RU2237296C2 - Method for encoding speech with function for altering comfort noise for increasing reproduction precision - Google Patents

Abstract

FIELD: radio communications technology.

SUBSTANCE: method includes altering values of comfort noise parameters, usually used for generating comfort noise, quality of comfort noise is increased, generation of comfort noise is performed by means of device for speech decoding during time spans, in which speech is absent. Values of comfort noise parameter are altered in response to receiving information about changeability, matching parameter of comfort noise. Altered values of comfort noise parameters are then used for generation of comfort noise.

EFFECT: higher system capacity.

2 cl, 10 dwg

Description

This application for invention, in accordance with Section 35 of USC 119 (e) (1) of the U.S. Code, has priority in the process of simultaneous consideration of provisional patent application US No. 60 / 109,555 with a filing date of November 23, 1998.

FIELD OF TECHNOLOGY

The invention relates, in General, to the process of encoding speech, in particular to a process for encoding speech, in which for periods of time in which there is no speech activity, create artificial background noise.

BACKGROUND OF THE INVENTION

Radio transmitting and receiving devices are usually equipped with speech encoding and decoding devices, respectively, which interact in such a way that they allow the transmission of speech signals between a given transmitter and receiver via a radio link. The combination of a speech encoding device and a speech decoding device is often referred to as a speech codec. A typical example of a communication device, which typically comprises a radio transmitter equipped with a speech encoding device, and a radio receiver equipped with a speech decoding device, is a mobile radiotelephone (e.g., a cellular telephone).

In conventional devices based on block coding of speech, the incoming speech signal is divided into blocks called frames. For regular telephony with a 4 kHz bandwidth, the frame length is usually 20 milliseconds (ms) or 160 samples. Frames are further divided into sub-frames, the length of which is usually 5 ms or 40 samples.

Conventional linear prediction coding devices through synthesis analysis (LPAS) use speech-conditioned models. Model parameters describing the human vocal tract, pitch of the speech signal, etc. extracted from the input speech signal. The calculation of slowly varying parameters is usually carried out for each frame. Examples of such parameters are, among other things, PCB parameters (short-term prediction, which characterize the speech path of the device generating the speech signal. One example of PCB parameters are linear prediction coefficients (PPC), which display the shape of the spectrum of the input speech signal. Examples of more rapidly changing The parameters are, among other things, pitch parameters and shape / gain corrections, which are usually calculated for each subframe.

Quantization of the extracted parameters is carried out using the appropriate known methods of scalar and vector quantization. PCB parameters, for example, linear prediction coefficients, are often converted into a more convenient form for quantization, for example, in the form of frequencies with a linear spectrum (LS). After quantization, the parameters are transmitted over the communication channel to the decoding device.

In a conventional LPAS decoding device, operations opposite to the above are usually performed and a speech signal is synthesized. Subsequent filtering of the synthesized speech signal is usually used to improve perceptual quality.

For many common types of background noise, a reasonably good signal model is obtained at a significantly lower bit rate than that needed for the speech signal. In existing mobile communication systems use this fact and carry out the appropriate adjustment of the bit rate during the presence of background noise. In conventional systems that employ continuous transmission methods, the variable bit rate (CSP) speech coding apparatus can use the lowest bit rate. In conventional discontinuous transmission (PRP) schemes, the transmitter stops transmitting encoded frames of the speech signal when there is no talking activity. At identical or unequal intervals (usually every 500 ms), the transmitter transmits the appropriate parameters of the speech signal to generate comfortable noise in the decoding device. The coding of these parameters to generate comfortable noise (GKSH) is usually carried out in the form of frames, which are sometimes called frames of the silence descriptor (DTSH). The decoding device in the receiver uses the comfort noise parameters adopted in the DTS frames to synthesize artificial noise using a conventional algorithm for introducing comfort noise (VCS).

After comfortable noise is generated in the decoding device of a conventional PDP system, noise is usually perceived as highly static and very different from background noise generated in the active (non-PDP) mode. The reason for this perception is that the transmission of DTL frames during PDP is performed less frequently than ordinary frames of a speech signal. In LPAS codecs-decoders having a PDP mode, the spectrum and energy of background noise are usually estimated (for example, averaged) over several frames, and then the calculated parameters are quantized and transmitted over the communication channel to the decoding device. Figure 1 shows an example of a known device for encoding comfort noise, which creates the above estimated parameters of the background noise (comfort noise). The transmission of quantized comfort noise parameters is usually carried out every 100-500 ms.

Transmitting DTS frames with a low update rate instead of transmitting conventional speech frames has a double advantage. Due to the reduction in power consumption, the battery life is extended, for example, in a mobile radio transceiver, and the interference caused by the transmitter is reduced, thereby providing a higher system throughput.

In a conventional decoding device, reception and decoding of comfort noise parameters can be implemented as shown in FIG. Since the decoding device does not accept new comfort noise parameters as often as the parameters of the speech signal are received, in order to ensure a smooth change in the parameters during the synthesis of comfort noise, the block 23 usually interpolates the comfort noise parameters received in the DTS frames. In the synthesis operation, denoted by 25, the decoding device introduces into the synthesis filter 27 excitation in the form of a gain-normalized random noise (eg, white noise) and interpolated spectrum parameters. As a result, the created comfort noise S _c (n) will be perceived as highly stationary (“static”) regardless of the change in the properties of the background noise s (n) in the encoding device (see FIG. 1). This problem is more pronounced for highly changing background noise, such as street noise and inarticulate noise from many sources (for example, noise in a restaurant), but it also occurs in the case of car noise.

One of the traditional approaches to solving this problem of “static” comfort noise is to simply increase the update rate of comfort noise parameters during PDP (for example, using a higher transmission rate of DTS frames). With such a solution, problems arise such as an increase in battery power consumption (for example, in a mobile communications transceiver), due to the need for more frequent operation of the transmitter, and a decrease in system throughput due to an increase in the transmission speed of DTS frames. Therefore, in traditional systems, they usually agree with the inevitability of static background noise.

Therefore, it is desirable to eliminate the above disadvantages associated with the traditional method of generating comfortable noise.

According to the invention, comfort noise parameters obtained by the traditional generation method are changed based on their properties of real background noise perceived by the encoding device. Comfort noise generated from the changed parameters is perceived as less static compared to comfort noise generated in the traditional way and more similar to real background noise existing in the encoding device.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a schematic representation of a process for obtaining comfort noise parameters in a conventional speech encoding device.

Figure 2 - schematic representation of the process of generating comfortable noise in a conventional speech decoding device.

Figure 3 is a block diagram of a device for changing comfort noise parameters used in generating comfort noise according to the invention.

Figure 4 is an example of an embodiment of the change device of Figure 3.

5 is an example of an embodiment of the variability estimation apparatus of FIG. 4.

FIG. 5 is an example of a SELECT signal control device of FIG. 5.

FIG. 6 is an example embodiment of the change device of FIG. 3 to FIG. 5, wherein the variability estimation device of FIG. 5 is implemented in part in an encoding device and in part in a decoding device.

Fig.7 is an exemplary embodiment of operations that can be performed by the change device of Fig.3 - Fig.6.

Fig. 8 is an example implementation of the evaluation calculation step of Fig. 7.

Fig.9 is a block diagram of a voice communication system in which embodiments of the change device of Fig.3 - Fig.8 can be implemented.

DETAILED DESCRIPTION

Figure 3 shows a device 30 for changing comfort noise parameters for changing comfort noise parameters according to the invention. According to figure 3, the input 33 of the device changes 30 serves the usual interpolated parameters of comfort noise, such as spectral and energy parameters obtained from the output of the interpolator 23 of figure 2. Also at the input 31 of the device 30 changes provide spectral and energy parameters corresponding to the background noise entering the encoding device. The change device 30 changes the obtained comfort noise parameters based on the background noise parameters obtained from the input 31, and creates the modified comfort noise parameters at the output 35. The modified comfort noise parameters can then be fed, for example, to the comfort noise synthesis unit 25 of FIG. 2 for use in conventional comfort noise synthesis operations. The modified comfort noise parameters generated at the output 35 allow the comfort noise generation in the synthesis unit 25 to more accurately reproduce the actual background noise entering the speech encoding device.

Figure 4 shows an example of an embodiment of a device 30 for changing comfort noise parameters of Figure 3. The change device 30 comprises a variability estimation device 41 connected to an input 31 that receives spectral and energy parameters of background noise. The device 41 estimates variability evaluates the characteristics of the variability of the parameters of the background noise and outputs 43 the information characterizing the variability of the parameters of the background noise. The variability information may indicate the degree of variability of the parameter relative to its average value, for example, the variance of the parameter or the maximum deviation of the parameter from its average value.

The information on the variability obtained at the output 43 may also indicate the correlation characteristics, the evolution of the parameter in time, or other criteria for assessing the variability of the parameter in time. Examples of information on time variability can be given, including simple evaluation criteria, such as the rate of change of a parameter (fast or slow changes), the variance of a parameter, the maximum deviation from the average value, other statistical evaluation criteria characterizing the variability of a parameter, and more complex evaluation criteria, such as autocorrelation characteristics and filter coefficients of a predictor with automatic regression (AR), the estimated calculation of which is performed but starting from the parameter values. One example of a simple criterion for evaluating the rate of change is to calculate the frequency of crossing through zero, that is, how many times the parameter changes sign when viewing a sequence of parameter values from the first parameter value to the last parameter value. The information obtained at the output 43 of the evaluation device 41 is fed to the input of the combining device 45, which combines the information received at the output 43 with the interpolated comfort noise parameters received at the input 33 to form the modified comfort noise parameters at the output 35.

FIG. 5 shows an example embodiment of the variability estimation apparatus 41 of FIG. 4. The evaluation device of FIG. 5 comprises an average variability determination device 51 connected to input 31 for receiving spectral and energy parameters of background noise. As described above, the apparatus 51 for determining the average value of the variability can determine the average values of the characteristics of the variability.

For example, if the background noise buffer 37 of FIG. 3 contains 8 frames and 32 subframes, then the variability analysis of buffered spectral and energy parameters can be performed as follows. The calculation of the average (or average) value of the buffered spectral parameters (as is usually done in PRP coding devices to create DTT frames) can be performed and its subtraction from the buffered values of the spectral parameter, whereby a vector of spectrum deviation values is obtained. In the same way, the calculation of the average value of the buffered energy parameters over a subframe can be performed (as is usually done in PRP coding devices to create DTT frames), and then its subtraction from the buffered values of the energy parameter in a subframe, whereby the energy deviation vector is obtained. Consequently, the spectrum and energy deviation vectors contain spectral and energy parameters with the mean value excluded. The spectrum and energy deviation vectors are transmitted from the variability determination device 51 to the storage device 55 for the deviation vector via the communication path 52.

In order that the coefficient calculating device 53 can obtain the background noise parameters, it is also connected to the input 31. An exemplary coefficient calculating device 53 performs conventional evaluative calculations by the AP method using the corresponding spectral and energy parameters. The filter coefficients obtained from the estimated calculations by the AP method are transmitted from the coefficient calculation device 53 to the filter 57 through the communication path 54. By means of the filtering coefficients calculated in the device 53, for example, corresponding pole filters for spectral and energy parameters can be set.

In one embodiment, the coefficient calculator 53 performs first-order AP estimation calculations for both spectral and energy parameters, calculating filter coefficients a1 = Rxx (1) / Rxx (0) for each parameter in the usual way. The values of Rxx (0) and Rxx (1) are the usual values of the autocorrelation function of a particular parameter:

In these calculation operations, Rxx is a background noise parameter (e.g. spectral or energetic). A positive value of a1 in the general case means that the parameter changes slowly, and a negative value in the general case means that it changes quickly.

According to one embodiment, for each frame of spectral parameters and for each subframe of energy parameters, for example, the component x (k) of the corresponding deviation vector can be selected randomly (by means of the SELECT input signal in memory 55) and filtered by a filter 57 s using appropriate filtering coefficients. Then, the output signal from the filter is normalized by multiplying it by a constant scaling factor by means of a scaling device 59, for example, a multiplying device. The normalized output signal, indicated in Fig. 5 as xp (k), is supplied to the input 43 of the combining device 45 in Fig. 4.

In one embodiment, which is shown schematically in FIG. 5A, the zero crossing frequency determination device 50 is connected to an input 31 for receiving buffered parameters from block 37. The determining device 50 determines the corresponding zero and frequency transition frequencies of the spectral and energy parameters. That is, for the sequence of energy parameters accumulated in the buffer 37, as well as for the sequence of spectral parameters accumulated in the buffer 37, the zero crossing frequency determination device 50 determines for the corresponding sequence how many times the value of the corresponding parameter changes sign when viewing the buffered sequence from the first parameter values to the last parameter value. Then, this information on the zero-crossing frequency can be used in block 56 to control the SELECT signal from FIG. 5.

For example, for a given deviation vector, the control of the SELECT signal can be performed to randomly select the components x (k) of the deviation vector relatively more often (with the frequency of each frame or subframe), if the zero crossing frequency corresponding to this parameter is relatively high (which indicates greater variability of the parameter), and to randomly select the components x (k) of the deviation vector, it is relatively less frequent (for example, with a lower frequency than the frequency of each frame or subframe) moreover, if the frequency of crossing through zero corresponding to this parameter is relatively low (which indicates a very small variability of the parameter). In other embodiments, the frequency of component selection x (k) of a given deviation vector may be set to a predetermined desired value.

The combining device of FIG. 4 combines the normalized output signal xp (k) with the usual comfort noise parameters. The combination of spectral parameters is carried out on a frame, and energy parameters on a subframe. In one example, the combiner 45 may be an adder that simply adds the signal xp (k) to the normal comfort noise parameters. Therefore, the normalized output signal xp (k) (Fig. 5) can be considered as a disturbing signal, which is used in the combining device 45 to exert a disturbing effect on the usual comfort noise parameters received at input 33 to form altered (or resulting disturbing effect) of the comfort noise parameters supplied to the input of the comfort noise synthesis unit 25 (see Fig. 2 - Fig. 4).

The use of comfort noise parameters obtained as a result of the disturbing effect in a conventional synthesis unit 25 can be performed in a conventional manner. Due to the perturbation effect on the usual parameters, the generated comfortable noise will change in a quasi-random manner, which significantly improves the perception quality for a very changing background, for example, for slurred noise from many sources and street noise, as well as car noise. As one example, the disturbing signal xp (k) can be expressed as follows:

xp (k) = β _x · (b0 _x · x (k) –a1 _x · γ _x · (xp (k-1)),

where β _x is the scale factor, b0 _x and a1 _x are the filtering coefficients, and γ _x is the coefficient of expansion of the frequency band.

The dotted line in FIG. 5 shows an embodiment in which the filtering operation is omitted, and the disturbing signal xp (k) contains the normalized components of the deviation vector.

In some embodiments, the implementation of the change device 30 of FIG. 3 to FIG. 5 is completely implemented as part of a speech decoding device (see FIG. 9), and in other embodiments, the implementation of the change device of FIG. speech decoding (see dashed lines in FIG. 9). In those embodiments in which the change device 30 is completely implemented in the decoding device, the decoding device should recognize the background noise parameters shown in FIG. This can be done by buffering in block 37 the required number (frames and subframes) of spectral and energy parameters obtained from the encoding device through the transmission channel. In the PDP scheme, in order to decide that the buffer 37 contains only the parameters associated with background noise, the implicit information typically found in the decoding device can be used. For example, if the buffer 37 can buffer N frames, and if in the PDP mode, after N speech fragments and before the transmission is interrupted, N blocking frames are used during the transition (which is common), then, as is known, these last N frames before switching The PDP mode contains only the spectral and energy parameters of background noise. As described above, these background noise parameters can then be used in the change device 30.

In those embodiments in which the change device 30 is distributed between the encoding device and the decoding device, the average variability value determining device 51 and the coefficient calculating device 53 may be provided in the encoding device. Therefore, in such embodiments, the communication paths 52 and 54 are similar to the conventional communication channel used to transmit conventional comfort noise parameters from the encoding device to the decoding device (see FIGS. 1 and 2). In particular, as shown in FIG. 6, paths 52 and 54 pass through a quantization device (see also FIG. 1), a communication channel (see also FIGS. 1 and 2) and an inverse quantization operation block (see also figure 2), to the storage device 55 and to the filter 57 (see also figure 5). As for the information on the average value of the variability and AR filter coefficients, known methods for quantizing scalar values and AR filter coefficients can be applied.

Using conventional means in the encoding device, it is known that the spectral and energy parameters of background noise are available for processing by means of the average variability determination device 51 and coefficient calculating device 53, since the same spectral and energy parameters are usually used in the encoding device to create ordinary parameters comfortable noise. In traditional encoding devices, the average energy value and the average spectral distribution value are usually calculated over several frames and these average spectral and energy parameters are transmitted to the decoding device as comfort noise parameters. Since it is required to transmit filter coefficients obtained from coefficient calculator 53 and deviation vectors obtained from mean variability determination device 51 from an encoding device to a decoding device over a transmission channel, as shown in FIG. 6, it is necessary to provide additional bandwidth in the case where the change device is distributed between the encoding device and the decoding device. Otherwise, when the change device is completely implemented in the decoding device, no additional bandwidth extension is required to implement it.

7 shows examples of the above operations that can be performed by the change device in the embodiments of FIGS. 3-5. At step 71, it is determined whether the available spectral and energy parameters (for example, in buffer 37 of FIG. 3) correspond to speech or background noise. If the available parameters correspond to background noise, then at step 73, estimated values of the background noise characteristics are calculated, for example, the average value of the variability and the degree of variability over time. After that, at step 75, a disturbance is introduced into the interpolated comfort noise parameters in accordance with the estimated values of the background noise characteristics. The perturbation process in step 75 is performed until, in step 77, background noise is detected. If speech activity is detected in step 77, then go to step 71, where they expect to receive additional background noise parameters.

On Fig shows an example of operations that can be performed when performing at step 73 of the evaluation calculation (Fig.7). In the processing in step 81, N frames and kN subframes that correspond to the aforementioned N buffered frames are taken into account. In one embodiment, N = 8, a k = 4. At step 83, a spectrum deviation vector having an N component is obtained, and at step 85, an energy deviation vector having a kN component is obtained. At step 87, a component is selected (e.g., randomly) from each deviation vector. At 89, filter coefficients are calculated and the selected components of the vectors are filtered accordingly. At step 88, the filtered vector components are normalized, creating a disturbing signal, which is used at step 75 (Fig.7). The dashed line in Fig. 8 corresponds to the embodiments indicated in Fig. 5 by a dashed line, that is, embodiments in which filtering is omitted, and normalized components of the deviation vector are used as perturbation parameters.

FIG. 9 shows an example of a voice communication system in which embodiments of the device for changing comfort noise parameters of FIGS. 3 to 8 can be implemented. The transmitter (TX) comprises a speech encoding device 91 that communicates with a voice decoding device 93 in a receiver ( CRM) through transmission channel 95. The transmitter of FIG. 9, the receiver of FIG. 9, or both of them may be part of, for example, a radiotelephone or other component of a radio communication system. Channel 95 may comprise, for example, a radio channel. As shown in FIG. 9, the embodiments of the change device of FIGS. 3-8 can be implemented in a decoding device, or, as described above with reference to FIGS. 5 and 6, can be distributed between the encoding device and the decoding device (see dashed lines).

It will be apparent to those skilled in the art that the embodiments shown in FIGS. 3 through 9 can be easily implemented in conventional speech encoding-decoding devices, for example, by appropriate modifications of software and / or hardware.

The invention described above improves the naturalness of background noise (without highlighting additional bandwidth or additional power costs in some embodiments). This allows you to implement in devices encoding-decoding of speech more invisible and, therefore, more acceptable for the human ear, switching between the modes of speech presence and absence of speech.

Although the detailed description has been given above for embodiments of the present invention, serving as examples, this does not limit the scope of the invention, which can be practiced in various embodiments.

Claims (26)

1. A method for generating comfort noise in a speech decoding device (93), which receives speech signal and noise information from a communication channel (95), including generating a plurality of values (33) of spectral and energy parameters of comfort noise typically used in a decoding device (93) speech for generating comfortable noise, characterized in that it includes the steps in which information (31) on the variability is obtained, which indicates the degree of variability of the spectral and energy parameters (37) of the background noise experienced the coding device, in response to the variability information, change (30) the values (33) of the spectral and energy parameters of the comfort noise to generate changed values (35) of the spectral and energy parameters of the comfort noise and use the changed values (35) of the spectral and energy parameters of the comfort noise to generate comfortable noise (25), and the variability information indicates how the spectral and energy parameters of the background noise change with respect to at least one th of the following variables: time and the average values of the background noise parameters. 2. The method according to claim 1, characterized in that said step of obtaining includes calculating average values (51) of spectral and energy parameters of background noise from a plurality of values of spectral and energy parameters (37) of background noise and subtracting the average value from each corresponding parameter value ( 37) background noise to form a plurality of deviation values (55). 3. The method according to claim 2, characterized in that said step of changing includes randomly selecting one of the mentioned deviation values, scaling the random deviation value by multiplying by a scale factor to form a scaled deviation value, and combining the scaled deviation value with the corresponding value ( 33) comfort noise parameter to form one of the changed values (35) of the comfort noise parameter. 4. The method according to claim 1, characterized in that the speech decoding device (93) is provided in the radio communication device. 5. The method according to claim 4, characterized in that the device (93) for decoding speech is provided in a cell phone. 6. The method according to claim 1, characterized in that the said receiving step includes receiving information (31) on the variability in the speech decoding apparatus (93) independently of the communication channel (95). 7. The method according to claim 1, characterized in that the said receiving step includes receiving, by the speech decoding apparatus (93), information on variability from the speech encoding apparatus (91) through the communication channel (95). 8. The method according to claim 1, characterized in that the variability information (31) includes variability information with respect to the average value indicating how the spectral and energy parameters of the background noise change relative to the average values of the spectral and energy parameters (37) of the background noise . 9. The method according to claim 8, characterized in that said obtaining step includes using a plurality of values of each of the background noise parameters (37) to calculate average values (51) of the background noise parameters (37) over a time interval and comparing average values of at least with some of the corresponding values (37) of the background noise parameters for generating background noise parameter values with the average value excluded. 10. The method according to claim 9, characterized in that said step of obtaining comprises using a plurality of background noise parameter values (37) to calculate filtering coefficients (89) and filtering at least some of the background noise parameter values (37) with excluded average value in accordance with the filtering coefficients. 11. The method according to claim 10, characterized in that the said use phase includes calculating filter coefficients of a filter of a predictive device with automatic regression. 12. The method according to claim 8, characterized in that the variability information (31) includes time variability information indicating how the background noise parameters (37) change over time. 13. The method according to claim 1, characterized in that the variability information (31) includes time variability information indicating how the background noise parameters change over time. 14. Device for generating comfort noise parameters (33) used to generate comfort noise in a speech decoding device (93) that receives speech signal and noise information from a communication channel (95) containing a first input (33) for supplying a plurality of values ( 33) spectral and energy parameters of comfortable noise, usually used in a speech decoding device (93) for generating comfortable noise, and a second input (31) for supplying spectral and energy parameters (37) of background noise, characterized in that It contains a change device (30) connected to the first (33) and second (31) inputs and responsive to the variability characteristics of the spectral and energy parameters (51) of the background noise, for changing values (33) of the spectral and energy parameters of the comfort noise to form altered the values (30) of the spectral and energy parameters of the comfort noise, and the output (35) connected to the device (30) changes for issuing the changed values (30) of the spectral and energy parameters of the comfort noise for their use and generating comfortable noise, wherein the variability characteristics of the spectral and energy parameters of the background noise indicate how the spectral and energy parameters of the background noise change with respect to at least one of the following values: time and average values of the background noise parameters. 15. The device according to 14, characterized in that the change device (30) comprises a variability estimation device (41) connected to a second input (31) and responsive to background noise parameters (37) to generate said variability information. 16. The device according to p. 15, characterized in that the device for assessing variability contains a device (51) for determining the average value of variability for generating information (31) about variability relative to the average value indicating how the background noise parameters (37) change with respect to the average values of the parameters (37) background noise. 17. The device according to clause 16, wherein the device (51) for determining the average value of the variability is contained in the speech decoding device (93). 18. The device according to p. 16, characterized in that the device (51) for determining the average value of the variability is contained in the speech encoding device (91), which communicates with the speech decoding device (93) through the communication channel (95). 19. The device according to clause 16, characterized in that the device (51) for determining the average value of the variability in response to receiving a plurality of background noise parameter values (37) calculates the average values of the background noise parameters (37) over a time interval (51) and additionally compares the average values with at least some of the values of the background noise parameters (37) to generate the values of the background noise parameters (37) with the excluded average value. 20. The device according to claim 19, characterized in that the variability information (31) includes time variability information indicating how the background noise parameters (37) change over time. 21. The device according to claim 20, characterized in that the variability estimation device (41) comprises a coefficient calculation device (53), which, in response to receiving a plurality of background noise parameter values (37), calculates filtering coefficients (89), said information Variability over time includes filter coefficients. 22. The device according to item 21, wherein the filter coefficients are filter coefficients of a filter of a predictive device with automatic regression. 23. The device according to item 21, characterized in that it contains a filter (57) connected to a device (53) for calculating coefficients for receiving filter coefficients from it and connected to a device (51) for determining the average value of the variability for filtering, at least some of the values of the background noise parameters with the excluded average value in accordance with the filtering coefficients. 24. The device according to item 21, wherein the coefficient calculation device is contained in a speech decoding device. 25. The device according to item 21, wherein the coefficient calculation device (53) is contained in a speech encoding device (91) that communicates with a speech decoding device (93) through a communication channel (95). 26. The device according to p. 15, characterized in that the information (31) about the variability includes information about the variability in time, indicating how the parameters of the background noise change over time.

Images