This variant which uses the decoded linear prediction filter has the advantage of providing a weighting function relatively smooth and slowly varying from one frame to another for a stationary signal; on the other hand, it has the disadvantage of staining the residue by the (extended) response of the H (z) filter. Other embodiments of the calculation of the gain are possible, without changing the principle of the invention.

In a variant of the embodiment thus described, the gain is calculated according to the sinusoidal model. The tonal components that are extracted and synthesized have the parameters (a1, k., 6e) with i E / 1, where / is the number of valid sinusoids (after elimination of invalid sinusoids that are not counted and after appropriate indexing) and k1 <k2 <<k1. Under these conditions we introduce an attenuation parameter in dB x (k), k = 0, ..., N / 2, such that: z (k) = sine Tc k-k. , with i such that kk, ± 1 ki 2 with the following conventions: 1 (0 = -k1 and k1 + 1 = K An example for this attenuation parameter is shown in Figure 10a for the following example with I = 6: ki = 0, ..., 6 = {5.25, 15.1, 35.5, 120, 180, 240} It can be seen that the attenuation parameter has a value that remains close to zero around the frequencies. attenuation can be any positive and normalized function at a maximum value of 1 (without loss of generality because it is sufficient for normalization) with a minimum (close to 0) at the ends of [ki, k, ± 11 and a unique maximum within the interval [ki, k + 1]. In a further variant, the attenuation parameter in dB is given by: 2 x (k) _ 4 (k -k) (kr kti) (Lis + d- k) if ke [k'k, ± 1] with the following conventions: 1 (0 = -k1 and k1 + 1 = K -k1) In this case, the attenuation parameter is a parabolic parcel function, where the vertices of each parable have a the value of 1 and the values are zero for the lines k that correspond computed can be saturated so that: g (k) = {flux (golH (e12'kul, g imii) k = 0, ..., K 72 ( exactly) at the frequency of one of the sinusoids. An example for this attenuation parameter is shown in FIG. 10b for the same frequencies k as before. In an advantageous embodiment, the gain g (k) is given by: 2z (k) g (k) = 10 where 2 is the attenuation level in dB desired - for example 2 = 6 dB. Block 310 thus performs a classification (step E310) of the audiofrequency signal x (n). The classification serves to detect the signals that are targeted by the post-processing, essentially the musical signals with a harmonic structure; thus on the other signals, the post-processing must remain "transparent", in particular for the speech signals and the transient signals (attacks of musical instruments for example). The details of the classification go beyond the object of the invention described here. Nevertheless, several examples of classification implementation are presented. In the main embodiment, the classification comprises two possible classes: "music" and "other". This classification is derived from ITU-T Recommendation G.720.1 "Generic sound activity detector (GSAD)" with the following adaptations: o The G.720.1 standard is used in broadband (at 16 kHz); the GSAD algorithm works with frames of 10 ms and without pre-exploration or 20 "lookahead" in English. It is applied to the decoded signal on 4 frames (ie 40 ms) in order to cover a temporal support including the signal analyzed. o If the 4 decisions obtained in the current frame are "music", block 310 classifies the current frame as "music", otherwise as "other". In a variant of the embodiment, the classification will be implemented as in the previously cited article by Vaillancourt et al. In this case, the signal is assumed by default to be a music signal, and the maximum attenuation level gn-'n is adapted according to a parameter of variation of energy oE calculated according to the principle of the article of T. Vaillancourt et al. cited previously in order to quantify the degree of variation of the spectral energy in the FFT domain: ## EQU1 ## where Et-1 is an average of the energy differences between frames adjacent. A var (adaptive) threshold is then applied to the o-E parameter to determine the signal stability level. The details of the o-E calculation are not repeated here, but they can be found in the article by T. Vaillancourt et al. 2 It is assumed here that the calculation of oE leads to a maximum attenuation level between 0 and 12 dB, as in the article by T. Vaillancourt et al. In this case the value of is given by the level of maximum attenuation provided by the classification.

In another variant of the embodiment, the classification information could be available to the decoder as a multiplexed parameter in the bit stream. Block 306 of FIG. 3 performs a step of combining the attenuated residue and the signal originating from the spectrum of tonal components. In this embodiment, step E306 recombines the attenuated residual spectrum and the spectrum of sinusoids identified as follows: T (k) = S (k) + Rajk), k = 0, ..., K 1 2 The spectrum the post-processed signal T (k), k = 0, ..., K 12, is extended in T (k), k = 0, ..., K -1 with T (k) = T * (K - k), k = K 1 2 + 1,, K -1 where * is the complex conjugate. To obtain the time signal t (m) the signal T (k) is transformed by the block 307 which comprises an inverse FFT (step E307) of length K-512, followed by a circular permutation of N / 2 temporal samples to the right performed to compensate for the "phase recentering" when it is applied to the analysis, as described above as the main embodiment. This amounts to writing the inverse Fourier Transform as follows :: 1 K-1 2z / c N) t (M) = - T (k) .e K 2, = ... K -1 K k = 0 In a variant of the invention where the phase recentering is not applied to the analysis, the frequency-time transform (block 307) for the synthesis consists of a simple inverse FFT: 1 K-1 2z / ct (m) = - In the two described embodiments, the inverse FFT is calculated according to the method described in the HV article. Sorensen et al., Entitled "Real-valued fast Fourier transform algorithm," IEEE Trans. on Signal Processing, Vol.35, No.6, pp 849-863, 1987. The synthesis window ws (m) is defined as 0 m = 0, ..., D -1 (vs (m) = wola ( m -D) wa (m) m = D, N -1 'm = 0, ..., N -1 where the quantity D = N - L - L01 of introduced and ignored samples has been introduced for convenience null by the synthetic windowing, and where the overlap overlapping window (111) = L + Lola -1 (OLA for Overlap-Add) is defined by: vola (m) = ((m = 0, Lola -1 -1 1-cos -Tt- (m + 0 .5) = Lola - - - L -1 2 Lola // m = L,, L + Lola -1 1 1 ((1+ cos -Tt- ( (m-L) +0.5) 2 \ Lola with 1.01a = 40 and L = 320. The windowswoia (m) and ws (m) are further illustrated in Figures 6b and 6c respectively, as illustrated in Figure 11, the synthesis window is intended to simultaneously translate the temporal envelope rectification and to prepare the addition-overlay method, and the vector of samples resulting from this windowing is defined: tw (m) = ws (m) t (m In practice that the first D samples do not need to be calculated as null. The last Lola samples, affected by the decreasing slope of the window vola (in), can be used only for the synthesis of the following frame, and are stored for this purpose in a vector tsave (n) = t. (n + N - D), n = 0, ..., Lola For the synthesis of the current frame, we combine the samples ts ("") (n) aPv, saved to the previous frame (or else set to zero if the postprocessing has not been applied to the previous frame) and a portion of the vector tw (m), as follows, completing the overlap-addition process: It. (n + D) + ts ( 'a'rvee ") (n) n = 0, ..., Lola -1 It is verified that, in view of the addition-recovery process, the L post-processed signal samples xps (n) are obtained with a delay of Lola samples compared to the current frame of L samples considered input of the post-processing Thus, in the embodiment described, the post-processing induces an additional algorithmic delay of Lola = 40 samples, or 2.5 ms. The main embodiment is based on a spectral attenuation of the sinusoidal modeling residue, the attenuation described above corresponds to a convection. olution x post (n) = t ola (n + D) n = L, L -1 circular (multiplication of the residual by a gain in the domain of the discrete Fourier transform). However, it is known to those skilled in the art that a gain function having a too strong dynamic can cause a circular convolution result very different from a linear convolution. In one variant of the embodiment, a zero padding and an overlap-add / save technique may be implemented to take into account the non-zero length of the impulse response. associated with the gain function achieving the spectral attenuation. In general this variant causes an additional time delay.

In a second embodiment, the steps of obtaining the residue, attenuation and combination of the processing method as described with reference to Figure 3, can be performed in the time domain. Figure 12 illustrates this process as well as the corresponding processing device 1200. In this figure, the blocks 300, 301, 302, 303 and 310 remain identical to those described above with reference to FIG. 3. The signal modeled in sum of sine still called spectrum of the tonal components (S (k)) is in this first embodiment restored in the time domain (block 1201) by an inverse transform step E1201 and its time envelope rectified by an appropriate windowing, leaving aside the portion located on the overlap area (the last Lola samples) for processing the next frame, and suitably combining the same saved portion with the previous frame to reconstruct the start of the frame (with a Lola delay). The residue is then calculated in step E1202 by block 1202 by simple subtraction between the delayed decoded audio signal (x (n)) of Lda samples and the signal from the spectrum of the tonal components (s (n)) after transformation into 1201. To this residue, a spectral attenuation (block 1203) is then applied by temporal filtering (step E1203) whose parameter design is controlled by the sinusoidal modeling module (block 303) and / or decoded parameters (p. decod.) from the bit stream and / or the classification module (block 310).

The step of combining E1204 of the thus attenuated residue (ratt (n)) and the temporal signal of the tonal components (s (n)) is performed by block 1204 to obtain the processed time signal (xpost (n)). In the same way as the method described with reference to FIG. 3, the processing carried out on the signal makes it possible to reduce the coding noise that could subsist on the musical signals during the decoding performed by decoders more adapted to speech signals. This method therefore provides both better speech and music signals, while keeping existing speech coders and decoders.

FIG. 13 represents an example of a hardware embodiment of a processing device according to the invention. This may be an integral part of an audio decoder or equipment receiving decoded audio signals. This type of device comprises a PROC processor cooperating with a memory block BM having a memory storage and / or work MEM. The memory block may advantageously comprise a computer program comprising code instructions for implementing the steps of the processing method in the sense of the invention, when these instructions are executed by the processor PROC, and in particular the identification steps of the processor. tonal components in the audio frequency signal represented in the frequency domain, calculating a spectrum of the tonal components identified from a sinusoidal model, obtaining a residual by the difference between the decoded audio-frequency signal and a signal from the spectrum of tonal components, spectral attenuation of the residue and combination of the attenuated residue and the signal from the spectrum of tonal components.

Typically, the description of FIG. 3 or FIG. 12 shows the steps of an algorithm of such a computer program. The computer program can also be stored on a memory medium readable by a reader of the device or downloadable in the memory space thereof. The memory MEM stores the pattern P (k) used in the method according to one embodiment of the invention and generally, all the data necessary for the implementation of the method. Such a device comprises an input module E adapted to receive an audiofrequency signal (x (n)) and an output module S adapted to transmit the processed signal xpost (n). In a possible embodiment, the device thus described may also include the decoding functions in addition to the processing functions according to the invention.

Claims (11)

REVENDICATIONS1. A method of processing a decoded audio frequency signal, characterized in that it comprises the following steps: identification (E303a) of tonal components in the audiofrequency signal represented in the frequency domain; calculating (E303b) a spectrum of tonal components identified from a sinusoidal model; obtaining (E304, E1202) a residue by the difference between the decoded audio frequency signal and a signal from the spectrum of the tonal components; spectral attenuation (E305, E1203) of the residue; and combining (E306, E1204) the attenuated residue and the signal from the spectrum of the tonal components. 2. Method according to claim 1, characterized in that the steps of obtaining the residue, attenuation and combination are performed in the frequency domain. 3. Method according to claim 1, characterized in that the steps of obtaining the residue, attenuation and combination are performed in the time domain. 4. Method according to claim 1, characterized in that the identification of the tonal components is performed by an interpolation method on a predetermined number of frequency lines. 5. Method according to claim 1, characterized in that the identification of the tonal components uses a spectral pattern (P (k)) pre-calculated and stored in memory. 6. Method according to claim 1, characterized in that the calculation of the spectrum of the tonal components uses a spectral pattern (P (k)) pre-calculated and stored in memory. 7. Method according to claim 1, characterized in that the spectral attenuation is performed according to a classification data of the audiofrequency signal. 8. Method according to claim 3, characterized in that the spectral attenuation is performed by filtering the residue in the time domain. 9. Device for processing a decoded audio-frequency signal, characterized in that it comprises: an identification module (303) of tonal components in the audio frequency signal represented in the frequency domain; a calculation module (303) for a spectrum of tonal components identified from a sinusoidal model; a module for obtaining (304, 1202) a residue by the difference between the decoded audio frequency signal and a signal from the spectrum of the tonal components; a spectral attenuation module (305, 1203) of the residue; and a combination module (306, 1204) of the attenuated residue and the signal from the spectrum of the tonal components. 10. Audiofrequency signal decoder characterized in that it comprises a processing device according to claim 9. 11. Computer program comprising code instructions for the implementation of the steps of the processing method according to one of claims 1 to 8, when these instructions are executed by a processor. 25