EP1395981B1 - Device and method for processing an audio signal - Google Patents

Abstract

The invention concerns audio signal processing, comprising: a first processing of an audio source signal, using at least a mathematical transform applied on first sequences of samples obtained by applying first segmentation windows on the audio source signal; and a second audio processing applied on second sequences of samples obtained by applying second segmentation windows on the signal delivered by the first step; the two successive first windows and/or the two successive second windows overlapping, the overlaps being such that the segmentations are synchronous.

Description

The present invention relates to the field of audio signal processing.

More specifically, the invention relates, in particular, the reduction or cancellation of noise in an audio signal processed by a digital communication device, for example of the digital telephone type and / or radio mobile phones type hands-free.

When digital audio communication devices are used in a noisy environment (typically inside a car), the latter can greatly disrupt an audio signal and consequently degrade the quality of a communication.

According to the known techniques, this problem is remedied by inserting noise attenuators or cancellers, acting on the signal picked up by a microphone, before a specific processing of the audio signal.

According to a first known technique, a cancellation and echo or noise reduction device is inserted between a microphone intended to pick up an audio signal and a device for processing the audio signal. This device improves the useful signal to noise ratio or reduces the echo so that the signal can be processed later under optimized conditions. Nevertheless, this technique of the prior art requires a specific dedicated device, which has the disadvantage of resulting in additional costs and increased complexity of use.

According to a second known technique, the noise reduction function, based on the use of a Fast Fourier Transform (FFT) applied to a continuous stream of voice samples, is integrated in the digital communication device. At first, the sample flow is split into 256 sample windows obtained by applying a formatting window, with the windows overlapping by half (the first 128 samples of a window corresponding to the last 128 samples from the previous window). An FFT is applied to each window then the result of the FFT is processed by a function of cancellation or reduction of noise or echo.

Then, the result of this function is processed by an inverse fast Fourier transform (or IFFT) to reconstruct a flow of voice samples that can be processed by a voice processing function.

A disadvantage of this technique of the prior art as illustrated in the document WO98 / 06090 is that it is relatively complex to implement.

The invention in its various aspects is intended to overcome these disadvantages of the prior art.

More specifically, an object of the invention is to provide a method and an audio processing device in a device that allows a reduction in the complexity of a processing based on a mathematical transformation applying to blocks of data while optimizing audio processing applying to audio frames.

Another object of the invention is to optimize the integration of processing based on a mathematical transformation and audio processing.

An object of the invention is also to optimize the delays of these treatments.

Another object of the invention is to reduce the computing power necessary for these treatments.

• une première étape de traitement d'un signal audio source, mettant en oeuvre au moins une transformation mathématique appliquée sur des premières séquences d'échantillons obtenues par l'application de premières fenêtres de segmentation sur le signal audio source ; et
• une deuxième étape de traitement audio, appliquée sur des secondes séquences d'échantillons obtenues par l'application de secondes fenêtres de segmentation sur le signal délivré par la première étape, les secondes fenêtres de segmentation étant distinctes des premières fenêtres de segmentation ;

remarquable en ce que deux premières fenêtres successives et/ou deux secondes fenêtres successives se chevauchent, les chevauchements étant tels que les segmentations soient synchrones.For this purpose, the invention as defined in claims 1 and 11 proposes a method for processing an audio signal, comprising:

• a first step of processing a source audio signal, implementing at least one mathematical transformation applied to first sample sequences obtained by the application of first segmentation windows on the source audio signal; and
• a second audio processing step, applied on second sample sequences obtained by the application of second segmentation windows on the signal delivered by the first step; second segmentation windows being distinct from the first segmentation windows;

remarkable in that two successive windows and / or two successive second windows overlap, the overlaps being such that the segmentations are synchronous.

Thus, the audio processing steps can be implemented sequentially or in a multitasking environment. Moreover, this implementation is facilitated by the use of memory with a predictable, accurate and economical dimensioning.

According to one particular characteristic, the method is remarkable in that the second segmentation windows are successive frames.

Thus, according to the invention, the processing times of the process are optimized.

According to one particular characteristic, the method is remarkable in that the last sample of a first sequence is also the last sample, after the first step, of the corresponding second sequence.

Thus, preferably the second stage of audio processing is performed without unnecessary waiting to optimize the overall time of audio processing.

• d'une première fenêtre intermédiaire à reconstruction parfaite et possédant des propriétés spectrales adaptées à la ou aux transformations mathématiques ; et
• d'une deuxième fenêtre intermédiaire rectangulaire.

According to one particular characteristic, the method is remarkable in that each first segmentation window is a window with perfect reconstruction obtained by convolution:

• a first intermediate window with perfect reconstruction and having spectral properties adapted to the mathematical transformation or transformations; and
• a second rectangular intermediate window.

Thus, the portions of the first overlapping segmentation windows are perfectly reconstructed, which makes it possible to have a recombination of the signals during the first relatively simple treatment.

In addition, the first intermediate window being adapted to the mathematical transformation or transformations (it is notably an attenuation of the second lobe of the window relatively strong while the main lobe remains flat), the quality of the corresponding treatment is optimized.

In addition, the second intermediate window being rectangular, the corresponding sample processing is simple and effective.

• une sous-étape de traitement prédéterminé appliquée à la première séquence;
• une sous-étape de transformation mathématique inverse appliquée aux échantillons traités de la première séquence; et
• une étape d'addition des échantillons vocaux issus de la sous étape de transformation mathématique inverse appliquée à la première séquence et des échantillons vocaux correspondants, issus de la sous étape de transformation mathématique inverse appliquée à la première séquence précédente.

According to a particular characteristic, the method is remarkable in that the first processing step applied to each first sequence further comprises:

• a predetermined processing sub-step applied to the first sequence;
• a substep of inverse mathematical transformation applied to the processed samples of the first sequence; and
• a step of adding the voice samples from the inverse mathematical transformation sub-step applied to the first sequence and corresponding speech samples from the inverse mathematical transformation sub-step applied to the first preceding sequence.

According to a particular characteristic, the method is remarkable in that the predetermined processing sub-step comprises a reduction or cancellation of noise in the audio signal.

• une réduction ou une annulation d'écho dans le signal audio ;
• une reconnaissance vocale dans le signal audio.

According to a particular characteristic, the method is remarkable in that the predetermined processing substep comprises at least one treatment belonging to the group comprising:

• an echo reduction or cancellation in the audio signal;
• voice recognition in the audio signal.

Thus, the method advantageously combines treatments such as reduction and / or cancellation of noise and / or echo and / or voice recognition in a device (for example of the telephone, personal computer or remote control type) which allows a reduction of the complexity while optimizing the effectiveness of these treatments and / or a strong integration of the device (which allows, as a consequence, a reduction in costs and energy consumption this which is relatively important especially for battery-operated communications devices).

• les transformations rapides de Fourrier (FFT) et leurs variantes ;
• les transformations rapides de Hadamard (FHT) et leurs variantes ; et
• les transformations en cosinus discrètes (DCT) et leurs variantes .

According to one particular characteristic, the method is remarkable in that said mathematical transformation or said transformations belong to the group comprising:

• fast Fourier transformations (FFT) and their variants;
• fast transformations of Hadamard (FHT) and their variants; and
• discrete cosine transformations (DCT) and their variants.

Thus, the invention advantageously makes it possible to use one or more mathematical transformations adapted to the first audio processing, these transformations applying to blocks of a size different from the size of the second segmentation windows.

According to a particular characteristic, the method is remarkable in that the source audio signal is a voice signal.

The invention is thus well suited to the second audio processing when it is specific to speech such as, for example, voice coding ("vocoding") and / or voice compression for storage and / or remote transmission.

• des premiers moyens de traitement d'un signal audio source, mettant en oeuvre au moins une transformation mathématique appliquée sur des premières séquences d'échantillons obtenues par l'application de premières fenêtres de segmentation sur le signal audio source ; et
• des deuxièmes moyens de traitement audio, appliquées sur des secondes séquences d'échantillons obtenues par l'application de secondes fenêtres de segmentation sur le signal délivré par la première étape, les secondes fenêtres de segmentation étant distinctes des premières fenêtres de segmentation ;

remarquable en ce que deux premières fenêtres successives et/ou deux secondes fenêtres successives se chevauchent, les chevauchements étant tels que les segmentations soient synchrones.The invention also relates to a device for processing an audio signal, comprising:

• first means for processing a source audio signal, implementing at least one mathematical transformation applied to first sample sequences obtained by the application of first segmentation windows on the source audio signal; and
• second audio processing means, applied on second sample sequences obtained by the application of second segmentation windows on the signal delivered by the first step, the second segmentation windows being distinct from the first segmentation windows;

remarkable in that two successive windows and / or two successive second windows overlap, the overlaps being such that the segmentations are synchronous.

• une première étape de traitement d'un signal audio source, mettant en oeuvre au moins une transformation mathématique appliquée sur des premières séquences d'échantillons obtenues par l'application de premières fenêtres de segmentation sur le signal audio source ; et
• une deuxième étape de traitement audio, appliquée sur des secondes séquences d'échantillons obtenues par l'application de secondes fenêtres de segmentation sur le signal délivré par la première étape, les secondes fenêtres de segmentation étant distinctes des premières fenêtres de segmentation ;

deux premières fenêtres successives et/ou deux secondes fenêtres successives se chevauchant, les chevauchements étant tels que les segmentations soient synchrones.The invention further relates to a computer program product comprising program elements recorded on a medium readable by at least one microprocessor, wherein the program elements control the microprocessor (s) to perform:

• a first step of processing a source audio signal, implementing at least one mathematical transformation applied to first sample sequences obtained by the application of first segmentation windows on the source audio signal; and
• a second audio processing step, applied to second sample sequences obtained by the application of second segmentation windows on the signal delivered by the first step, the second segmentation windows being distinct from the first segmentation windows;

first two successive windows and / or two successive second windows overlapping, the overlaps being such that the segmentations are synchronous.

In addition, the invention relates to a computer program product, which is remarkable in that the program comprises instruction sequences adapted to the implementation of an audio processing method as described above when the program is executed on a computer.

The advantages of the audio signal processing device, and computer program products are the same as those of the audio signal processing method, they are not detailed further.

• la figure 1 présente un synoptique général d'un radiotéléphone, conforme à l'invention selon un mode particulier de réalisation ;
• la figure 2 illustre les traitements successifs effectués par le radiotéléphone de la figure 1, sur un signal vocal ;
• la figure 3 présente un algorithme d'annulation ou de réduction de bruit, selon la figure 2;
• la figure 4 présente un traitement vocal appliquer à une trame, selon la figure 2;
• la figure 5 décrit un fenêtrage du flux d'échantillons tel qu'effectué par les traitements des figures 3 et 4 ;
• la figure 6 illustre une fenêtre de mise en forme connue en soi;
• la figure 7 illustre une fenêtre de mise en forme, optimisée et utilisée dans les opérations de fenêtrage de la figure 3 selon un mode préférentiel de l'invention ; et
• la figure 8 décrit plus précisément un traitement de type réduction de bruit présentée à la figure 3.

Other characteristics and advantages of the invention will appear more clearly on reading the following description of a preferred embodiment, given as a simple illustrative and nonlimiting example, and the appended drawings, among which:

• Figure 1 shows a general block diagram of a radiotelephone, according to the invention according to a particular embodiment;
• FIG. 2 illustrates the successive processing carried out by the radiotelephone of FIG. 1, on a voice signal;
• FIG. 3 presents a cancellation or noise reduction algorithm, according to FIG. 2;
• Figure 4 shows a voice processing applied to a frame, according to Figure 2;
• Figure 5 depicts a windowing of the sample stream as performed by the processes of Figures 3 and 4;
• FIG. 6 illustrates a shaping window known per se;
• FIG. 7 illustrates a formatting window, optimized and used in the windowing operations of FIG. 3 according to a preferred embodiment of the invention; and
• FIG. 8 more precisely describes a noise reduction type processing shown in FIG. 3.

• des traitements basés sur une FFT notamment des traitements d'annulation ou de réduction de bruit; et
• de traitement vocal de type codage de la parole.

The general principle of the invention is based on synchronization:

• FFT-based treatments including noise cancellation or noise reduction treatments; and
• speech coding type speech processing.

Indeed, FFT and IFFT deal with windows with a power of 2 samples (typically 128 or 256).

On the other hand, the speech coding takes into account windows that do not have the same size (typically the voice processing in the context of the GSM considers windows of 160 samples).

In the case, for example, of a GSM radio telephone complying with the standards published by the ETSI (European Telecommunication Standard Institute), the voice signal is sampled at a frequency of 8 kHz before being transmitted in 20 ms form. compressed to a recipient.

• « Full Rate (FR) speech transcoding » (GSM06.10) (ou « codage de parole à plein débit » en français) ;
• « Half Rate (HR) speech transcoding » (GSM06.20) (ou « codage de parole à demi-débit » en français);
• « Enhanced Full Rate (EFR) speech transcoding » (GSM06.60) (ou « codage de parole à plein débit amélioré » en français) ; et
• « Adaptive Multi-Rate (AMR) speech transcoding » (GSM 06.90) (ou « codage de parole à débit multiple adaptatif » en français).

It is noted that, according to the GSM standard, speech coding is performed on frames of 160 samples by a vocoder. This coding which is a function of the desired flow rate is specified in particular in the following documents:

• "Full Rate (FR) speech transcoding" (GSM06.10) (or "full-rate speech coding" in French);
• "Half Rate (HR) speech transcoding" (GSM06.20) (or "half-rate speech coding" in French);
• "Enhanced Full Rate (EFR) speech transcoding" (GSM06.60) (or "enhanced full-rate speech coding" in French); and
• "Adaptive Multi-Rate (AMR) speech transcoding" (GSM 06.90) (or "adaptive multiple rate speech coding" in French).

According to the state of the art, considering a window of 160 vocally processed samples, the noise reduction and / or echo cancellation and / or cancellation device processes a window of length 256 which can intersect up to three windows. length 160. It is, among other things, the asynchronism inherent in this state-of-the-art technique that makes these processes complex and requires an oversizing of the memories and the computing power and / or the clock. a DSP (Signal Processing Processor) of the "Digital Signal Processor" used for the calculations.

According to the invention, the two types of processing are synchronized by systematically matching the end of a cancellation or noise reduction and / or echo window with a voice processing frame and preferably with the end of a voice processing frame. Thus, if the noise reduction or cancellation windows have a size equal to 256 samples and if the speech processing frames have a size equal to 160 samples, a reduction or echo cancellation window will contain the a full voice processing frame and 96 samples (ie 256 minus 160) from the previous window.

Thus, the synchronism between the noise reduction or cancellation windows and the speech processing frames is maintained and the overall processing times are optimized.

• à reconstruction parfaite, c'est-à-dire que la somme des amplitudes de deux fenêtres se recouvrant est toujours égale à 1 (sur la partie qui se recouvre) ;
• une fenêtre de longueur 256 avec un recouvrement de 96 de chaque coté.

According to the invention, a formatting window (adapted to associated voice frames of 160 samples and to 256-point FFTs) is preferably:

• with perfect reconstruction, that is to say that the sum of the amplitudes of two overlapping windows is always equal to 1 (on the overlapping part);
• a window of length 256 with a cover of 96 on each side.

Such a window is, for example, obtained by the convolution of a Hanning window of width 97 (denoted Hanning (97)) with a rectangular window of width 160 (denoted Rect (160)).

A 256-point FFT is then applied to each 256-sample window synchronized to frames of 160 samples. The implementation of FFT is well known to those skilled in the art and is particularly detailed in the book "Numerical Recipes in C, 2 ^nd edition" (or in French "Digital Recipes in C language, ^2nd edition") written by Press WH, Teukolsky SA, Vetterling WT and Flannery BP and published in 1992 by Cambridge University Press.

Then, we apply a noise reduction algorithm, of any type known per se, before performing an inverse transform operation (denoted IFFT) on the block of 256 samples considered.

Blocks of 256 samples are thus treated successively. After the IFFT operation, the first 96 processed samples of the current window are added to the last 96 processed samples from the previous window. After addition, the first 160 samples of the current window are transmitted to the vocoder to be processed according to the speech coding methods known per se, in accordance, where appropriate, with the applicable standard.

In connection with FIG. 1, a radiotelephone embodying the invention is presented.

Figure 1 schematically illustrates a general block diagram of a radiotelephone, according to the invention according to a preferred embodiment.

• un microphone 107 ;
• un convertisseur Analogique/Numérique 108 ;
• un haut-parleur 109 ;
• un convertisseur Numérique/Analogique 110 ;
• un processeur de traitement du signal (DSP) 104 ;
• une mémoire non volatile 105 ;
• une mémoire vive 106 ;
• une interface radio 111 ;
• une unité 112 de gestion et de contrôle des échanges des trames de données et de protocoles ; et
• une interface de relation homme/machine (typiquement un clavier et un écran) 113.

The radiotelephone 100 comprises interconnected by an address and data bus 103:

• a microphone 107;
• an Analog / Digital converter 108;
• a speaker 109;
• a digital / analog converter 110;
• a signal processing processor (DSP) 104;
• a non-volatile memory 105;
• a random access memory 106;
• a radio interface 111;
• a unit 112 for managing and controlling exchanges of data frames and protocols; and
• a man / machine interface (typically a keyboard and a screen) 113.

Each of the elements illustrated in Figure 1 is well known to those skilled in the art. These common elements are not described here.

It is further observed that the word "register" used throughout the description designates in each of the memories mentioned, as well a low capacitance memory area (a few binary data) a large capacity memory area (for storing a program whole or all of a sequence of transaction data).

• le programme de fonctionnement du DSP 104 dans un registre « prog » 308 ;
• une valeur L (valant typiquement 256), représentant une première taille de fenêtre de segmentation correspondant à un nombre de points pris en compte par une FFT dans un registre 115 ;
• une valeur L' (valant typiquement 160), représentant une deuxième taille de fenêtre correspondant à une taille de trame traitées par un vocodeur dans un registre 115 ; et
• des valeurs α, β, γ, κ et β _f utilisées pour la réduction de bruit dans le signal.

The nonvolatile memory 105 (or ROM) keeps in registers which for convenience have the same names as the data they retain:

• the operating program of the DSP 104 in a " prog " register 308;
• a value L (typically 256), representing a first segmentation window size corresponding to a number of points taken into account by an FFT in a register 115;
• a value L '(typically 160), representing a second window size corresponding to a frame size processed by a vocoder in a register 115; and
• values α, β, γ, κ and β _f used for noise reduction in the signal.

• un registre 117 dans lequel sont conservées des valeurs d'échantillons bruités du signal reçu ;
• un registre 118 dans lequel sont conservées des valeurs d'échantillons traités ; et
• une séquence d'échantillons traités destinée à un vocodeur.

RAM 106 stores data, variables, and intermediate processing results and includes:

• a register 117 in which noisy sample values of the received signal are stored;
• a register 118 in which are kept processed sample values; and
• a sequence of processed samples for a vocoder.

The DSP is particularly suitable for processing Fourier transform and speech coding. It may be used, for example, a DSP core manufactured by the company "DSP GROUP" (registered trademark) under the reference "OAK" (registered trademark).

Figure 2 illustrates the successive processing carried out by the radiotelephone of Figure 1, on a voice signal.

• d'un signal vocal pouvant être affecté d'un écho (symbolisé par la somme du signal produit 200 et du signal produit retardé) ; et
• d'un bruit 202

Note that the signal entering the microphone 107 is the sum 203:

• an echo-capable speech signal (symbolized by the sum of the produced signal 200 and the delayed product signal); and
• of a noise 202

The noisy signal picked up by the microphone 107 is delivered to the Analog / Digital converter 204 where it is converted into a series of digital samples during a step 204. According to the GSM standard, it is noted that the sampling is typically done at frequency equal to 8kHz.

Then, in a step 205, the digital sample sequence is processed.

Then, in a step 206, L 'frames (160) of processed samples are encoded by a vocoder according to a method known per se (typically as specified in the GSM standard).

Then, during a step 207, "vocoded" frames are formatted by the unit 112 to be transmitted by the radio module 111 according to techniques known per se (for example, according to the GSM standard).

FIG. 3 presents a cancellation or noise reduction algorithm implemented in the processing step 205 of FIG. 2.

During an initialization step 300, the DSP 104 initializes in the RAM 106, a first block of 96 zero samples corresponding to the last samples received as well as all the variables necessary for the proper operation of the processing 205.

Then during a step 301, the DSP 104 stores in the RAM 106 following the previously received samples a sequence of 160 incoming samples from the converter 108.

Then, in a step 302, the DSP 104 applies a segmentation window of length 256 to the sequence of the last 256 received samples. (Note that this window is illustrated below with reference to Figure 7)

A mathematical transformation of 256-point FFT type is then applied to the sequence obtained by applying the segmentation window.

Then, during a step 303, a noise reduction type processing (specified later with reference to FIG. 8) is applied to the sequence resulting from the mathematical transformation.

Then, during a step 304, a transformation inverse to that of the step 302, of the IFFT type is applied to the processed sequence.

Then, during a step 305, the DSP 104 adds, if necessary (that is to say after a first iteration), the last 96 samples of the preceding processed sequence to the first 96 processed samples of the current sequence. .

Then, during a step 306, the sequence or frame formed of the first 160 processed current samples is transmitted to the vocoder.

Then, during a step 307, the 160 received samples corresponding to the 160 samples transmitted during the step 305 are erased from the memory 106.

Then, step 301 is repeated.

FIG. 4 presents a coding of the speech, implemented in step 206 of FIG.

During an initialization step 400, the DSP 104 initializes in the RAM 106, all the variables necessary for the proper functioning of the coding 206.

Then during a step 401, the DSP 104 stores in the RAM 106 a frame of 160 samples transmitted during the step 307.

Then, during a step 402, the DSP 104 applies speech coding processing to the frame of 160 samples according to a technique known per se.

Then, during a step 403, the coded frame is formatted and transmitted to the unit 102 to be transmitted to a recipient.

Then, during a step 404, the frame of 160 samples is erased from the RAM 106.

Then, operation 401 is reiterated.

Figure 5 depicts a window of the sample sequences as performed by the treatments of Figures 3 and 4.

In a first graph, the curve 500 of the intensity 503 of the signal received directly from the converter 108 as a function of time t 502 is shown.

On a second graph, the curve 500 of the intensity 504 of the signal processed in step 205 is shown as a function of time t 502.

It is noted in the first graph that the time is divided into successive windows 505 and 506 of length L equal to 256, overlapping over a length L "equal to 96 and obtained during step 302.

It is also noted on the second graph that the time is divided into successive frames 507 and 508 of length L 'equal to 160, which do not overlap and are obtained during the transmission step 306.

The segmentation of the signal is such that the windows 505 (respectively 506) and 507 (respectively 502) are perfectly synchronous.

Thus, according to the preferred embodiment, the windows 505 (respectively 506) and 507 (respectively 502) end on the same sample before or after treatment (according to the steps 303, 304 and 305).

In this way, the overlap is on a length equal to L ".

Figure 6 illustrates a shaping window known per se.

The graph giving the amplitude 602 of a window as a function of the rank of a sample 601 is represented by Hanning windows 603 and 604 of length 256 with an overlap of 128.

Note that according to this division known per se, the windowing can in no way be synchronous with a frame segmentation of 160 samples.

FIG. 7 illustrates shaping windows 700 and 701, optimized according to the invention (corresponding to the windows 505 and 506 respectively of FIG. 5, but represented more precisely).

As before, the graph gives the amplitude 602 of a window as a function of the rank of a sample 601.

It is noted that windows 700 and 701 are Hanning windows obtained by convolution of an intermediate Hanning window of length 97 with a rectangular window of length 160. Thus, with the successive offsets of the windows, 160 parallels are obtained. windows with perfect reconstruction.

FIG. 8 specifies step 303 of noise reduction type processing as illustrated with reference to FIG. 3.

• « Spectral substraction based on minimum statistics » (en français « soustraction spectrale basée sur des statistiques minimum ») écrit par R. Martin et publié dans le document « Signal Processing VII : Théories and applications, 1994, EURASIP » aux pages 1182 à 1185 ;
• « Computationally efficient speech enhancement by spectral minima tracking in subbands » (en français « amélioration de la parole efficace pour le calcul par la recherche de minima spectraux dans des sous-bandes »), écrit par G. DOBLINGER et publié dans les comptes-rendus (pages 1513 à 1516) de la conférence « ESCA. EUROPSPEECH'95, 4th European Conference on speech communication and technology » ; et
• « A combination of noise reduction and improved echo cancellation » (en français « une combinaison de réduction de bruit et d'annulation d'écho améliorée ») publié en Allemagne dans la collection « Fachgebiet Theorie der Signale » par l'université de technologie de Darmstadt .

This noise reduction treatment is described in particular in the following documents:

• "Spectral substraction based on minimum statistics" (in French "spectral subtraction based on minimum statistics") written by R. Martin and published in the document "Signal Processing VII: Theories and Applications, 1994, EURASIP" on pages 1182 to 1185 ;
• "Computationally efficient speech enhancement by spectral minima tracking in subbands", written by G. DOBLINGER and published in the Proceedings (pages 1513 to 1516) of the conference "ESCA. EUROPSPEECH'95, 4th European Conference on speech communication and technology " and
• "A combination of noise reduction and improved echo cancellation" published in Germany in the "Fachgebiet Theorie der Signal" collection by the University of Technology of Darmstadt .

After being processed according to step 302, a frame 801 having 256 spectral components corresponding to a noisy speech signal is processed according to the treatment 303 described hereinafter.

We denote X _k (m) the ^kth component of the m ^th frame of the noisy speech signal.

During an operation 802, the DSP 104 converts the components of the rectangular coordinate frame 801 to polar coordinates to separate the phase of the spectral amplitude.

During the different treatments, only the spectral amplitude will be modified, the phase remaining unchanged.

pour m>1
avec une valeur pour le coefficient « d'oubli » α comprise entre 0,7 et 0,9 ce qui permet d'assurer une recherche adéquate du spectre de parole stationnaire à courts termes.During a step 803, the power P _xk (m) of the short-term signal is first estimated according to the following relationships:
P _xk (1) = (1- α) | X _k ( 1 ) | ² (to which a correction value may be added to improve the speed of convergence of the estimate); $$P_{\mathit{xk}}\left(m\right)={\mathit{\alpha }\mathit{P}}_{\mathit{xk}}\left(m-\mathit{1}\right)+\left(1-\mathit{\alpha }\right)|X_{\mathrm{k}}\left(m\right){|}^2$$

for m > 1
with a value for the "forgetting" coefficient α of between 0.7 and 0.9, which makes it possible to ensure an adequate search of the stationary short-term speech spectrum.

• leur simplicité de calcul ; et
• le fait qu'aucun délai de mesure n'est introduit.

These relationships have two advantages:

• their simplicity of calculation; and
• the fact that no measurement delay is introduced.

According to an alternative embodiment, an improved noise reduction algorithm is used. Nevertheless, the introduction of additional delay in this algorithm required a larger memory size for the storage of spectral components with complex values.

et pour m strictement supérieur à 1 (m>1) :
si P_nk (m-1) < P_xk (m)
alors $P_{\mathit{nk}}\left(m\right)={\mathit{\gamma P}}_{\mathit{nk}}\left(m-1\right)+\frac{1-\mathit{\gamma }}{1-\mathit{\beta }}\left(P_{\mathit{nk}}\left(m\right)-\mathit{\beta }{P}_{\mathit{xk}}\left(m-1\right)\right);$

sinon P_nk(m) = P_xk(m) ; Then, the spectral power P _nk (m) of the noise is estimated according to the following nonlinear estimator (which does some sort of search for the temporal minima of P _xk (m)) : $${\mathit{P}}_{\mathit{nk}}\left(\mathit{1}\right)\mathit{=}{\mathit{<figure-callout\; id="1"\; label="P\; xk"\; filenames="imgf0002.png,imgf0005.png"\; state="\{\{state\}\}">P\; </figure-callout>}}_{\mathit{xk}}\left(\mathit{1}\right)\mathit{;}$$

and for m strictly greater than 1 ( m > 1):
if P _nk ( m -1) < P _xk ( m )
so $P_{\mathit{nk}}\left(m\right)={\mathit{\gamma P}}_{\mathit{nk}}\left(m-1\right)+\frac{1-\mathit{\gamma }}{1-\mathit{\beta }}\left(P_{\mathit{nk}}\left(m\right)-\mathit{\beta }{P}_{\mathit{xk}}\left(m-1\right)\right);$

otherwise P _nk (m) = P _xk (m);

si g_k(m) > β _f
et g_k(m) = β _f sinonThen, during a step 806, the DSP 104 calculates a gain factor g _k (m) at real values according to the following relationships: $${\mathrm{boy\; Wut}}_k\left(m\right)=1-\sqrt{\frac{{\mathit{\kappa P}}_{\mathit{nk}}\left(m\right)}{{\mathit{P}}_{\mathit{xk}}\left(m\right)}}$$

if g _k (m) > β _f
and g _k (m) = β _f otherwise

The coefficient κ is an overestimation factor of the noise that is introduced to obtain better performances of the noise reduction algorithm.

β _f corresponds to a spectral value floor. β _f limits the attenuation of the noise reduction filter to a positive value to leave a minimum noise in the signal.

Then, during a step 807, the DSP 104 multiplies the amplitude | X _k ( m ) | by the corresponding gain factor g _k (m) to obtain the improved signal amplitude | Y _k ( m ) | according to the following relation:
| Y _k ( m ) | = g _k (m) . | X _k ( m ) | for the values of k between 1 and 256.

Then, during a polar-to-rectangular coordinate conversion step 808, the DSP 104 constructs the signal 809 with reduced noise from the amplitude | Y _k ( m ) | determined during step 807 and the phase of the signal extracted during step 802.

The signal 809 is then processed according to the inverse Fourier transformation step 304.

Of course, the invention is not limited to the embodiments mentioned above.

In particular, a person skilled in the art can provide any variant in the application of the invention which is not limited to mobile telephony (in particular GSM, UMTS, IS95, etc.) but extends to any type of device comprising audio coding after or before a mathematical transformation on an incoming audio signal.

In addition, the invention applies not only to the processing of voice source signals but extends to any type of audio processing.

According to the invention, the mathematical transformation applied is in particular of any type applying to blocks of samples of a particular length which is not equal to the size of the frames processed according to an audio processing or which is not a multiple or divider adjacent to this frame size. Thus, the invention extends to the case where the size of the audio frames is equal to 160 or more generally is not a power of 2 and where a mathematical transformation applies to block sizes of length 256, 128, 512 or more generally 2 "(where n represents an integer) in particular an FFT, an FHT (of the English" Fast Hadamard Transform "or, in French" Fast Hadamard Transform ") or a DCT (of the English" Discrete Cosine Transform: "or, in French," transformed into discrete cosine ") or variants of these transformations (obtained, for example, by combining one or more of these transformations with one or more other transformations) ...

In addition, the invention applies to any type of processing associated with the mathematical transformation and performed before or after a speech coding step, particularly in the case of voice recognition or cancellation and / or reduction. echo.

It will be noted that the invention is not limited to a purely material implantation but that it can also be implemented in the form of a sequence of instructions of a computer program or any form mixing a material part and a part software. In the case where the invention is partially or totally implemented in software form, the corresponding instruction sequence can be stored in a removable storage means (such as for example a floppy disk, a CD-ROM or a DVD-ROM) or no, this storage means being partially or completely readable by a computer or a microprocessor.

Claims (11)

1. Method of processing an audio signal, comprising:

   - a first stage (205) of processing a source audio signal, implementing at least one mathematical transformation which is applied to first sequences of samples which are obtained by application of first segmentation windows (505, 506, 700, 701) to the said source audio signal; and

   - a second stage (206) of audio processing, applied to second sequences of samples which are obtained by applying second segmentation windows (507, 508) to the signal which the said first stage supplies, the length of the said second segmentation windows being distinct from the length of the said first segmentation windows;

   characterized in that two successive first windows and/or two successive second windows straddle each other, the straddlings being such that the segmentations are synchronous and the segmentations are synchronised to the end of the said first and second windows.
2. Method according to Claim 1, characterized in that the said second segmentation windows are successive frames.
3. Method according to one of Claims 1 and 2, characterized in that the last sample of a first sequence is also the last sample, after the said first stage, of the corresponding second sequence.
4. Method according to Claims 1 to 3, characterized in that each said first segmentation window (700, 701) is a perfect reconstruction window which is obtained by convolution:

   - of a first intermediate window with perfect reconstruction and having spectral properties which are suitable for the said mathematical transformation(s); and

   - of a second rectangular intermediate window.
5. Method according to one of Claims 1 to 4, characterized in that the said first processing stage, which is applied to each first sequence, additionally comprises:

   - a predetermined processing sub-stage (303) which is applied to the said first sequence;

   - an inverse mathematical transformation sub-stage (304) which is applied to the processed samples of the said first sequence; and

   - a stage of addition (305) of voice samples from the said inverse mathematical transformation sub-stage, applied to the said first sequence, and corresponding voice samples from the said inverse mathematical transformation sub-stage, applied to the preceding first sequence.
6. Method according to Claim 5, characterized in that the said predetermined processing sub-stage includes a reduction or cancellation of noise in the said audio signal.
7. Method according to one of Claims 5 and 6, characterized in that the said predetermined processing sub-stage includes at least one process which is part of the group including:

   - reduction or cancellation of echo in the said audio signal;

   - voice recognition in the said audio signal.
8. Method according to one of Claims 1 to 7, characterized in that the said mathematical transformation(s) belong to the group including:

   - fast Fourier transformations (FFT) and their variants;

   - fast Hadamard transformations (FHT) and their variants; and

   - discrete cosine transformations (DCT) and their variants.
9. Method according to one of Claims 1 to 8, characterized in that the said source audio signal is a voice signal.
10. Device for processing an audio signal, comprising:

    - first means for processing a source audio signal, implementing at least one mathematical transformation which is applied to first sequences of samples which are obtained by application of first segmentation windows to the said source audio signal; and

    - second means for audio processing, applied to second sequences of samples which are obtained by applying second segmentation windows to the signal which the said first stage supplies, the length of the said second segmentation windows being distinct from the length of the said first segmentation windows;

    characterized in that two successive first windows and/or two successive second windows straddle each other, the straddlings being such that the segmentations are synchronous and the segmentations are synchronised to the end of the said first and second windows.
11. Computer program product which can be downloaded from a communication network and/or recorded on a medium which can be read by a computer and/or executed by a processor, characterized in that it includes program code instructions for implementing the method of processing an audio signal according to at least one of Claims 1 to 9.

Images