EP1383110A1 - Method and device for wide band speech coding, particularly allowing for an improved quality of voised speech frames - Google Patents

Abstract

The speech is sampled so as to obtain successive vocal frames each comprising a predetermined number of samples, and for each vocal frame parameters of a linear prediction model with code excitation are determined, these parameters comprising a digital word of long-term excitation (vi) extracted from an adaptive coded repertoire (DLT) and an associated long-term gain (Ga), as well as a short-term excitation word (cj) extracted from an algebraic coded repertoire ( DCT) using digital linear prediction filtering (FP), and associated short-term gain (Gc). The adaptive coded directory is updated from the extracted long-term excitation word and the extracted short-term excitation word, and the state of the linear prediction filter (FP) is updated with the word d short-term excitation filtered by a filter (FLT1) of order greater than or equal to 1, the coefficients of which depend on the value of the long-term gain, so as to weaken the contribution of the short-term excitation when the gain of long-term excitement is above a predetermined threshold. <IMAGE>

Description

The invention relates to speech encoding / decoding extended band, in particular but not limited to telephony mobile.

In wideband, the bandwidth of the speech signal is between 50 and 7000 Hz.

Successive speech sequences sampled at one predetermined sampling frequency, for example 16 kHz, are processed in a coding device using a prediction linear excitation by coded sequences (ACELP: “algebraic-code-excited linear-prediction ”), well known to those skilled in the art, and described in particular in recommendation ITU-TG 729, version 3/96, titled “speech coding at 8 kbit / s by prediction linear with excitation by coded sequences with algebraic structure conjugate ”.

We will now briefly recall, referring to the Figure 1, the main features and functionality of such coder, the skilled person can refer for all practical purposes, for more details, see the above-mentioned recommendation G 729.

The prediction coder CD, of the ACELP type, is based on the linear predictive coding model with code excitation. The coder operates on vocal superframes equivalent for example to 20 ms of signal and each comprising 320 samples. The extraction of the linear prediction parameters, i.e. the coefficients of the linear prediction filter, also called short-term synthesis filter 1 / A (z), is carried out for each speech superframe. On the other hand, each superframe is subdivided into 5 ms frames comprising 80 samples. All the frames, the speech signal is analyzed to extract the parameters of the CELP prediction model (that is to say, in particular, a long-term digital excitation word v _i extracted from an adaptive coded DLT directory, also called "adaptive long term dictionary", an associated long term gain Ga, a short term excitation word c _j , extracted from a DCT algebraic coded repertoire, also called "fixed coded repertoire" or "short term dictionary algebraic ", and an associated short-term gain Gc).

These parameters are then coded and transmitted.

On reception, these parameters are used, in a decoder, to retrieve the excitation and predictive filter parameters. We then reconstitutes speech by filtering this excitation flow in a short-term synthesis filter.

While the Adaptive DLT Dictionary Contains Words digital representative of representative tone delays of past excitations, the short-term dictionary DCT is founded on an algebraic structure using a permutation model intertwined with Dirac pulses. In this coded directory, which contains innovative excitations also called excitations algebraic or short-term, each vector contains a certain number of non-zero pulses, for example four, each of which can have amplitude +1 or -1 with predetermined positions.

The CD encoder processing means include functionally of the first MEXT 1 extraction means intended to extract the word long-term excitement, and second MEXT2 extraction means intended to extract the word short-term excitement. Functionally, these means are made for example in software within a processor.

These extraction means include a predictive filter FP having a transfer function equal to 1 / A (z), as well as a filter FPP perceptual weighting with a transfer function W (z). The perceptual weighting filter is applied to the signal to model the perception of the ear.

Furthermore, the extraction means include means MECM intended to perform a minimization of a square error average.

The linear prediction FP synthesis filter models the spectral envelope of the signal. Linear predictive analysis is performed all superframes, so as to determine the linear predictive filter coefficients. These are converted to spectral line pairs (LSP: “Line Spectrum Pairs”) and digitized by predictive vector quantization in two stages.

Each 20 ms speech superframe is divided into four frames of 5 ms each containing 80 samples. The settings Quantized LSPs are transmitted to the decoder once per superframe while long term and short term parameters are passed at each frame.

The coefficients of the linear prediction filter, quantified and not quantified, are used for the most recent frame of a super-frame, while the other three frames of the same super-frame use an interpolation of these coefficients. Tonal delay open loop is estimated every two frames based on the perceptually weighted voice signal. Then, the following operations are repeated at each frame:

The long-term target signal X _LT is calculated by filtering the sampled speech signal s (n) by the perceptual weighting filter FPP.

We then subtract the response to null input of the weighted synthesis filter FP, FPP, so that get a new long-term target signal.

The impulse response of the weighted synthesis filter is calculated.

A closed loop tonal analysis using a minimization of the mean square error is then carried out in order to determine the long-term excitation word v _i and the associated gain Ga, by means of the target signal and the impulse response, by searches around the value of the tone delay in open loop.

The long-term target signal is then updated by subtracting the filtered contribution y from the adaptive coded directory DLT and this new short-term target signal X _ST is used when exploring the fixed coded directory DCT in order to determine the password. short term excitation C _j and the associated gain G _c . Again, this closed loop search is performed by minimizing the mean square error.

Finally, the long-term adaptive DLT dictionary as well that the memories of the filters FP and FPP, are updated by means of the long term and short term excitation words so determined.

The quality of a CELP algorithm strongly depends on the richness of the DCT algebraic excitation dictionary. If the efficiency of such an algorithm is indisputable for band signals narrow bandwidth (300-3400 Hz), problems arise for wideband signals.

The inventors have in fact observed that even with a very rich algebraic dictionary, the algorithm for encoding the speech produces a reconstructed signal corrupted by different kinds noises and in particular a “whistling” noise which taints voiced speech frames.

This high frequency noise comes from excitation at short term that introduces unwanted artifacts. Two types of solutions to solve this problem have already been proposed in the literature.

• dans l'article de Gerson et Jasiuk, intitulé « Techniques for Improving the Performance of CELP-Type Speech Coders », IEEE, Journal on Selected Areas In Communications, Vol. 10, N°5, Juin 1992, pages 858-865, ou bien
• dans l'article de Miki et autres, intitulé « A Pitch Synchronous Innovation CELP (PSI-CELP) Coder for 2-4 kbit/s », Proc., IEEE Int. Conf. Acoustics, Speech, and Signal Processing, ICASSP'94, Adelaide, South Australia, 1994, Vol.II, pages 113-116,

propose de rendre la contribution à court terme périodique. Un autre type de solution, décrit par exemple

• dans l'article de Taniguchi, Johnson et Ohta, intitulé « Pitch Sharpening for Perceptually Improved CELP, and the Sparse-Delta Codebook for Reduced Computation », Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Processing, ICASSP'91, Toronto, Canada 1991, pages 241-244, ou
• dans l'article de Shoham, intitulé « Constrained-Stochastic Excitation Coding of Speech At 4,8 kb/s », Advances in Speech Coding, B.S. Atal, V. Cuperman, and A. Gersho, Eds., Dordrecht, The Netherlands, Kluwer, 1991, pages 339-348,

propose un contrôle de façon adaptative du gain à court terme.A first type of solution, described for example

• in the article by Gerson and Jasiuk, entitled "Techniques for Improving the Performance of CELP-Type Speech Coders", IEEE, Journal on Selected Areas In Communications, Vol. 10, N ° 5, June 1992, pages 858-865, or else
• in the article by Miki et al., entitled "A Pitch Synchronous Innovation CELP (PSI-CELP) Coder for 2-4 kbit / s", Proc., IEEE Int. Conf. Acoustics, Speech, and Signal Processing, ICASSP'94, Adelaide, South Australia, 1994, Vol. II, pages 113-116,

proposes to make the short-term contribution periodic. Another type of solution, described for example

• in the article by Taniguchi, Johnson and Ohta, entitled “Pitch Sharpening for Perceptually Improved CELP, and the Sparse-Delta Codebook for Reduced Computation”, Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Processing, ICASSP'91, Toronto, Canada 1991, pages 241-244, or
• in Shoham's article, "Constrained-Stochastic Excitation Coding of Speech At 4.8 kb / s", Advances in Speech Coding, BS Atal, V. Cuperman, and A. Gersho, Eds., Dordrecht, The Netherlands, Kluwer, 1991, pages 339-348,

offers adaptive short-term gain control.

The invention also provides a control type solution gain, but totally different from that described in particular in the articles of Taniguchi and others and of Shoham.

The invention also aims to control regardless of short-term and long-term distortions.

The invention therefore provides a speech encoding method with wide band, in which the speech is sampled so as to obtain successive voice frames each comprising a predetermined number of samples, and for each voice frame, we determines parameters of a linear prediction model at excitation by code, these parameters comprising a numeric word of long-term excitement extracted from an adaptive coded repertoire and a associated long-term gain, as well as a word of short-term excitement extract from an algebraic coded repertoire and short-term gain associated, and we update the adaptive coded directory from the word excerpt long term excitement and short term excitement word extract.

According to a general characteristic of the invention, the method includes an update of the state of the linear prediction filter with the word short-term excitation filtered by an order filter greater than or equal to 1, for example an impulse response filter finite of order 1, whose coefficients depend on the value of the gain in the long run, so as to weaken the contribution of excitement to short term when the gain of long term excitement is greater at a predetermined threshold, for example equal to 1.

In other words, the solution according to the invention consists here to weaken the contribution of short-term excitement if the gain of long-term excitement is important. However, this is the contribution of undiminished short-term excitement which is stored in the adaptive dictionary for updating. So the reduction occurs only on exit. Preserving the magnitude of the short-term contribution to be stored is important, since the richness of the adaptive dictionary is thus preserved for the lowest frequencies.

Of course, the gain correction must also be applied when reconstructing the signal at the decoder.

According to an embodiment of the invention in which the filter is of order 1 and its transfer function equal to B0 + B1 z ^-1 , the first coefficient B0 of the filter is equal to 1 / (1 + β. min (Ga, 1)), and the second coefficient B1 of the filter is equal to β.min (Ga, 1) / (1 + β.min (Ga, 1)), where β is a real number with a lower absolute value at 1, Ga is the long-term gain and min (Ga, 1) designates the minimum value between Ga and 1.

According to a variant of the invention, the extraction of the word of long-term excitement using a first filter of perceptual weighting including a first weighting filter formant, we perform the short excitation word extraction term using the first perceptual weighting filter cascaded to a second perceptual weighting filter comprising a second formantic weighting filter. The denominator of the transfer function of the first formantic weighting filter equals the numerator of the second weighting filter formant.

Thus, according to this variant, the use of two filters of weighting different formant allows to control regardless of short-term and long-term distortions. The short-term weighting filter is cascaded to the filter of long-term weighting. Also, tying the denominator of the long-term weighting filter in the numerator of the short-term weighting allows these two to be controlled separately filters and also allows a clear simplification when these two filters are cascaded.

It is planned to update the status of the two filters perceptual weighting with the filtered short-term excitation word by the filter of order greater than or equal to 1.

• des moyens d'échantillonnage aptes à échantillonner la parole de façon à obtenir des trames vocales successives comportant chacune un nombre prédéterminé d'échantillons,
• des moyens de traitement aptes à chaque trame vocale, à déterminer des paramètres d'un modèle de prédiction linéaire à excitation par code, ces moyens de traitement comportant des premiers moyens d'extraction aptes à extraire un mot numérique d'excitation à long terme d'un répertoire codé adaptatif et à calculer un gain à long terme associé, et des deuxièmes moyens d'extraction aptes à extraire un mot d'excitation à court terme d'un répertoire codé algébrique et à calculer un gain à court terme associé, et
• des premiers moyens de mise à jour aptes à mettre à jour le répertoire codé adaptatif à partir du mot d'excitation à long terme extrait et du mot d'excitation à court terme extrait.

The subject of the invention is also a device for encoding speech with a wide band, comprising

• sampling means able to sample the speech so as to obtain successive speech frames each comprising a predetermined number of samples,
• processing means suitable for each voice frame, in determining parameters of a linear prediction model with code excitation, these processing means comprising first extraction means capable of extracting a digital word of long-term excitation d '' an adaptive coded repertoire and calculating an associated long-term gain, and second extraction means capable of extracting a short-term excitation word from an algebraic coded repertoire and calculating an associated short-term gain, and
• first updating means capable of updating the adaptive coded directory on the basis of the extracted long-term excitation word and the extracted short-term excitation word.

According to a general characteristic of the invention, the first extraction means include a digital filter of linear prediction, and the device comprises second means update capable of updating the filter status of the linear prediction with short term excitation word filtered by a filter of order greater than or equal to 1 whose coefficients depend on the value of the long-term gain, so as to weaken the contribution of short-term excitement when the gain of long-term excitement is above a predetermined threshold.

According to one embodiment of the invention, the first extraction means include a first weighting filter perceptual with a first weighting filter formantic, by the fact that the second means of extraction include the first perceptual weighting filter cascaded to a second perceptual weighting filter comprising a second formantic weighting filter, and the denominator of the transfer function of the first formantic weighting filter equals the numerator of the second weighting filter formant.

The invention also relates to a terminal of a system wireless communication, such as a mobile phone cell, incorporating a device as defined above.

• la figure 1, déjà décrite, illustre schématiquement un dispositif d'encodage de la parole, selon l'art antérieur ;
• la figure 2 illustre schématiquement un mode de réalisation d'un dispositif d'encodage, selon l'invention, et la figure 2a illustre schématiquement un mode de réalisation d'un décodeur correspondant ;
• la figure 3 illustre schématiquement un autre mode de réalisation d'un dispositif d'encodage, selon l'invention ; et
• la figure 4 illustre schématiquement l'architecture interne d'un téléphone mobile cellulaire incorporant un dispositif de codage, selon l'invention.

Other advantages and characteristics of the invention will appear on examining the detailed description of embodiments and implementation, in no way limiting, and the appended drawings, in which:

• Figure 1, already described, schematically illustrates a speech encoding device, according to the prior art;
• FIG. 2 schematically illustrates an embodiment of an encoding device according to the invention, and FIG. 2a schematically illustrates an embodiment of a corresponding decoder;
• Figure 3 schematically illustrates another embodiment of an encoding device according to the invention; and
• FIG. 4 schematically illustrates the internal architecture of a cellular mobile telephone incorporating a coding device according to the invention.

The encoding device, or CD encoder, according to the invention, as illustrated in FIG. 2, differs from that of the prior art as illustrated in FIG. 1 by the fact that the CD encoder further comprises second updating means MAJ2 able to update the state of the linear prediction filter FP and the state of the perceptual weighting filter FPP with the short-term excitation word c _j filtered by a filter of order greater than or equal to 1 FLT1 which is for example here a filter of order 1 with finite impulse response.

The coefficients of this first order filter depend on the value long-term gain Ga, so as to weaken the contribution of short-term excitement when gaining long-term excitement Ga is greater than a predetermined threshold, for example equal to 1.

As an example, the transfer function of the filter FLT1 is equal to B0 + B1 z ^-1 and the first coefficient of the filter B0 can be determined by formula (I) below. 1 / (1 + 0.98 min (Ga, 1)) while the second coefficient of filter B1 can be determined by formula (II) below. 0.98 min (Ga, 1) / (1 + 0.98 min (Ga, 1))

However, it is the short-term contribution not weakened (gain Gc) which is stored in the adaptive DLT dictionary for its update.

Thus, the attenuation occurs only on the signal of output and keeping the excitement contribution short term to store keeps the richness of the dictionary adaptive for the lowest frequencies.

Naturally, the filtering of the excitation must also be applied for updating the memory status of the filters in the DCD decoder, as shown diagrammatically in FIG. 2a.

The variant embodiment illustrated in FIG. 2 allows eliminating hissing type noise on voiced speech frames.

The FPP perceptual weighting filter uses the masking properties of the human ear compared to the spectral envelope of the speech signal, whose shape is a function resonances of the vocal tract. This filter allows you to assign more importance of the error appearing in the spectral valleys by compared to formic peaks.

In the variant illustrated in FIG. 2, the same FPP perceptual weighting filter is used for short-term research and for long-term research. The transfer function W (z) of this FPP filter is given by the formula (III) below. W ( z ) = AT ( z / γ 1 ) AT ( z / γ 2 ) in which 1 / A (z) is the transfer function of the predictive filter FP and γ1 and γ2 are the perceptual weighting coefficients, the two coefficients being positive or zero and less than or equal to 1 with the coefficient γ2 less than or equal to the coefficient γ1.

In general, the perceptual weighting filter consists of a formantic weighting filter and a weighting of the slope of the spectral envelope of the signal (tilt).

In this case, assume that the weighting filter perceptual is only formed by the weighting filter formant whose transfer function is given by the formula (III) above.

The spectral nature of the long-term contribution is different from the short-term contribution. Therefore, it is advantageous to use two formantic weighting filters different, allowing independent control of distortions at short term and long term.

Such an embodiment is illustrated in FIG. 3, in which, compared to Figure 2, the unique FPP filter was replaced by a first formantic weighting filter FPP1 for long-term research, cascaded with a second filter of FPP2 formant weighting for short-term research.

Since the short-term weighting filter FPP2 is cascaded to the long-term weighting filter, the filters appearing in the long-term research loop should also appear in the short-term research loop.

The transfer function W ₁ (z) of the formantic weighting filter FPP1 is given by formula (IV) below. W 1 ( z ) = AT ( z / γ 11 ) AT ( z / γ 12 ) while the transfer function W ₂ (z) of the formantic weighting filter FPP2 is given by the formula (V) below. W 2 ( z ) = AT ( z / γ 21 ) AT ( z / γ 22 )

Furthermore, the coefficient γ ₁₂ is equal to the coefficient γ ₂₁ . This allows a clear simplification when cascading these two filters. Thus, the filter equivalent to the cascade of these two filters has a transfer function given by the formula (VI) below. AT ( z / γ 11 ) AT ( z / γ 22 )

Furthermore, if the value 1 is used for the coefficient γ ₁₁ , then the synthesis filter FP (having the transfer function 1 / A (z)) followed by the long-term weighting filter FPP1 and the weighting filter FPP2 is then equivalent to the filter whose transfer function is given by formula (VII) below. 1 AT ( z / γ 22 )

This further reduces the complexity of the excitation extraction algorithm.

As an indication, one can for example use for the coefficients γ ₁₁ , γ ₂₁ = γ ₁₂ and γ ₂₂ , the respective values 1; 0.1 and 0.9.

The invention advantageously applies to telephony mobile, and in particular to all remote terminals belonging to a wireless communication system.

Such a terminal, for example a TP mobile telephone, such as that illustrated in FIG. 4, conventionally comprises a antenna connected via a DUP duplexer to a chain reception CHR and a CHT transmission chain. A baseband processor BB is connected to the chain respectively of reception CHR and to the chain of transmission CHT by via analog digital ADCs and analog digital DACs.

Conventionally, the processor BB performs processing in baseband, including DCN channel decoding, followed by DCS source decoding.

For transmission, the processor performs source coding CCS followed by CCN channel coding.

When the mobile phone incorporates an encoder according to the invention, it is incorporated within the coding means of CCS source, while the decoder is incorporated within the means DCS source decoding.

Claims (12)

Wideband speech encoding method, in which speech is sampled so as to obtain successive speech frames each comprising a predetermined number of samples, and for each speech frame parameters of a linear prediction model are determined with code excitation, these parameters comprising a long term digital excitation word (v _i ) extracted from an adaptive coded directory (DLT) and an associated long term gain (Ga), as well as an excitation word short-term (cj) extracted from an algebraic coded repertoire (DCT) using digital linear prediction filtering (FP), and an associated short-term gain (Gc), and the adaptive coded repertoire is updated from of the extracted long-term excitation word and of the extracted short-term excitation word, characterized in that the method comprises updating the state of the linear prediction filter (FP) with the excitation word short t erme filtered by a filter of order greater than or equal to 1 (FLT1) whose coefficients depend on the value of the long-term gain, so as to weaken the contribution of the short-term excitation when the gain of the excitation at long term is above a predetermined threshold. Method according to claim 1, characterized in that the predetermined threshold is equal to 1. Method according to claim 2, characterized in that the filter (FLT1) is of order 1 and its transfer function equal to B0 + B1 z ^-1 , in that the first coefficient B0 of the filter is equal to 1 / ( 1 + β.min (Ga, 1)), and the second coefficient B1 of the filter is equal to β.min (Ga, 1) / (1 + β.min (Ga, 1)), where β is a real number with an absolute value less than 1, Ga is the long-term gain and min (Ga, 1) designates the minimum value between Ga and 1. Method according to one of the preceding claims, characterized in that the extraction of the long-term excitation word is carried out using a first perceptual weighting filter (FPP1) comprising a first formant weighting filter, by the fact that the short term excitation word is extracted using the first perceptual weighting filter (FPP1) cascaded to a second perceptual weighting filter (FPP2) comprising a second formantic weighting filter, and by the fact that the denominator of the transfer function of the first formant weighting filter is equal to the numerator of the second formant weighting filter. Method according to claim 4, characterized in that it includes an update of the state of the two perceptual weighting filters (FPP1, FPP2) with the short-term excitation word filtered by said filter of order 1 . Wideband speech encoding device comprising sampling means able to sample the speech so as to obtain successive speech frames each comprising a predetermined number of samples, processing means suitable for each voice frame, in determining parameters of a linear prediction model with code excitation, these processing means comprising first extraction means (MEXT1) capable of extracting a digital excitation word with long term of an adaptive coded repertoire and calculating an associated long term gain, and second extraction means (MEXT2) able to extract a short term excitation word from an algebraic coded repertoire and calculating a gain associated short-term, and first updating means (MAJ) capable of updating the adaptive coded directory on the basis of the extracted long-term excitation word and the extracted short-term excitation word, characterized by the fact that the first extraction means comprise a digital linear prediction filter (FP), and by the fact that the device comprises second updating means (MAJ2) capable of carrying out an updating of the state of the linear prediction filter with the short-term excitation word filtered by a filter (FLT1) of order greater than or equal to 1, the coefficients of which depend on the value of the long-term gain, so as to weaken the contribution of short-term excitation when the gain of long-term excitation is greater than a predetermined threshold. Device according to claim 6, characterized in that the predetermined threshold is equal to 1. Device according to Claim 7, characterized in that the filter (FLT1) is of order 1 and its transfer function equal to B0 + B1 z ^-1 , in that the first coefficient B0 of the filter is equal to 1 / (1 + β.min (Ga, 1)), and the second coefficient B1 of the filter is equal to β.min (Ga, 1) / (1 + β.min (Ga, 1)), where β is a number real with an absolute value less than 1, Ga is the long-term gain and min (Ga, 1) designates the minimum value between Ga and 1. Device according to one of Claims 6 to 8, characterized in that the first extraction means comprise a first perceptual weighting filter (FPP1) comprising a first formantic weighting filter, in that the second extraction means include the first perceptual weighting filter cascaded to a second perceptual weighting filter (FPP2) having a second formantic weighting filter, and by the fact that the denominator of the transfer function of the first formantic weighting filter is equal to the numerator of the second formantic weighting filter. Device according to claim 9, characterized in that the second updating means are capable of updating the state of the two perceptual weighting filters with the short-term excitation word filtered by said filter d 'order 1. Terminal of a wireless communication system, characterized in that it incorporates a device according to one of claims 6 to 10. Terminal according to claim 11, characterized in that it forms a cellular mobile telephone.

Images