FR2796189A1 - AUDIO CODING AND DECODING METHODS AND DEVICES - Google Patents

Abstract

The encoder estimates a fundamental frequency (F0) of the audio signal, determines a spectrum of the audio signal by a frequency domain transform of a frame of the audio signal, and includes in the digital stream transmitted to the decoder encoding data of a harmonic component of the audio signal, comprising data representative of spectral amplitudes associated with frequencies multiple of the estimated fundamental frequency. The harmonic component coding data further comprises, for at least one of the multiple frequencies of the estimated fundamental frequency, data (iDELTAphi) relating to the phase of the spectrum of the audio signal in the vicinity of this multiple frequency.

Description

AUDIO CODING AND DECODING METHODS AND DEVICES

  The present invention relates to the field of coding audio signals. It applies in particular, but not exclusively, to speech coding, in narrowband or wideband, in various ranges of coding rate. The design of an audio codec is primarily aimed at providing a good compromise between the flow rate of the stream transmitted by the encoder and the quality of the signal.

  audio that the decoder is able to reconstruct from that stream.

  In this respect, encoder families have been developed based on an analysis of the audio signal in the spectral domain: the encoder estimates a fundamental frequency of the signal, representing its pitch, and the spectral analysis consists of determining parameters representing the harmonic structure of the signal at frequencies which are integer multiples of this fundamental frequency. A modeling of the non-harmonic component, or voiceless, can also be performed in the spectral domain. The parameters transmitted to the decoder typically represent the spectrum module of the voiced and unvoiced components. There is added information representing either voiced / unvoiced decisions relating to different portions of the spectrum, or information on the probability of voicing of the signal, allowing the decoder to determine in which portions of the spectrum it must use the voiced component. or the

unvoiced component.

  These families of encoders include encoders of the type MBE ("Multi-Band Excitation"), or the type of encoders STC ("Sinusoidal Transform Coder"). As a reference, mention may be made of US patents

  4,856,068, 4,885,790, 4,937,873, 5,054,072, 5,081,681, 5,195,166, 5,216,747,

  226,084, 5,226,108, 5,247,579, 5,473,727, 517,511, 5,630,011, 5,630,012,

  649,050, 5,651,093, 5,664,051, 5,664,052, 5,684,926, 5,701,390, 5,715,365,

  749,065, 5,752,222, 5,765,127, 5,774,837 and 5,890,108. An object of the present invention is to allow, in an analysis scheme for analysis in the spectral domain, to improve the modeling of the

  phases of the signal spectrum by the decoder.

  The invention thus proposes a method of encoding an audio signal, in which a fundamental frequency of the audio signal is estimated, determining a spectrum of the audio signal by a transformation in the -2frequential domain of a frame of the audio signal, and encoding data of a harmonic component of the audio signal, including data representative of spectral amplitudes associated with multiple frequencies of the estimated fundamental frequency, are included in an output digital stream. According to the invention, said data representative of spectral amplitudes associated with multiple frequencies of the fundamental frequency are obtained by means of cepstral coefficients calculated by transforming in the cepstral domain a compressed upper envelope of the spectrum of the audio signal, and the coding data. of the harmonic component furthermore comprise, for at least one of the multiple frequencies of the estimated fundamental frequency, data relating to the phase of the spectrum of the audio signal at the

  neighborhood of said multiple frequency.

  Another aspect of the present invention relates to a method of decoding an input digital stream representing an encoded audio signal, in which encoding data of a fundamental frequency of the signal is received in the input digital stream. audio and encoding data of a harmonic component comprising cepstral coefficient quantization data representing an upper spectral envelope of the audio signal, and generating a spectral estimate of the harmonic component whose modulus is determined by said spectral envelope greater than the frequencies multiples of the fundamental frequency. According to the invention, in order to determine the phase of the spectral estimation of the harmonic component, the cepstral coefficients are transformed by scaling and smoothing in the cepstral domain, and minimum phases of the audio signal are calculated at frequencies that are multiples of the fundamental frequency. on the basis of transformed cepstral coefficients. According to the invention, it is also possible to determine a variation of the phase of the spectral estimation of the harmonic component on at least one frequency interval surrounding a frequency that is a multiple of the fundamental frequency on the basis of quantization parameters included in the data. coding of the

harmonic component.

  The invention also proposes an encoder and an audio decoder

  comprising means for the implementation of the above methods.

  Other features and advantages of the present invention

  will appear in the following description of examples of non

  with reference to the accompanying drawings, in which: FIG. 1 is a block diagram of an audio coder according to the invention; Figures 2 and 3 are diagrams illustrating the formation of the audio signal frames in the encoder of Figure 1 - Figures 4 and 5 are graphs showing an example of the audio signal spectrum and illustrating the extraction of the upper and lower envelopes of this spectrum; FIG. 6 is a block diagram of an example of quantization means that can be used in the encoder of FIG. 1; FIG. 7 is a block diagram of means that can be used to extract

  parameters relating to the phase of the non-

  harmonic in a variant of the encoder of Figure 1; FIG. 8 is a block diagram of an audio decoder corresponding to the coder of FIG. 1; FIG. 9 is a flowchart of an exemplary procedure for smoothing spectral coefficients and extracting minimum phases implemented in the decoder of FIG. 8; FIG. 10 is a block diagram of spectral analysis and mixing modules of harmonic and non-harmonic components of the audio signal; FIGS. 11 to 13 are graphs showing examples of non-linear functions that can be used in the analysis module of the figure; FIGS. 14 and 15 are diagrams illustrating one way of temporally synthesizing the signal frames in the decoder of FIG. 8; FIGS. 16 and 17 are graphs showing windowing functions that can be used in the synthesis of the frames according to FIGS. 14 and 15; Figures 18 and 19 are block diagrams of interpolation means for use in an alternative embodiment of the encoder and the decoder; FIG. 20 is a block diagram of interpolation means that can be used in another embodiment of the encoder; FIGS. 21 and 22 are diagrams illustrating another way of temporally synthesizing the signal frames in the decoder of FIG. 8, using interpolation of parameters; Figures 23 to 25 are block diagrams of post-processing means variants of the cepstral coefficients representing the upper envelope of the signal spectrum in the encoder of Figure 1, and Figure 26 is a partial block diagram of 'a decoder associated with a

encoder according to Figure 25.

  The encoder and decoder described below are digital circuits which, as is usual in the field of audio signal processing, can be implemented by programming a digital signal processor.

  (DSP) or an application specific integrated circuit (ASIC).

  The audio coder shown in FIG. 1 processes an input audio signal x which, in the nonlimiting example considered hereinafter, is a speech signal. The signal x is available in digital form, for example at a sampling frequency Fe of 8 kHz. It is for example delivered by an analog-digital converter processing the amplified output signal of a microphone. The input signal x can also be formed from another

  version, analog or digital, coded or not, of the speech signal.

  The encoder comprises a module 1 which forms successive audio signal frames for the different processes performed, and an output multiplexer 6 which delivers an output stream O containing for each frame sets of quantization parameters from which a decoder will be able to of

  synthesize a decoded version of the audio signal.

  The structure of the frames is illustrated in FIGS. 2 and 3. Each frame 2 is composed of a number N of consecutive samples of the audio signal x. The successive frames have mutual time offsets corresponding to M samples, so that their overlap is of L = N-M samples of the signal. In the example considered, where N = 256, M = 160 and L = 96, the duration of the frames 2 is N / Fe = 32 ms, and one frame is

formed all M / Fe = 20 ms.

  Conventionally, the module 1 multiplies the samples of each frame 2 by a windowing function fA, preferably chosen for its good spectral properties. The samples x (i) of the frame being numbered from i = 0 to i = N-1, the analysis window fA (i) can thus be a Hamming window, of expression: fA (i) = 0 , 54 + 0.46.cos (27 (Ni) 2 (1) or a Hanning window of expression: f 2 (i) ((i 2 i (N 1) / 2 (2) or a window of Kaiser, of expression [X (i- (N-1) / 2)

I0 o _ -

  N fA (i) = l (3) (where oo is a coefficient for example equal to 6, and lo (.) Denotes the function of

Bessel of index 0.

  The encoder of FIG. 1 carries out an analysis of the audio signal in the spectral domain. It comprises a module 3 which calculates the Fast Fourier Transform (TFR) of each signal frame. The signal frame is formatted before being submitted to the TFR module 3: the module 1 adjoins N = 256 samples to zero in order to obtain the maximum resolution of the Fourier transform, and it also performs a circular permutation of the 2N = 512 samples to compensate for the phase effects resulting from the analysis window. This modification of the frame is illustrated in FIG. 3. The frame whose fast Fourier transform is calculated on 2N = 512 points starts with the N / 2 = last 128 weighted samples of the frame, followed by the N = 256 samples at zero, and ends with the N / 2 = 128 first

  weighted samples of the frame.

  The TFR 3 module obtains the spectrum of the signal for each frame whose module and phase are respectively denoted | X | and (px, or | X (i) 1 and (px (i) for frequency indices i = 0 to i = 2N-1 (thanks to the symmetry of the Fourier transform and the frames, we can confine ourselves to values for

0 <i <N).

  A fundamental frequency detector 4 estimates for each signal frame a value of the fundamental frequency Fo. The detector 4 can apply any known method of analyzing the speech signal of the frame to estimate the fundamental frequency F0, for example a method based on the autocorrelation function or the AMDF function, possibly preceded by a module. bleaching by linear prediction. The estimation can also be carried out in the spectral domain or in the cepstral domain. Another possibility is to evaluate the time intervals between the consecutive breaks of the speech signal attributable to closures of the speaker's glottis occurring during the duration of the frame. Well known methods that can be used to detect such micro-breaks are described in the following articles: M. Basseville et al., "Sequential detection of abrupt changes in spectral characteristics of digital signals" (IEEE Trans. IT-29, No. 5, pages 708-723); R. Andre-Obrecht, "A new statistical approach for the automatic segmentation of continuous speech" (IEEE Trans., Acous., Speech and Sig Proc., Vol 36, No. 1, January 1988); and C. MURGIA et al., "An algorithm for the estimation of glottal closure instants using the sequential detection of abrupt changes in speech signais" (Signal

  Processing VII, 1994, pages 1685-1688).

  The estimated fundamental frequency F0 is quantized, for example scalar, by a module 5, which supplies the output multiplexer 6 with an iF index for quantizing the fundamental frequency for each frame

of the signal.

  The encoder uses cepstral parametric modelings to represent an upper envelope and a lower envelope of the audio signal spectrum. The first step of the cepstral transformation consists in applying to the signal spectrum module a spectral compression function, which can be a logarithmic or root function. The module 8 of the encoder thus operates, for each value X (i) of the signal spectrum (0 <i <N), the following transformation: LX (i) = LoglX (i) l) (4) in the case of a logarithmic compression or LX (i) = IX (i) | (5) in the case of root compression, y being an exponent between 0

and 1.

  The compressed spectrum LX of the audio signal is processed by a module 9 which extracts spectral amplitudes associated with the harmonics of the signal corresponding to the multiples of the estimated fundamental frequency F0. These amplitudes are then interpolated by a module 10 in order to obtain a 7 -

  compressed upper envelope denoted LXsup.

  It should be noted that the spectral compression could similarly be performed after the determination of the amplitudes associated with the harmonics. It could also be done after the interpolation, which would only change the shape of the interpolation functions. The module 9 for extracting the maxima takes into account the possible variation of the fundamental frequency on the analysis frame, errors that can be made by the detector 4, as well as inaccuracies related to the discrete nature of the frequency sampling. For this, the search for the amplitudes of the spectral peaks does not consist simply in taking the values LX (i) corresponding to the indices i such that i.Fe / 2N is the frequency closest to a harmonic of frequency k.F0 (k > 1). The spectral amplitude chosen for a harmonic of order k is a local maximum of the modulus of the spectrum near the frequency k.F0 (this amplitude is obtained directly in compressed form when the spectral compression 8 is

  performed before extraction of maxima 9).

  FIGS. 4 and 5 show an example of the shape of the compressed spectrum LX, where it can be seen that the maximum amplitudes of the harmonic peaks do not necessarily coincide with the amplitudes corresponding to the integer multiples of the estimated fundamental frequency F0. The flanks of the peaks being quite steep, a small positioning error of the fundamental frequency Fo, amplified by the harmonic index k, can strongly distort the estimated upper envelope of the spectrum and cause a poor modeling of the signal's signal structure. . For example, taking the spectral amplitude directly for the frequency 3.F0 in the case of FIGS. 4 and 5 would produce a significant error in the extraction of the upper envelope in the vicinity of the harmonic of order k = 3, then that it is an energetically important zone in the drawn example. By interpolating from the true maximum, we avoid this kind

  estimation error of the upper envelope.

  In the example shown in FIG. 4, the interpolation is performed between points whose abscissa is the frequency corresponding to the maximum of the amplitude of a spectral peak, and whose ordinate is this maximum, before or

after compression.

  -8 - The interpolation performed to compute the upper envelope LX_sup is a simple linear interpolation. Of course another form

  interpolation could be used (eg polynomial or spline).

  In the preferred variant shown in FIG. 5, the interpolation is performed between points whose abscissa is a frequency k.Fo which is a multiple of the fundamental frequency (in fact the closest frequency in the discrete spectrum) and whose ordinate is the maximum amplitude, before or after

  compression, the spectrum in the vicinity of this multiple frequency.

  Comparing FIGS. 4 and 5, it can be seen that the extraction mode according to FIG. 5, which repositioned the peaks on the harmonic frequencies, led to a better accuracy on the amplitude of the peaks that the decoder will attribute to the multiple frequencies of the fundamental frequency. There may be a slight shift in frequency of the position of these peaks, which is not perceptually very important and is not avoided either in the case of Figure 4. In the case of the figure 4, the anchor points for the interpolation are confused with the vertices of the harmonic peaks. In the case of FIG. 5, it is required that these anchor points be precisely at the frequencies that are multiples of the fundamental frequency, their amplitudes

corresponding to those of the peaks.

  The search interval of the amplitude maximum associated with a harmonic of rank k is centered on the index i of the frequency of the TFR closest to k.F0, that is to say i = 2Nk Fe + 2, where Laj denotes the integer equal to or immediately smaller than the number a. The width of this search interval depends on the sampling frequency Fe, the size 2N of the TFR and the range of possible variation of the fundamental frequency. This width is typically of the order of ten frequencies with the examples of values previously considered. It can be made adjustable according to the

  value F0 of the fundamental frequency and the number k of the harmonic.

  In order to improve the resolution in the low frequencies and thus to represent more accurately the amplitudes of the harmonics in this zone, a non-linear distortion of the frequency scale is operated on the upper envelope compressed by a module 12 before the module 13 performs the inverse fast Fourier transform (TFRI) providing the

Cepstral coefficients cxsup.

  Nonlinear distortion can more effectively minimize the modeling error. It is for example carried out according to a frequency scale of Mel or Bark type. This distortion may possibly depend on the estimated fundamental frequency F0. Figure 1 illustrates the case of the Mel scale. The relation between the frequencies F of the linear spectrum, expressed in hertz, and the frequencies F of the scale Mel is the following one: F 1000 oo F iogo (2) x log1 0 1+ (6) = og, o (2) x 1000) In order to limit the transmission rate, a truncation of the cepstral coefficients cx_sup is performed. The TFRI 13 module needs to calculate only a cepstral vector of NCS cepstral coefficients of orders 0 to

  NCS-1. For example, NCS can be equal to 16.

  A post-filtering in the cepstral domain, called post-liftrage, is applied by a module 15 to the compressed upper envelope LX_sup. This post-scaling corresponds to a manipulation of the cepstral coefficients cx_sup delivered by the TRFI module 13, which corresponds approximately to a post-filtering of the harmonic part of the signal by a transfer function having the classical form: H (z) = ( 1-.zl A (z / y1) (7) 1 (z) /1t-Az/2) o A (z) is the transfer function of a linear prediction filter of the audio signal, i1 and Y2 are coefficients between 0 and 1, and p. is a coefficient of pre-emphasis possibly zero. The relation between the post-lifted coefficient of order i, noted cp (i), and the corresponding cepstral coefficient c (i) = cxsup (i) delivered by the module 13 is then: cp (O) = c (o) i_ il) c (i) _ por i0 (8) Cp (i) = (1 + - Y1) C (i) i for i> 0 The optional pre-emphasis coefficient p. can be controlled by posing as a constraint to preserve the value of the cepstral coefficient cx_sup (1) relative to the slope. Indeed, the value c (1) = cxsup (1) of a white noise filtered by the pre-emphasis filter corresponds to the pre-emphasis coefficient. We can choose the latter as follows:

p = (/ 2-1) .c (1).

  After the post-lifter 15, a normalization module 16 modifies again

- 10 -

  the cepstral coefficients by imposing the exact modeling constraint of a point of the initial spectrum, which is preferably the most energetic point among the spectral maxima extracted by the module 9. In practice, this

  normalization only changes the value of the coefficient cp (0).

  The normalization module 16 operates in the following way: it recalculates a value of the spectrum synthesized at the frequency of the maximum indicated by the module 9, by Fourier transform of the truncated and post-lifted cepstral coefficients, taking into account the nonlinear distortion of the frequency axis; it determines a gain of normalization gN by the logarithmic difference between the value of the maximum provided by the module 9 and this

  recalculated value; and he adds the gain gN to the post-filtered cephalal coefficient cp (0).

  This standardization can be seen as part of post-layering.

  The post-lifted and normalized cepstral coefficients are quantified by a module 18 which transmits quantization indices

  corresponding icxs to the output multiplexer 6 of the encoder.

  The module 18 can operate by vector quantization from cepstral vectors formed of post-lifted and normalized coefficients, denoted here cx [n] for the n-rank signal frame. By way of example, the cepstral vector cx [n] of NCS = 16 cepstral coefficients cx [n, 0], cx [n, 1], ..., cx [n, NCS-1] is distributed in four sub-groups. cepstral vectors each containing four consecutive order coefficients. The cepstral vector cx [n] can be processed by the means represented in FIG. 6, forming part of the quantization module 18. These means implement, for each component cx [n, i], a predictor of the form: cxp [n, i] = (1 - o (i)). rcx [n, i] + ca (i). rcx [n-1, i] (9) where rcx [n] designates a residual prediction vector for the n-rank frame

  whose components are respectively denoted rcx [n, 0], rcx [n, 1], ...

  rcx [n, NCS-1], and ca (i) denotes a prediction coefficient chosen to be representative of a supposed inter-frame correlation. After quantization of the residues, this residual vector is defined by: rcx [n, icx [n, x [n i] - cl (i). rcxq [n-1, i] (10) 2 - cf (i) o rcx_q [n-1] denotes the quantized residual vector for the frame of rank nl, the components of which are respectively denoted rcxq [n, 0], rcxq [n, 1], -11 -

rcxq [n, NCS-1].

  The numerator of the relation (10) is obtained by a subtractor 20 whose components of the output vector are divided by the quantities 2-c (i) into 21. For the purpose of the quantification, the residual vector rcx [n] is subdivided into four subvectors, corresponding to the subdivision into four cepstral subvectors. On the basis of a dictionary obtained by prior learning, the unit 22 proceeds to the vector quantization of each sub-vector of the residual vector rcx [n]. This quantization can consist, for each sub-vector srcx [n], of selecting in the dictionary the quantized sub-vector srcxq [n] which minimizes the squared error srcx [n] -srcx_q [n] 2. The set icxs quantization index icx, corresponding to the addresses in the sub-vector dictionaries

  quantized residuals srcxq [n], is supplied to the output multiplexer 6.

  The unit 22 also delivers the values of the quantized residual sub-vectors, which form the vector rcxq [n]. This latter is delayed by one frame at 23, and its components are multiplied by the coefficients ca (i) at 24 to supply the vector at the negative input of the subtractor 20. This latter vector is, on the other hand, supplied to a adder 25, whose other input receives a vector formed by the components of the quantized residual rcxq [n] respectively multiplied by the quantities 1-c (i) at 26. The adder 25 thus delivers the

  cepstral vector quantified cx_q [n] that will recover the decoder.

  The prediction coefficient c (i) can be optimized separately for each of the cepstral coefficients. Quantization dictionaries can

  also be optimized separately for each four cepstral subvectors.

  On the other hand, it is possible, in a manner known per se, to normalize the cepstral vectors before applying the prediction / quantization scheme, starting from

the variance of the cepstres.

  It should be noted that the above scheme of quantization of cepstral coefficients may be applied only for some frames only. For example, a second quantization mode can be provided, as well as a selection process for the one of the two modes, which minimizes a least squares criterion with the cepstral coefficients to be quantized, and transmits with the quantization indices of the frame a bit indicating which one

modes has been selected.

  The quantified cepstral coefficients cxsup_q = cxq [n] supplied by -12Iadditionneur 25 are addressed to a module 28 which recalculates the spectral amplitudes associated with one or more of the harmonics of the fundamental frequency F0 (FIG. 1). These spectral amplitudes are for example calculated in compressed form, by applying the Fourier transform to the quasispersed coefficients, taking into account the non-linear distortion of the frequency scale used in the cepstral transformation. The amplitudes thus recalculated are supplied to an adaptation module 29 which compares them with amplitudes of maxima determined by the extraction module 9. The adaptation module 29 controls the post-lifter 15 so as to minimize a module gap. between the spectrum of the audio signal and the corresponding module values calculated at 28. This module difference can be expressed as a sum of absolute values of amplitude differences, compressed or not, corresponding to one or more of the harmonic frequencies. This sum can be weighted according to the amplitudes

  spectrum associated with these frequencies.

  Optimally, the module gap taken into account in the adaptation

  Post-Lifting would take into account all the harmonics of the spectrum.

  However, in order to reduce the complexity of the optimization, the module 28 can resynthesize the spectral amplitudes only for one or more multiple frequencies of the fundamental frequency F 0, selected on the basis of the magnitude of the spectrum modulus in absolute value. The adaptation module 29 can for example consider the three most spectral peaks

  intense in calculating the module deviation to be minimized.

  In another embodiment, the adaptation module 29 estimates a spectral masking curve of the audio signal by means of a psychoacoustic model, and the frequencies taken into account in the calculation of the module deviation to be minimized are selected on the basis of the importance of the modulus of the spectrum relative to the masking curve (for example, the three frequencies for which the spectrum modulus is greater than the masking curve) can be taken. Different classical methods

  can be used to calculate the masking curve from the audio signal.

  One can for example use the one developed by J.D. Johnston ("Transform Coding of Audio Signals Using Perceptual Noise Criteria", IEEE Journal on

  Selected Area in Communications, Vol. 6, No. 2, February 1988).

  In order to adapt the post-lifter, the module 29 can use a filter identification model. A simpler method consists in predefining a set of post-scaling parameter sets, that is to say a set of pairs y1, 12 in the case of post-scaling according to the relations (8), to be performed the operations incumbent on the modules 15, 16, 18 and 28 for each of these sets of parameters, and to retain that of the sets of parameters which leads to the minimum module gap between the signal spectrum and the recalculated values. The quantization indices provided by the module 18 are then

  those that relate to the best set of parameters.

  By a process analogous to that of extracting the coefficients cx_sup representing the compressed upper envelope LXsup of the signal spectrum, the coder determines coefficients cx_inf representing a compressed lower envelope LX_inf. A module 30 extracts from the compressed spectrum LX spectral amplitudes associated with frequencies located in areas of the intermediate spectrum with respect to the frequency multiples of the frequency

fundamental estimated F0.

  In the example illustrated in FIGS. 4 and 5, each amplitude associated with a frequency located in an intermediate zone between two successive harmonics k.F0 and (k + 1) .F0 simply corresponds to the frequency modulus for the frequency (k + 1/2) .F0 located in the middle of the interval separating the two harmonics. In another embodiment, this amplitude could be an average of the spectrum modulus on a small beach surrounding this

frequency (k + 112) .F0.

  A module 31 interpolates, for example linearly, the spectral amplitudes associated with the frequencies located in the zones

  intermediates to obtain the compressed lower envelope LX_inf.

  The cepstral transformation applied to this compressed lower envelope LX_inf is performed according to a frequency scale resulting from a nonlinear distortion applied by a module 32. The TFRI module 33 calculates a cepstral vector of NCI cepstral coefficients cx_inf of orders 0 to NCI-1 representing the lower envelope. NCI is a number that can be

  significantly smaller than NCS, for example NCI = 4.

  The non-linear transformation of the frequency scale for the cepstral transformation of the lower envelope can be performed to a finer scale at the higher frequencies than at the lower frequencies, which

- 14 -

  advantageously makes it possible to model the unvoiced components of the signal at high frequencies. However, to ensure homogeneity of representation between the upper envelope and the lower envelope, it may be preferable to adopt in the module 32 the same scale as in the module 12 (Mel in the example). The cepstral coefficients cx_inf representing the compressed lower envelope are quantized by a module 34, which can operate in the same way as the module 18 for quantizing the cepstral coefficients representing the compressed upper envelope. In the case considered, where we limit ourselves to NCI = 4 cepstral coefficients for the lower envelope, the vector thus formed is subjected to vector quantization of prediction residue, performed by means identical to those shown in FIG. subdivision into sub-vectors. The quantization index icx = icxi determined by the vector quantizer 22 for each frame relative to the

  coefficients cxinf is supplied to the output multiplexer 6 of the encoder.

  The encoder shown in FIG. 1 does not include any device

  particular for coding the phases of the spectrum to the harmonics of the audio signal.

  On the other hand, it comprises means 36-40 for encoding a temporal information related to the phase of the non-harmonic component.

  represented by the lower envelope.

  A spectral decompression module 36 and a TFRI module 37

  form a temporal estimate of the non-component

  harmonic. The module 36 applies a reciprocal decompression function of the compression function applied by the module 8 (that is to say an exponential or a power function 1ly) to the compressed lower envelope LXinf produced by the interpolation module 31. This provides the modulus of the estimated frame of the non-harmonic component, whose phase is taken equal to that (Px of the spectrum of the signal X on the frame.) The inverse Fourier transform performed by the module 37 provides the estimated frame of the

non-harmonic component.

  The module 38 subdivides this estimated frame of the non-component

  harmonic in several time segments. The frame delivered by the module 37 consisting of 2N = 512 weighted samples as illustrated by FIG. 3, the module 38 considers only the N / 2 = 128 first samples and the N / 2 = 128 last samples, and subdivides them, for example in eight segments of 32 consecutive samples each representing 4 ms of signal. For each segment, the module 38 calculates the energy equal to the sum of the squares of the samples, and forms a vector El formed of eight positive real components equal to the eight calculated energies. The largest of these eight energies, denoted EM, is also determined to be provided, with the vector E1, to a normalization module 39. This module divides each component of the vector E1 by EM, so that the standardized vector Emix is composed of eight components between 0 and 1. It is this normalized vector Emix, or weighting vector, which is subjected to the quantization by the module 40. This can perform a vector quantization with a dictionary determined during a test. prior learning. The quantization index iEm is provided by the module 40 to the output multiplexer 6 of the encoder. FIG. 7 shows an alternative embodiment of the means employed by the coder of FIG. 1 for determining the energy-weighting Emix vector of the frame of the non-harmonic component. The spectral decompression modules 36, 37 and TFRI operate as those having the same references in FIG. 1. A selection module 42 is added to determine the value of the spectrum module subjected to the inverse Fourier transform 37. Based on the estimated fundamental frequency FO, the module 42 identifies harmonic regions and non-harmonic regions of the audio signal spectrum. For example, a frequency will be considered as belonging to a harmonic region if it is in a frequency interval centered on a harmonic k.F0 and width

  corresponding to a synthesized spectral line width, and a non-region

  harmonic otherwise. In the non-harmonic regions, the complex signal subjected to the TFRI 37 is equal to the value of the spectrum, that is to say that its modulus and its phase correspond to the values | X | and (px provided by the TFR module 3. In the harmonic regions, this complex signal has the same phase (Px as the spectrum and a modulus given by the lower envelope after spectral decompression 36). This procedure according to FIG. 7 provides

  more accurate modeling of non-harmonic regions.

  The decoder shown in FIG. 8 comprises an input demultiplexer 45 which extracts from the bit stream (D, coming from an encoder according to FIG. 1) the indexes iF, icxs, icxi, iEm of quantization of the fundamental frequency FO, cepstral coefficients representing the compressed upper envelope, coefficients representative of the compressed lower envelope, and the Emix weighting vector, and distributes them respectively to modules 46, 47, 48 and 49. These modules 46-49 comprise Quantization dictionaries similar to those of the modules 5, 18, 34 and 40 of Figure 1, in order to restore the values of the quantized parameters The modules 47 and 48 have dictionaries to form the quantized prediction residuals rcxq [n], and they deduce the quantified cepstral vectors cxq [n] with elements identical to the elements 23-26 of FIG. 6. These quasified cepstral vectors cx_q [n] provide the cepstral coefficients cxsup_q and cx_inf_q

processed by the decoder.

  A module 51 calculates the fast Fourier transform of the cepstral coefficients cx_sup for each signal frame. The frequency scale of the resulting compressed spectrum is modified non-linearly by a module 52 applying the reciprocal non-linear transformation of that of the module 12 of FIG. 1, and which provides the LX_sup estimate of the compressed upper envelope. . A spectral decompression of LXsup, operated by a module 53, provides the upper envelope Xsup containing the estimated values of the

  multi-frequency spectrum module of the fundamental frequency F0.

  The module 54 synthesizes the spectral estimation Xv of the harmonic component of the audio signal, by a sum of spectral lines centered on the multiple frequencies of the fundamental frequency F0 and whose

  Amplitudes (in modulus) are those given by the upper envelope X_sup.

  Although the digital input stream O does not have specific information on the phase of the signal spectrum at the harmonics of the fundamental frequency, the decoder of FIG. 8 is able to extract information on this phase from cepstral coefficients cx_sup_q representing the compressed upper envelope. This phase information is used to assign a phase (p (k) to each of the spectral lines determined by the

  module 54 in the estimation of the harmonic component of the signal.

  As a first approximation, the speech signal can be considered to be at minimum phase. On the other hand, it is known that minimal phase information can easily be deduced from cepstral modeling. This minimum phase information is therefore calculated for each harmonic frequency. The minimum phase hypothesis means that the energy of the synthesized signal is located at the beginning of each period of the frequency

fundamental F0.

  To be closer to a real speech signal, a little dispersion is introduced by means of a specific post-ligation of the cepstres during the synthesis of the phase. With this post-stretching, carried out by the module 55 of FIG. 8, it is possible to accentuate the formanic resonances of the envelope and thus to control the dispersion of the phases. This postlifter is for example

of the form (8).

  To limit the phase breaks, it is preferable to smooth the post-lifted cepstral coefficients, which is done by the module 56. The module 57 deduces from the post-lifted cepstral coefficients and smoothed the minimum phase assigned to each spectral line representing a harmonic peak of the spectrum. The operations carried out by the minimum phase smoothing and extraction modules 56, 57 are illustrated by the flowchart of FIG. 9. The module 56 examines the variations of the cepstral coefficients to apply a smaller smoothing in the presence of sudden variations. only in the presence of slow variations. For this, it performs the smoothing of the cepstral coefficients by means of a forgetting factor Xc chosen according to a comparison between a threshold dth and a distance d between two successive sets of post-lifted cepstral coefficients. The threshold dth is itself adapted according to

  variations of the cepstral coefficients.

  The first step 60 consists in calculating the distance d between the two successive vectors relative to the n-1 and n frames. These vectors, noted here cxp [n-1] and cxp [n], correspond for each frame to the set of NCS post-lifted cepstral coefficients representing the compressed upper envelope. The distance used may notably be the Euclidean distance

  between the two vectors or a quadratic distance.

  Two smoothings are first performed, respectively by means of forgetting factors Xmin and Xmax, to determine a minimum distance dmin and a maximum distance dmax. The threshold dth is then determined in step 70 as being situated between the minimum and maximum distances dmin, dmax:

  dth = .dmax + (1-3) .dmin, the coefficient f being for example equal to 0.5.

-18 -

  In the example shown, the forgetting factors Xmin and Xmax are themselves

  same selected from two distinct values, respectively; min2 and Xmax1, Xmax2 between 0 and 1, the indices Xmini, Xmax1 being each substantially closer to 0 than the indices Xrmin2, Xmax2 'If d> dmin (test 61), the forgetting factor? min is equal to minl (step 62); otherwise it is taken equal to kmin2 (step 63). In step 64, the minimum distance dmin is taken equal to Xmin.dmin + (1-Xmin) .d. If d> d max (test 65), the forgetting factor, max is equal to X max (step 66); otherwise it is taken equal to Xmax2 (step 67). At step 68, the

  minimum distance dmax is taken equal to Xmax dmax + (1-Xmax) .d.

  If the distance d between the two consecutive cepstral vectors is greater than the threshold dth (test 71), the forgetting factor Xc is taken to have a value Xc1 relatively close to 0 (step 72). It is considered in this case that the corresponding signal is of the non-stationary type, so that it is not necessary to keep a large memory of the previous cepstral coefficients. If d <dth, we adopt in step 73 for the forgetting factor Xc a value Xc2 less close to 0 in order to further smooth the cepstral coefficients. The smoothing is performed at step 74, where the vector cx1 [n] of smoothed coefficients for the current frame n is determined by: cx1 [n] = Xc.cxl [n-1] + (1-Xc) cxp [ n] (11) The module 57 then calculates the minimum phases p (k) associated with the harmonics k.Fo. In a known way, the minimal phase for a harmonic of order k is given by: NCS-1 p (k) = - 2., Cxl [n, m] .sin (2nmkFo / Fe) (12) m = 1

  o cxl [n, m] denotes the smoothed cepstral coefficient of order m for the n-frame.

  In step 75, the harmonic index k is initialized to 1. To initialize the calculation of the minimum phase assigned to the harmonic k, the phase p (k) and the cepstral index m are initialized respectively at 0 and 1 at step 76. At step 77, the

  module 57 adds to the phase p (k) the quantity -2.cxl [n, m] .sin (2 = mk.Fo / Fe).

  The cepstral index m is incremented in step 78 and compared to NCS in step 79.

  Steps 77 and 78 are repeated as long as NCS. When m = NCS, the calculation of the minimum phase is completed for the harmonic k, and the index k is incremented in step 80. The calculation of minimum phases 76-79 is repeated

  for the next harmonic as k.F0 <Fe / 2 (test 81).

  In the exemplary embodiment according to FIG. 8, the module 54 takes into account a constant phase over the width of each spectral line, equal to the minimum phase (p (k) supplied for the corresponding harmonic k by the

module 57.

  The estimate Xv of the harmonic component is synthesized by summation of spectral lines positioned at the harmonic frequencies of the fundamental frequency F0. During this synthesis, it is possible to position the spectral lines on the frequency axis with a resolution greater than the resolution of the Fourier transform. For this, a reference spectral line is precalculated once and for all according to the higher resolution. This calculation can consist of a Fourier transform of the analysis window fA with a transform size of 16384 points, providing a resolution of 0.5 Hz per point. The synthesis of each harmonic line is then performed by the module 54 by positioning the high resolution reference line on the frequency axis, and by sub-sampling this reference spectral line to reduce to the 16.625 Hz resolution of the Fourier transform on

  512 points. This makes it possible to accurately position the spectral line.

  For the determination of the lower envelope, the TFR module 85 of the decoder of FIG. 8 receives the quantized cepstral coefficients cx_infq of orders 0 to NCI-1, and it advantageously completes them by the NCS-NCI coefficients cepstral cxsup_q d NCI order to NCS - 1 representing the upper envelope. In fact, it can be estimated as a first approximation that the rapid variations of the compressed lower envelope are well reproduced by those of the compressed upper envelope. In another embodiment, the TFR 85 module could consider only the NCIs

cepstral parameters cxinfq.

  The module 86 converts the frequency scale reciprocally from the conversion operated by the module 32 of the encoder, in order to restore the estimate LX_inf of the compressed lower envelope, submitted to the spectral decompression module 87. At the output of the module 87 , the decoder has a lower envelope X_inf comprising the values of the spectrum module -20 -

  in the valleys between the harmonic peaks.

  This envelope X_inf will modulate the spectrum of a noise frame whose phase is processed according to the quantified weighting vector Emix extracted by the module 49. A generator 88 delivers a normalized noise frame whose segments of 4 ms are weighted in a module 89 according to the standardized components of the Emix vector provided by the module 49 for the current frame. This noise is a high-pass filtered white noise to account for the low level that normally has the unvoiced component at low frequencies. From the energy-weighted noise, the module 90 forms frames of 2N = 512 samples by applying the analysis window fA, the insertion of 256 samples at zero and the circular permutation for the phase compensation according to what has been done. This is explained with reference to FIG. 3. The Fourier transform of the resulting frame is calculated by the

TFR module 91.

  The spectral estimate Xuv of the non-harmonic component is determined by the spectral synthesis module 92 which performs frequency-frequency weighting. This weighting consists of multiplying each complex spectral value provided by the TFR module 91 by the value of the lower envelope X_inf obtained for the same frequency by the

  spectral decompression module 87.

  The spectral estimates Xv, Xuv of the harmonic (voiced in the case of a speech signal) and non-harmonic (or non-voiced) components are combined by a mixing module 95 controlled by a module 96

  analyzing the degree of harmonicity (or voicing) of the signal.

  The organization of these modules 95, 96 is illustrated in FIG. 10. The analysis module 96 comprises a unit 97 for estimating a degree of voicing W depending on the frequency, from which four gains depending on the frequency are calculated. the frequency, namely two gains gv, guv controlling the relative importance of the harmonic and nonharmonic components in the synthesized signal, and two gains gv -p guv_, used to noise the phase of the

harmonic component.

  The degree of voicing W (i) is a continuously varying value between 0 and 1 determined for each frequency index i (0 <i <N) as a function of the upper envelope Xsup (i) and the lower envelope Xinf (i) obtained for this frequency i by the decompression modules 53,

- 21 -

  87. The degree of neighbor W (i) is estimated by unit 97 for each index of frequency i corresponding to a harmonic of fundamental frequency F0, namely i = 2NkF + 2A for k = 1,2, ... by an increasing function of the ratio between the upper envelope Xsup and the lower envelope X_inf at this frequency, for example according to the formula: W (i) = min {1, 10 og1 0 [x-sup (i) / Xinf (i)] (13) Vt-h- (Fo) -. The threshold Vth (F0) corresponds to the average dynamic calculated on a synthetic spectrum purely voiced at the fundamental frequency. It is

  advantageously chosen depending on the fundamental frequency F0.

  The degree of voicing W (i) for a frequency other than the harmonic frequencies is obtained simply as being equal to that

  estimated for the nearest harmonic.

  The gain gv (i), which depends on the frequency, is obtained by applying a

  non-linear function at the degree of voicing W (i) (block 98). This non-function

  linear, for example, has the form shown in FIG. 11i gv (i) = 0 if 0 <W (i) W1 gV (i) - W (i) W1 if W1 <W (i) <W2 (14)

W2 - W1

  gv (i) = 1 if W2 <W (i) <1 the thresholds W1, W2 being such that 0 <W1 <W2 <1. The guv gain can be calculated in a similar way to the gain gv (the sum of the two gains gv , guv being constant, for example equal to 1), or simply deduced from it by the relation guv (i) = 1 - gv (i), as schematized by the subtractor 99 on the

figure 10.

  It is interesting to be able to noise the phase of the harmonic component of the signal at a given frequency if the analysis of the degree of voicing shows that the signal is rather non-harmonic type at this frequency. For this, the phase q (pv of the mixed harmonic component is the result of a linear combination of the phases pv, (puv of the components

  harmonic and non-harmonic Xv, Xuv synthesized by the modules 54, 92.

  The gains gvqp, guv _q respectively applied to these phases are calculated from the degree of voicing W and weighted also according to the frequency index i, since the sounding of the phase is not really true.

  useful only beyond a certain frequency.

  A first gain gv1_, p is calculated by applying a nonlinear function to the degree of voicing W (i), as shown schematically by block 100 in FIG. 10. This non-linear function can have the form represented in FIG. 12: gvlq ( i) = G1 if 0 <W (i) <W3 gv1_c (i) = Gl1 + (1-G1) W (i) W3 siW3 <W (i) <W4 (15)

W4 -W3

  gv1cp (i) = 1 if W4 <W (i) <1 the thresholds W3 and W4 being such that 0 <W3 <W4 <1, and the minimum gain G1

being between 0 and 1.

  A multiplier 101 multiplies for each index frequency i the gain gv1 by another gain gv2 p depending only on the frequency index i, to form the gain gvqP (i). The gain gv2_, p (i) depends non-linearly on the frequency index i, for example as indicated in FIG. 13 gv2 (i) = 1 if 0 <i <il gv2 q (i) = 1- (1) -G2) j it if it <i <i2 (16) i2_- -i1 it gv2_p (i) = G2 if i2 <i <1 the indexes he and i2 being such that 0 <il <i2 <N, and the gain with G2 being between 0 and 1. The gain guv _p (i) can be simply calculated as 1 - gv _, (i) = 1 gvl_ (p (i) .gv2_qp (i) (subtracter 102 of the figure The complex spectrum Y of the synthesized signal is produced by the mixing module 95, which carries out the following mixing relation, for 0 <i <N: Y (i) = gv (i) i). XMi). exp [P'v (i)] + guv (i). Xuv (i) (17) with qPV (i) = gv_ (i). (Pv (i) + guv _ (p (i). (Puv (i) (18) where pv (i) designates the argument of the complex number Xv (i) provided by the module 54 for the index frequency i (block 104 of FIG. 10), and (puv (i) designates

- 23 -

  The argument of the complex number Xuv (i) provided by the module 92 (block 105 of FIG. 10). This combination is achieved by the multipliers 106110 and the

  adders 111 -112 shown in FIG.

  The mixed spectrum Y (i) for 0 i <2N (with Y (2N-1-i) = Y (i)) is then transformed in the time domain by the TFRI module 115 (FIG. 8). Only the first N / 2 = 128 and the N / 2 = last 128 samples of the frame of 2N = 512 samples produced by the module 115 are retained, and the inverse circular permutation of that illustrated in FIG. the

  synthesized frame of N = 256 samples weighted by the analysis window fA.

  The frames successively obtained in this manner are finally processed by the temporal synthesis module 116 which forms the audio signal

decoded x.

  The time synthesis module 116 performs an overlapping sum of frames modified with respect to those successively evaluated at the output of the module 115. The modification can be seen in two

  steps illustrated respectively in Figures 14 and 15.

  The first step (FIG. 14) consists in multiplying each frame 2 'delivered by the TFRI module 115 by a window 1 / fA inverse to the analysis window fA used by the module 1 of the encoder. The samples of the frame

  2 "that result are therefore weighted uniformly.

  The second step (FIG. 15) consists of multiplying the samples of this frame 2 "by a synthesis window fs satisfying the following properties: fs (N-L + i) + fs (i) = A for 0 <i <L ( 19) fs (i) = A for L <i <NL (20) where A denotes an arbitrary positive constant, for example A = 1. The synthesis window fs (i) increases progressively from 0 to A for i ranging from 0 to L. It is for example a raised half-sine A fs (i) = A. (1-cos [(i + 1/2) n / LD for0 <i <L (21) After reweighting each frame 2 by the synthesis window fs, the module 116 positions the successive frames with their time offsets of M = 160 samples and their time overlaps of L = 96 samples, then it performs the sum of the frames thus positioned in time. properties (19) and (20) of the synthesis window fs, each sample of the decoded audio signal x thus obtained is assigned a uniform overall weight, equal to A. This overall weight comes from the contribution of a single frame if the sample has in this frame a rank i such that L <i <N - L, and comprises the summed contributions of two successive frames if 0 <<<L <N <i <N. It is thus possible to perform the temporal synthesis in a simple manner even if, as in the case considered, the overlap L between two frames

  successive is smaller than half the size N of these frames.

  The two steps outlined above for modifying the signal frames can be merged in one step. It suffices to precompute a composite window fc (i) = fs (i) / fA (i), and to simply multiply the frames 2 'of N = 256 samples delivered by the module 115 by the composite window

  fc before performing the overlay summation.

  FIG. 16 shows the appearance of the composite window fc in the case where the analysis window fA is a Hamming window and the synthesis window fs

  the form given by relations (19) to (21).

  Other forms of the synthesis window fs satisfying relationships (19) and (20) can be employed. In the variant of FIG. 17, it is a piecewise affine function defined by: fs (i) = Ai / L for 0 <i <L (22) In order to improve the coding quality of the audio signal, the coder of FIG. 1 can increase the rate of frame formation and analysis, in order to transmit more quantization parameters to the decoder. In the frame structure shown in Fig. 2, a frame of N = 256 samples (32 ms) is formed every 20 ms. These frames of 256 samples could be formed at a higher rate, for example 10 ms, two successive frames then having an offset of M / 2 = 80

  samples and an overlay of 176 samples. In these conditions, it is possible to transmit the complete sets of

  quantization parameters iF, icxs, icxi, iEm for only one sub-

  set of frames, and transmit for the other frames parameters to perform adequate interpolation at the decoder. In the example envisaged above, the subset for which complete parameter sets are transmitted can be constituted by the rank frames

- 25 -

  integer n, whose periodicity is M / Fe = 20 ms, and the frames for which interpolation is performed can be those of half-integer n + 1/2

  which are shifted by 10 ms from the frames of the subassembly.

  In the embodiment illustrated in FIG. 18, the notations cxq [n-1] and cxq [n] denote determined quasified cepstral vectors, for two successive frames of entire rank, by the quantization module 18 and / or by the module of Quantization 34. These vectors comprise for example four consecutive cepstral coefficients each. They could also

  understand more cepstral coefficients.

  A module 120 interpolates these two cepstral vectors cxq [n-1] and cxq [n], in order to estimate an intermediate value cx_i [n-1/2]. The interpolation performed by the module 120 can be a simple arithmetic mean of the vectors cxq [n-1] and cxq [n]. As a variant, the module 120 could apply a more sophisticated interpolation formula, for example polynomial, also based on the cepstral vectors obtained for frames prior to the n-1 frame. On the other hand, if more than one interpolated frame is interposed between two consecutive frames of full rank,

  Interpolation takes into account the relative position of each interpolated frame.

  Using the means previously described, the coder calculates

  also the cepstral coefficients cx [n-1/2] relating to the half-rank weft

  whole. In the case of the upper envelope, these cepstral coefficients are those provided by the TFRI module 13 after post-scaling (for example with the same post-scaling coefficients as for the previous frame n-1) and normalization 16. In the case of the lower envelope, the coefficients

  cepstral cx [n-1/2] are those delivered by the TFRI 33 module.

  A subtractor 121 forms the difference ecx [n-1/2] between the cepstral coefficients cx [n-1/2] calculated for the half-integer field and the coefficients cxi [n-1/2] estimated by interpolation. This difference is provided to a quantization module 122 which addresses quantization indices icx [n-1/2] to the output multiplexer 6 of the encoder. The module 122 operates for example by vector quantization interpolation errors

* ecx [n-1/2] successively determined for frames of half-integer rank.

  This quantization of the interpolation error can be carried out by the coder for each of the NCS + NCI coefficients used by the decoder, or only for some of them, typically those of orders

- 26 -

the little ones.

  The corresponding means of the decoder are illustrated in FIG. 19.

  The decoder operates essentially as described with reference to FIG. 8 to determine the full rank signal frames. An interpolation module 124 identical to the encoder module 120 estimates the intermediate coefficients cxi [n-1/2] from the quantized coefficients cxq [n-1] and cx_q [n] provided by the module 47 and / or the module 48 from the indexes icxs, icxi extracted from the stream t. A parameter extraction module 125 receives the quantization index icx [n-1/2] from the input demultiplexer 45 of the decoder, and deduces therefrom the quantized interpolation error ecxq [n-1/2] from the same quantization dictionary as used by the encoder module 122. An adder 126 sums the cepstral vectors cx_i [n-1/2] and ecx_q [n-1/2] in order to provide the cepstral coefficients cx [n-1/2] which will be used by the decoder (modules 51- 57, 95, 96, 115 and / or modules 85-87, 92,

  95, 96, 115) to form the interpolated frame of rank n-1/2.

  If only some of the cepstral coefficients have been quantized for interpolation error, the others are determined by the

  decoder by simple interpolation, without correction.

  The decoder can also interpolate the other parameters F0, Emix used to synthesize the signal frames. The fundamental frequency F0 can be interpolated linearly, either in the time domain or (preferably) directly in the frequency domain. For the possible interpolation of the Emix energy weighting vector, the interpolation after denormalization and taking into account

time delays between frames.

  It should be noted that it is particularly advantageous, to interpolate the representation of the spectral envelopes, to perform this interpolation in the cepstral domain. Unlike interpolation performed on other parameters, such as LSP ("Line Spectrum Pairs") coefficients, the linear interpolation of the cepstral coefficients corresponds to the interpolation

  linear spectral amplitudes compressed.

  In the variant shown in FIG. 20, the coder uses the cepstral vectors cxq [n], cxq [n-1], ..., cxq [nr] and cxq [n-1/2] calculated for the last past frames. (r> 1) to identify an optimal interpolator filter which, when subjected to quantified cepstral vectors cxq [nr],

- 27 -

  cxq [n] relative to the frames of full rank, delivers an interpolated cepstral vector cxi [n-1/2] which has a minimal distance with the vector

  cx [n-1/2] calculated for the last frame of half-integer rank.

  In the example shown in FIG. 20, this interpolator filter 128 is present in the encoder, and a subtractor 129 subtracts its output cxi [n-1/2] from the calculated cepstral vector cx [n-1/2]. A minimization module determines the parameter set {P} of the interpolator filter 128, for which the ecx interpolation error [n-1/2] delivered by the subtractor 129 has a minimum standard. This parameter set {P} is addressed to a quantization module 131 which provides a quantization index corresponding to iP to

output multiplexer 6 of the encoder.

  As a function of the flow rate allocated in the stream Oe to the quantization indexes of the parameters {P} defining the optimal interpolator filter 128, it will be possible to adopt a more or less fine quantization of these parameters, or a more or less elaborate form of the interpolator filter, or still provide several interpolator filters quantized separately for different vectors of

cepstral coefficients.

  In a simple embodiment, the interpolator filter 128 is linear, with r = 1: cx i [n-1/2] = p.cx_q [n-1] + (1-p) .cx_q [n] (23)

  and the parameter set {P} is limited to the coefficient p between 0 and 1.

  From the quantization index P of the parameters {P} obtained in the bit stream (p, the decoder reconstructs the interpolator filter 128 (with quantization errors close), and processes the spectral vectors cxq [nr], ..., cxq [n] in order to estimate the cextral coefficients cx [n-1/2] used to synthesize the

frames of half-whole rank.

  In general, the decoder can use a simple interpolation method (without transmission of parameters by the coder for half-integer frames), an interpolation method taking into account an interpolation error. quantized (according to FIGS. 17 and 18), or an interpolation method with an optimal interpolator filter (according to FIG. 19) for evaluating the half-integer rank frames in addition to the rank frames

  integer evaluated directly as explained with reference to FIGS. 8-13.

  The temporal synthesis module 116 can then combine all of these

  frames evaluated to form the synthesized signal x in the manner explained above.

-28 -

  after referring to Figures 14, 21 and 22.

  As in the temporal synthesis method previously described, the module 116 performs an overlap sum of frames modified with respect to those successively evaluated at the output of the module 115, and this modification can be seen in two steps, the first of which is identical to that previously described with reference to Figure 14 (divide

  samples of the 2 'frame by the analysis window fA).

  The second step (FIG. 21) consists in multiplying the samples of the renormalized frame 2 "by a synthesis window fs satisfying the following properties fs (i) = 0 for 0 <i <N / 2 - M / p and N / 2 + M / p <i <N (24)

I I

  fs (i) + fs (i + M / p) = A for N / 2 - M / p <i <N / 2 (25) where A is an arbitrary positive constant, for example A = 1, and p is l integer such that the time offset between the successive frames (calculated directly and interpolated) is M / p samples, or p = 2 in the example described. The synthesis window fs (i) gradually increases for i ranging from N / 2 - M / p to N / 2. It is for example a sinusoid raised on the interval N / 2 - M / p <i <N / 2 + M / p. In particular, the synthesis window fs may be, over this interval, a Hamming window (as shown in FIG. 21)

or a Hanning window.

  FIG. 21 shows the successive frames 2 "repositioned in time by the module 116. The hatches indicate the portions removed from the frames (synthesis window at 0. It can be seen that by performing the overlap sum of the samples of the successive frames, the property (25)

  ensures a homogeneous weighting of the samples of the synthesized signal.

  As in the synthesis method illustrated in FIGS. 14 and 15, the weighting procedure of the Fourier transform inverse fields of the Y spectra can be carried out in a single step, with a composite window fc (i) = fs (i) / fA (i). Figure 22 shows the shape of the window

  composed fc in the case where the windows fA and fs are of type Hamming.

  As the temporal synthesis method illustrated in FIGS. 14 to 17, that illustrated by FIGS. 14, 21 and 22 makes it possible to take into account

- 29 -

  an overlap L between two analysis frames (for which the analysis is performed completely) less than half that size N of these frames. In general, the latter method is applicable when the successive analysis frames have mutual time offsets M of more than N / 2 samples (even possibly more than N samples if a very low flow rate is required), the interpolation leading to to a set of frames whose mutual time offsets are less than N / 2 samples. Interpolated frames may be subject to reduced transmission of coding parameters, as previously described, but this is not required. This embodiment makes it possible to maintain a relatively large interval M between two analysis frames, and thus to limit the required transmission rate, while limiting the discontinuities that may appear due to the size of this interval relative to the scales. typical of the variations in the parameters of the audio signal, especially the

  cepstral coefficients and the fundamental frequency.

  FIGS. 23 to 25 show other embodiments of the means used to process the cepstral coefficients cxsup delivered by the

  TFRI module 13 of Figure 1, representing the upper envelope.

  In all three cases, the post-lifter 15, normalization 16, quantization 18 and spectral amplitude calculation modules 28 are essentially identical to those previously described with reference to FIG. 1. In addition, post-scale modules 140, smoothing 141 and minimum phase extraction 142 are provided to process the post-lifted and quasified cpsup_q cepstral coefficients delivered by the quantization module 18. These modules 140-142 essentially operate

  as the corresponding modules 55-57 of the decoder of FIG. 8.

  In the embodiment shown in FIG. 23, the adaptation module 144 performs a function similar to that of the module 29 of FIG. 1. However, the adaptation is not carried out solely on the basis of the module of FIG.

  spectrum. The module 144 determines the best set of coefficients for the post-

  by minimizing the difference between the spectrum of the audio signal, in IXI module and Px phase, and complex values recalculated for one or more of the harmonics of the fundamental frequency. The modules of these latter complex values are given by the calculation module 28, and their phases -30 - correspond to the minimum phases p (k) provided by the extraction module 142. To achieve the adaptation, the module 144 can take into account any appropriate distance in the complex plane, for example the Euclidean distance. Thus, the adaptation of the post-lifter 15 by the module 144 takes into account, in a combined manner, frequency aspects of the signal, reflected by the module of the

  spectrum, and temporal aspects, reflected by the phase of the spectrum.

  As shown in dashed lines in FIG. 23, the post-lifter 140 can also be adaptive, the adaptation made by the module 144 bearing jointly on the two post-lifters 15, 140. In this case, the post-lifter 55 of the decoder (Figure 8) is adapted, as the post-lifter 140, according to iLIF parameters that the adaptation module 144 provides the multiplexer 6 for it to include in the digital stream c). Typically, a few sets of coefficients y1, Y2 are provided for the post-lifters 140 and 55, and the module 144 performs a comprehensive test of these different sets of coefficients to retain

  the one that minimizes the gap in the complex plane.

  In the example shown in FIG. 24, the adaptation module 29 of the post-lifter 15 is identical to that of FIG. 1. FIG. 24 shows a module 145 for estimating a masking curve allowing the module 29 to select, for the minimization of the difference in modulus, the harmonic frequency or frequencies which exceed the most of the masking curve calculated on the

  base of the spectrum in modulus | X |, as previously described.

  The post-lifter 140 of FIG. 24 is separately adapted by a module 146 which minimizes the differences between the phase (pX of the signal spectrum and the minimum phase rp (k) calculated by the module 142 for one or more of the harmonics Here again, the harmonics selected for the calculation of the minimized phase difference can be as a function of the masking curve estimated by the module 145. The module 146 supplies the output multiplexer 6 of the encoder with the parameters iLif which represent the optimall post-lifter

  140, for use in the post-filter 55 of the decoder.

  In the example illustrated in FIG. 25, the post-lifter 140 used in the calculation of the minimum phases is not adaptive. The minimum phases p (k) calculated by the module 142 for the harmonics of the fundamental frequency are compared with the phases (Px of the spectrum of the audio signal, and

- 31 -

  The phase difference is quantized by a module 148. The corresponding quantization indices iA (p are provided by the module 148 at

output multiplexer 6 of the encoder.

  In a decoder (FIG. 26) corresponding to an encoder according to FIG. 25, a module 149 exploits these quantization indices iA (p provided by the demultiplexer 45 to obtain the quantized phase difference values, which an adder 150 adds to the phases (p (k) calculated by the module 57 (the post-lifters 140 and 55 being identical) The phases provided by the adder 150 are then used by the module 54 which synthesizes the

  spectral lines of the harmonic component Xv.

  The quantized phase difference by the module 148, and that the decoder modules 149 and 150 use to correct the minimum phases p (k), can be of two types: it can represent, for each corresponding index frequency i at a harmonic of order k of the fundamental frequency F0 the difference between the phase px (i) of the spectrum of the signal at the frequency i and the minimum phase (p (k) calculated by the module 142 for the harmonic k; alternatively or cumulatively, this phase difference can represent the variation of the phase qx of the spectrum over the width of one or more spectral peaks corresponding to harmonics of the signal, this variation being relative to the minimum phase (p (k) assigned to In both cases, the peak or peaks for which the phase difference is quantized may be chosen as a function of the spectral energy represented by the upper envelope, which is available to the encoder and the decoder, which allows dec smell to determine which spectral line the gaps must

to be applied.

  In the first case, the phase differences may be subject to

  scalar quantification, or vector if they are grouped together for several peaks.

  In the second case, the variation of the phase (PX around the minimum phase p (k) over the width of a harmonic peak (determined by the width of the reference line used by the module 54) can be represented simply by the slope of a linear segment selected as being that which has a minimum quadratic distance with the curve of variation of

- 32 -

  phase of the spectrum over the width of the line, and possibly by an offset at the origin. These slopes can be subject to scalar quantification, or

  vector if we group them for several peaks.

  The quantization of phase variations on the harmonic peaks can relate to all the harmonic frequencies. Another possibility is to quantify several slopes each obtained by averaging slopes to harmonics on one or more subbands of the spectrum. This averaging can be weighted to take into account the energies relative to the different harmonic frequencies, represented by the upper envelope. The module 148 can also model the phase variation over the width of a peak by a more complex curve than a linear segment, for example a spline, the parameters of which are quantized to be transmitted to the

decoder.

  Another possibility is to carry out prior learning of harmonic phase models, representative of the variations in phase over the width of the peaks, observed on a corpus of reference signals. These models are stored in a dictionary stored by the modules 148 and 149. The encoder module 148 determines the indexes iAqp corresponding to the addresses of the models closest to the phase variations in the vicinity of the harmonic peaks considered, and the module 149 of the decoder recovers these

  models for the synthesis of the phase of the harmonic component.

-33 -

Claims (27)

  A method of coding an audio signal, in which a fundamental frequency (F0) of the audio signal is estimated, determining a spectrum of the audio signal by a frequency domain transform of a frame of the audio signal, and including in an output digital stream (1) coding data of a harmonic component of the audio signal, comprising data representative of spectral amplitudes associated with multiple frequencies of the estimated fundamental frequency, characterized in that said representative data of Spectral amplitudes associated with multiple frequencies of the fundamental frequency (F0) are obtained by means of cepstral coefficients calculated by transforming in the cepstral domain a compressed upper envelope (LXsup) of the spectrum of the audio signal, and that the coding data of the harmonic component further comprise, for at least one of the multiple frequencies of the estimated fundamental frequency, Data (iA (p iLif) relating to the spectrum phase   of the audio signal in the vicinity of said multiple frequency.   The method of claim 1, wherein the compressed upper envelope (LXsup) is determined by interpolating said spectral amplitudes associated with the frequency multiples of the frequency   fundamental (Fo) with application of a spectral compression function.   3. Method according to claim 1 or 2, in which the cepstral coefficients are transformed by scaling and smoothing in the cepstral domain, and minimum phases (p (k)) of the audio signal are calculated at frequencies that are multiples of the fundamental frequency ( Fo) on the basis of the transformed cepstral coefficients (cx1 [n]), 4. The method according to claim 3, wherein a phase difference between the spectrum of the audio signal and at least one frequency which is a multiple of the fundamental frequency (Fo) is quantized. ) and the corresponding computed minimum phase (p (k)), and quantization parameters (iAo) of said   phase difference in the coding data of the harmonic component. - 34 -   The method according to claim 3, wherein said scaling is adapted to minimize a phase difference between the spectrum of the audio signal at at least one frequency multiple of the fundamental frequency (F0) and the corresponding calculated minimum phase (rp ( k)), and include parameters (iLif) representative of the adapted scaling in the coding data of the harmonic component.   The method according to claim 5, wherein the minimized phase difference for scaling adaptation relates to at least one frequency multiple of the fundamental frequency (F0), selected on the basis of   The importance of the spectrum module in absolute value.   7. The method of claim 5, wherein a curve is estimated   spectral masking of the audio signal by means of a psycho-   acoustic noise, and the minimized phase difference for the adaptation of the scaling refers to at least one frequency multiple of the fundamental frequency (FO), selected on the basis of the importance of the spectrum modulus relative to the masking curve.   8. Process according to any one of claims 3 to 7, in   which one examines the variations of the cepstral coefficients (cxp [n]) to apply a smoothing less important in the presence of sudden variations than presence of slow variations.   9. The method according to claim 8, wherein the smoothing of the cepstral coefficients is performed by means of a forgetting factor (Xc) chosen as a function of a comparison between a threshold (dth) and a distance (d) between two games   successive cepstral coefficients (cxp [n-1], cxp [n]).   The method of claim 9, wherein said threshold (dth) is   adapted according to the variations of the cepstral coefficients (cxp [n]).   11. The method according to any one of claims 3 to 10, in   where the lifting is of the form cp (i) = [1 + y2i-Y1i] .c (i) - (, ui / i), where cp (i) and c (i) denote the cepstral coefficient of order i respectively before and after lifting, Y1 and y2 are coefficients between 0 and 1 and pL is a pre-emphasis coefficient.   12. The method of claim 11, wherein A = (Y2-y1) .c (1).   The method of any one of claims 1 to 12, wherein   which quantifies a variation of the phase of the spectrum of the audio signal on at least one frequency range surrounding a frequency that is a multiple of the fundamental frequency (F0), and includes parameters (iA (p) for quantization of said phase variation in the coding data of the harmonic component.   The method of claim 13, wherein the quantized phase variation is represented by the slope of the spectrum phase at the frequency   multiple of the fundamental frequency (F0).   The method of claim 13, wherein the quantized phase variation represents a phase model stored in a dictionary determined by learning.   16. Audio coder, comprising means for executing a method   according to any one of claims 1 to 15.   A method of decoding an input digital stream (d) representing an encoded audio signal, wherein inputting data in the input digital stream of data of a fundamental frequency (F0) of the audio signal and encoding data of a harmonic component comprising cepstral coefficient quantization data representing an upper spectral envelope (X_sup) of the audio signal, and generating a spectral estimate (Xv) of the harmonic component whose modulus is determined by said spectral envelope greater than the multiple frequencies of the fundamental frequency, characterized in that, in order to determine the phase (pv) of the spectral estimation of the harmonic component (Xv), the cepstral coefficients are transformed by scaling and smoothing in the cepstral domain, and calculates minimum phases (p (k)) of the audio signal at multiple frequencies - 36 -   of the fundamental frequency on the basis of the cepstral coefficients transformed (cxl [n]).   18. The method of claim 17, wherein, to obtain the phase (pv) of the spectral estimation (Xv) of the harmonic component, modifying at least one calculated minimum phase (qp (k)) on the database. phase difference quantification (iA & p) included in the data of coding of the harmonic component.   The method of claim 18, wherein the phase difference quantization data (iAp) is used to determine a change in the phase (dv) of the spectral estimate (Xv) of the harmonic component over at least one Frequency range surrounding a frequency   multiple of the fundamental frequency (F0).   The method of claim 19, wherein the phase variation determined over a frequency range is a linear variation.   represented by a slope quantization parameter.   21. The method of claim 19, wherein the phase variation determined over a frequency range represents a phase model.   memorized in a dictionary determined by learning.   22. Process according to any one of claims 17 to 21, in   which said scaling is adapted according to parameters (iLif) included in the   encoding data of the harmonic component.   23. Process according to any one of claims 17 to 22, in   which one examines the variations of the cepstral coefficients (cxp [n]) to apply a smoothing less important in the presence of sudden variations than presence of slow variations.   24. The method according to claim 23, wherein the smoothing of the cepstral coefficients (cxp [n]) is performed by means of a forgetting factor (Xc) chosen according to a comparison between a threshold (dth) and a distance (d) between two   successive sets of cepstral coefficients (cxp [n-1], cxp [n]).   25. The method of claim 24, wherein said threshold (dth) is   adapted according to the variations of the cepstral coefficients (cxp [n]).   26. A method of decoding an input digital stream (ó) representing a coded audio signal, in which input data of coding of a fundamental frequency (F0) of the audio signal and encoding data of a harmonic component comprising cepstral coefficient quantization data representing an upper spectral envelope (X_sup) of the audio signal, and generating a spectral estimate (Xv) of the harmonic component whose modulus is determined by said spectral envelope greater than the multiple frequencies of the fundamental frequency, characterized in that a variation of the phase (pv) of the spectral estimation of the harmonic component is determined on at least one frequency interval surrounding a frequency which is a multiple of the fundamental frequency on the Quantization parameter base (iA (p) included in the   encoding data of the harmonic component.   The method of claim 26, wherein the phase variation determined over a frequency range is a linear variation.   represented by a slope quantization parameter.   28. The method of claim 26, wherein the phase variation determined over a frequency range represents a phase model.   memorized in a dictionary determined by learning.   29. Audio decoder, comprising means for executing a method   according to any one of claims 17 to 28.