JPH02500620A - coded communication system - Google Patents

Abstract

In an LPC type coded communications system the excitation source is derived from previous filter outputs at the decoder; in one embodiment the speech output is used, in other embodiments, intermediate excitation outputs are used. To enable tracking the coder derives the filter parameters by using the same excitations, supplied by a local decoder, to synthesise locally the actual error produced at the decoder; the parameters are optimised iteratively by varying the delay of an FIR stage and deriving the actual error in a loop, and selecting the delay for minimum actual error. The IIR parameters may be calculated jointly with the other FIR parameters inside the loop, either for minimum prediction error or minimum actual error. The FIR stage may comprise several parallel FIRs, separately excited.

Description

[Detailed description of the invention] name of invention coded communication system The present invention relates to a system for transmitting coded audio (speech) signals. . The present invention also provides a transmitter/receiver t= Regarding machines.

In many conventional systems that transmit coded audio signals, the coding section extracts the characteristics of the filter from the input audio sample and sends it to part of the decoder. Some rob coders use the filter characteristics sent to them as decoding filters. It is used to configure the This decoding filter then uses the appropriate excitation It is excited by a signal source to generate a synthetic reproduction signal for the input audio signal.

In linear pre-el coding (LPG), the short-period spectral envelope of the audio signal is An all-pole liquor model using a finite number of periodically updated coefficients to model A sibu filter is used. These coefficients typically minimize the “prediction error” It is computed directly by solving a set of linear equations devised to give

(The above prediction error is the difference between the input audio signal and the predicted audio signal. Refers to measured value. ) In some LPG systems, the speaker's pitch can be determined from the above difference signal. Including a prediction stage on the coding side to remove the long periodicity corresponding to It's set. This prediction stage can be thought of as a type of filter, corresponding to Filter parameters are also sent to part of the decoder.

The prediction error is related to whether or not the predicted speech signal includes a pitch prediction. Regardless, the error will be different from the actual error. This is because this prediction is based on the excitation model. based on the actual excitation in some part of the decoder. It is the body.

Modern LPG systems include, for example, multiple pumping (MP), regular pulse pumping’ (RPE), and codebook excitation (CE) types of LPG systems. There is. The excitation used in part of the decoder is selected or be guided. In such systems, part of the decoder is a controllable excitation signal generator. In that case, the encoding section is the one that transmits the control signal to the decoder part. Must.

Some decoders are therefore equipped with two-stage filters (relatively short delay all-pole filters). (a filter with a relatively long delay) and an excitation generator.

The excitation control signal itself is generated in the coding section.

In MP type LPG, this is achieved through "analysis by synthesis." That is, The unmodified excitation pulse sequence is used locally in the synthesized speech to generate the actual error signal. Subtract the synthesized speech sequence from the actual input signal to create an audible error signal. Then, the weighted error signal is used as a constraint in an optimization closed loop. control to select the excitation sequence that minimizes the (weighted) actual error. do.

The present invention provides a method for encoding speech. This encoding method uses filter parameters. parameters are periodically extracted from the input audio signal of the encoder. The derived f Filter parameters are used to update the response of some filters in the decoder. Sent to part of the coder. In addition, the excitation input to the decoder part filter is , in the decoder part, is extracted from the audio output of the decoder part filter. .

This means that the data associated with the excitation sequence is no longer transmitted and some parts of the decoder You can reduce the bitrate by simply pulling the excitation from the previous output audio. I can do it. Preferably, at least some of the filter parameters are , to reduce the actual error that would be generated by the decoder part rather than the prediction error. The sea urchin is taken out. Note that the synthesis filter may be provided in the encoding section. , and the actual error can be determined by subtracting the locally synthesized audio from the input audio signal. You can take it out.

Preferably, the generation of parameters minimizes the actual error and reduces the total zeros of the filter. This is achieved by iteratively retrieving the stage's delay parameters. I also like it Alternatively, at least some of the other parameters are calculated within said loop. It will be done. Also preferably, all remaining parameters are These are optimized together by calculation. In other embodiments, these The optimization method uses the excitation of multiple parallel feedforward stages of the filter as Can be applied to systems using previous excitation rather than previously decoded speech. can. By adopting the claimed invention, it is possible to significantly improve the signal-to-noise ratio and downsize the device. This reduces the coding delay for a given signal-to-noise ratio. It can be made shorter.

Other forms of the invention are set out in the claims. From now on, this invention will be described by way of example with reference to the accompanying drawings, in which:

Brief description of the drawing FIG. 1 is a schematic diagram of a generalized transmitter with a coder according to the invention; FIG. 1a shows an embodiment of a transmitter with a coder according to the invention, FIG. 2 is a schematic diagram of a generalized receiver with a decoder according to the invention; FIG. 2a shows an embodiment of a receiver with a decoder according to the invention, FIG. 3 is a schematic diagram showing the components of the single excitation synthesis filter of FIG. 1 or 2; Figures 4a and 4b show the parameters of the synthesis filter of the coder according to the invention. We provide six algorithms that address how to use data most effectively. FIG. 5 shows the configuration of a composite excitation synthesis filter of a coder according to another embodiment of the present invention. A schematic diagram showing the elements, Figure 6 shows the parameters of the composite input synthesis filter effectively. Show examples of how to utilize Figures 7a and 7b show waveforms of excitation waves generated in an embodiment of the invention. This is a diagram.

The following description refers to the transmitter of FIG. (]i-0.1, 2...n-1 is shown. Each frame Based on the system, stage 1 of optimal use of filters uses the following ``synthetic analysis'' technology. , pull out the filter parameters. Excitation sequence source 2 generates the maximum The initial excitation sequence is a predetermined impulse response filter with a response of H(z). - B (z) and infinite impulse response (all poles) filter 1/A (z) Accordingly, the two-stage synthesis filter 3 is driven: to=1 This is the size of the folding frame.

A better measure of error should not emphasize the format region in the error spectrum. The error signal is frequency-weighted by a centrifugal function, and the weighting is performed by filter 5. filtered as desired. In an iterative closed loop, this signal is is minimized by calculator 6 to optimize the parameters of synthesis filter 3. . The synthesis filter 3 obtains values for [ak, [b], and dl. these The value of is quantized The signal is quantized by a unit 7 and passed through a coder 8. Coder 8 is a decoding station. code the parameters for transmission to the application. Also, the coder 8 for transmission to a local decoder 10 which is functionally equivalent to the decoder at the station. Related parameters are also coded.

In the local decoder 10, the parameters generate a local output frame. By the same excitation sequence [X,] the parameters were optimized to excitation sequence generator It is driven from the device 2.

This combined output is received and processed as required by the excitation adaptation calculator 12. Ru. This excitation adaptation calculator 12 receives a new value from the output of the decoder synthesis filter 11. An excitation sequence rx-s] is constructed. The new excitation sequence is first Use of the next input audio frame forming part of the coded speech (PDS) signal is passed through the excitation sequence generator 2.

The excitation adaptation calculator 12 and the local decoder 10 are connected to the main source as shown in FIG. 1a. It is not necessary in all embodiments of the present invention. In this simple example, the input sound full voice [y-] - the optimized filter parameters are received by the excitation sequence generator 2. obtained in a closed loop by using the excitation sequence supplied by Ru. If this is done, by using these optimized parameters (rather than the local decoder synthesis filter 11 as described above) It is synthesized by the router 3 and is supplied to the synthesized flare generator 2, which uses it as the next excitation sheet. It is given as kens [X,]. (The first frame of the input audio is encoded If so, it also remembers the required initial excitation sequence. ) In FIG. 2, the decoder unit in the far decoder 20 has a filter parameter. Perform the opposite operation to coder 8 to restore meters [a], [bk], and dl. cormorant. This filter parameter generates a pre-optimized placement at the transmitter. It passes through a synthesis filter 21 in order to The excitation sequence generator 23 generates an excitation sequence. The filter is driven by the sequence [X,]. This excitation sequence [X,] is The materials used in the transmitter is the same as the one used by the transmitter for the first frame of audio. is the initial excitation sequence equal to .

This output is also sent to the excitation sequence generator 23 in the same way as the local decoder 10. The new excitation sequence [X, ]″−1 is fed to the excitation adaptation calculator 22 in order to create .

As shown in FIG. 2a, in a single-person implementation, the excitation adaptation calculator 22 is not required. . In addition, the excitation sequence generator 23 also outputs the last decoded frame as described above. If we perform exactly the same synthesis using , the resulting updated excitation sequence is also the same, and the excitation sequence There is no need to transmit any performance data.

Incorrectly received filter parameters will result in an “incorrect” excitation sequence and this The excitation sequence is a sequence for optimizing the following parameter groups on the transmitter side. Therefore, it is desirable to match the transmission decoders. make this happen To do this, control bits are periodically sent to both decoders and each excitation sequence generator 23, restart the pre-programmed initial excitation sequence. Just give instructions.

Most or all of the functions described above can be achieved with a single digital processor. , does not require extra physical circuitry.

We now describe in detail the filter optimization stage 1 of the transmitter of Figures 1 and 1a. Ru. In many conventional systems using LPC coding techniques, the optimization process is performed by minimizing the value of the pre/l1ll (significant) error. child The significant error is the comparison between the current value of the audio signal and the predicted value based on past sample values. It can be obtained by

However, e is the prediction error of the ith sample.

Although this technique can be used to embody this invention, it is By passing the actual pressurized misuke filter 5 between the synthesized voice outputs of the filter 3, the obtained It is desirable to optimize the filter characteristics for each weight.

The weighted filter 5 is defined similarly to MPLPG or RPE-LPG. Therefore, the explanation of its operation will be omitted.

"Actual error" shall include the actual cognitively weighted difference.

is a nonlinear problem that involves complex calculations, and in the preferred embodiment of this invention, the filter - Characteristics are the parameters for the minimum prediction error energy e”i, which is a direct signal. Partially linearize the problem by first computing (using mean least squares) Selected by Using these values, the remaining values are obtained.

The values of other parameters can be calculated inside or outside the loop as follows: can.

In Figure 3, the synthesis filter 3 of filter optimization stage 1 is connected in series. It consists of two stages. Suna n-di That is, it consists of an infinite impulse response filter with transmission characteristic B(z)z The first filter circuit 31 and an infinite impulse having a transmission characteristic of 1/A(z) The second filter circuit 32 includes a second filter circuit 32 consisting of a frequency response filter. Therefore, the whole The filter impulse response H(Z) of is required in equation A. first The filter circuit 31 is a q1+1 coefficient circuit [b,] (i-0,11,,,q ) and q1+d are constructed with delay circuits, and the length of (d1+n-1) is delay, followed by a q1+1 stage non-recursive filter. There is. In reality, a small number of coefficient circuits (q1-0.1o "2") are used.

The number of delay circuits (and therefore delay lengths) used is limited to a predetermined maximum number N- in practical operation. 1 (this value may be smaller or larger than the frame size n).

Therefore, this first filter circuit 31 receives the excitation signal EX,3 and [X ,] contains the most recent ql+d1+n sample values of the sequence.

The second filter circuit has filter coefficients [ak (k-1, 21,,,p) consists of a recursive filter with an all-pole response defined by

The following description uses vector markings. Vector a indicates the [ak] coefficient group , vector b indicates a group of coefficients [b,].

For each received frame for the input sound, the filter optimization calculator 6 first: Calculate the [a,] coefficient of the average prediction error energy eTe.

In the preferred embodiment (see method 1 of FIGS. 4 and 5), each d1 between 0 and N-1 Compute the a vector by combining it with the b vector in the loop for the value, as described above. against the police to the minimum e The a vector and b vector to be minimized are obtained by the following equations.

However, X is the excitation sample X from -(n+d 1q) to -(dl+1) nx (ql + 1) matrix of k, and Y is yk sun from y to y Pull's (nxp) matrix-p n-2 It's Kusu.

The actual error energy iT″i! is minimized using one of the following three methods: be done.

The actual error energy is determined using y-'9. therefore,) I(Z) is determined within the synthesis analysis loop. In this loop, dl is N Varies within a range of values.

−, while i and 5 are optimized for the minimum predetermined 11 FI error energy 11 1. This will be described in FIG. 4 as Scheme 1.

In a second embodiment, the approach for evaluating E is a variation of Scheme 1 above. It is the shape. In this case, i and 6 are prediction errors; for the value of d that minimizes T is defined by equation (D). Once dl is optimized, vector b is re-evaluated (outside the optimization loop) using equation (F) to minimize the actual error. valued.

Here, m indicates the output of the 1/A(z) filter when its output is zero. Ru. Its memory consists of the most recent sample of the previous synthesized audio frame. be done. Q means a (n x n) spiral matrix, and qk is a synthetic search step. of the 1/A(z) filter using i and dl obtained from the previous analysis by is the second value of the impulse response. This is shown as the second scheme in Figure 4. It will be done.

In a third example, for each given value of dl, using equation (D): a (and b) are calculated to minimize the pre-ΔIIJ error. Stone value is actual is (re)calculated using (F) in a loop to minimize the error of .

Therefore, each d [(0(d,(N-1)) is divided by a composition loop defined for In the analysis, if i uses equation (D), the first ancient T ancient can also be calculated directly from equation (G). Ru.

The d1 value that minimizes the actual error energy, corresponding to a.

Five values are then selected to be optimal. This is shown in Figure 4 as Scheme 3. is shown.

For accuracy reasons, the above schemes 2 and 3 are based on the actual error rather than the prediction error energy i. 5 is chosen to be (re)defined by minimizing the difference energy i.

In another embodiment, i is 6, as shown in Figure 4.5 as Method II. After the assumption that it is equal to zero, outside the loop according to equation (H), independently of 6, , is first calculated to minimize the predicted energy.

E = (YTY)″l YTY’ (H) Next, i is given, and 6 and d1 parameter parameter can be optimized according to scheme 1.2.3 above. In particular, the minimum predicted The value of b for generating energy eTe is given by equation (J).

6 = (xTx) - 1x" (y - yM) (J) to Scheme 1.2.3 The corresponding approaches are shown as Schemes 4 and 5.6, respectively, in Figure S4. , which constitutes the 4.5.6 embodiment of the present invention.

The generated parameters are then quantized by a quantizer 7. This quantizer allocates available bits between parameters.

The excitation for the synthesis filter was explained using a single excitation sequence [X,] However, it is desirable to use several excitation sequences. Therefore, in Figure 1, The excitation sequence source 2 generates a plurality of different sequences [X, 1. [u,].

[V, ] is adapted to supply the synthesis filter 3. Ru.

The previously decoded audio y, input to is obtained by the excitation adaptation calculator 12 from the intermediate output of the controller 3. (or that All such sequences may be obtained as described below. ) Figure 5 Smell Therefore, filter 3 (of course, filter 11.21 in Fig. 1.2.28) is connected in series. It consists of two stages located at The number of first filter elements 51 is j. It consists of filters 51a, 51b, . . . , 51j, each filter having an excitation series. - Kens [X,].

[u　], EV-1], etc., and generates an output. Illustrated As shown in FIG. (2) Z-n-d2B (z), B (z) (3) It has Z-n-d3. And those B (2) The combined output is provided to a second filter element 32.

The second filter element 32 is a repeating filter with a response 1/A(z), as described above. It is a filter.

As shown, each filter 51a, 51b, etc. is a single excitation filter. is used, the actual error energy is minimized between the parameters. Optimized by This optimization is explained in the first to sixth embodiments. By partially linearizing (i.e. , some parameters can be calculated by taking the minimum Al11 energy. ), with reduced computational complexity.

In general, the filter parameters will allow the filter to grow and improve its accuracy. are optimized in order.

This is useful in transmission schemes, if a reduction in the proportion of bits is required. For example, only the earlier coefficients (eg (1), dl) need to be transmitted.

S b As shown in FIG. 6, filter optimization stage 6 in one embodiment used to adopt the calculation based on the first embodiment (method l) for the It will be done. In the first stage of the process, a, b, and dl were made to look like zero, and The actual error error according to scheme 1.2 or 3 in Fig. 4 can be calculated depending on the parameters of It is found by minimizing the energy.

In preferred embodiments, a and b may themselves be redefined in various ways. It can be used for

The first approach is based on Scheme 1 (or Scheme 2). If b is When computed independently of i and b, the minimum predicted energy e”e solution is: It is given by the formula below. .

b (2)-(uTu)-'uT(y-Ya-Xb(1)), excitation sample nx(ql+1) matrix of Luke, which corresponds to the equation F of the first stage. Ru.

−(2) is computed by being alternately combined with i or with i and b, and in this case In this case, a simple matrix representation corresponding to equation (D) in the first stage is used. It will be done.

can be calculated for Rugi. However, Koruru filter is selected be done.

Some of the filter parameters are special is reevaluated (out of the loop) for the values d□−d1′ and d2−d,,’ . In particular, the second analysis by the synthesis process is the filter for the minimum predicted energy. , < When evaluating the parameter, some or all of the filter parameters are For either the minimum predicted error energy or the minimum actual error energy, The re-limited calculation is as below.

Alternately, an approach based on the algorithm of Scheme 3 is adopted. for example, (i) b is a combination performed independently of a and b for N different values of d2. Analysis range by forming loop-T2 is selected for the value of d2-d2' which is the minimum energy ee.

(ii) b is concatenated with the minimum prediction error energy eTe, and 6(2) is the minimum Reevaluated for prediction errors.

(iii) b(2) is calculated by combining i and −(1) for the minimum prediction error. It will be done. The new i vector is preserved, while b and b are from equation (D3) Reevaluated for minimum prediction error.

Given the optimized values d1 and d2, i.e. du' and d2', the filter parameters Some of the meters are calculated using equation (D3) (not analyzed by the synthesis loop) On a single Ste-knob, re-optimization is performed, resulting in b and -(2) relimit.

- (3) and to calculate d3 and specify the previous filter An extension of the calculations described above to a third stage for recalculating the equations from one stage to the next with a total number of terms modified appropriately to include higher terms. It increases sequentially like an optimization process that unfolds over multiple pages.

In other embodiments of the invention, approaches 4.5 or 6 of the scheme of FIG. The search is performed as follows, where the coefficient [ak is the minimum value using the usual LPG solution method. It is first optimized for the prediction error energy iTi.

The method used in calculating the filter coefficients can be used in stage 1 A wide range of algorithms based on one or more of six schemes Selected from rhythm. Considering the second stage equation, the stage 1 equation and form It can be seen that the formulas are similar. The filter optimization calculator 6 includes stage 2 and its For subsequent stages, use the same method as for stage 1, extending as necessary. I can do things. Easily repeat programming techniques and, in turn, simplify systems. This is because it leads to simplification.

Choosing some of the filter parameters affects the minimum prediction error energy e. but other (e.g. inverse filter operations) [yi]) and can be calculated.

Adapting the H(Z) filter described above is only limited to the values of the set of coefficients. However, the structure of the filter is redefined, i.e. is important.

The operation of Excitation Adaptive Calculator] 2 will be described next. From the above application, the synthesis filter 21 is The excitation used at the receiving station to drive the It is clear that it must be possible. The reception filter 21 is When the local decoder filter 11 and optimization filter 3 in the transmitter are the same , the excitation must be the same as that for the transmitted parameters to be optimized. No.

transmitter 2 and receiver 23 with a pre-programmed initial sequence. This can be configured as described above by providing an excitation sequence generator that It will be done. The exact properties the sequence has are not critical, e.g. It is like an average Gaussian random sequence.

This single sequence is applied to a composite input synthesis filter 3 configured as shown in FIG. It can be used to drive each input. Three responses as mentioned above, Considering the first three filters 51as, 51bs, and 51c that waited for Here, n is the frame size of the analysis frame, and the values of parameters d1, d2, and d3 are , denote certain segments in the sequence [X,], which are H(Z) synthetic fi Provided as an excitation signal for each of the FIR signals 51a, 51b, and 51c of the router 3. be provided.

In the embodiment of the PDS of this invention, the additional signal calculation circuit]2 is simply the excitation fJ number. This is a link between the storage section and the audio output section. Use PDS signal as excitation sequence This increases the signal-to-noise ratio (SNR) at a given transmission rate, and/or reduce coding delays. It is extremely effective.

As mentioned above, it is of course possible to use other inputs instead of X (Z). It is. In particular, X (Z) is derived from the output of the first filter element 31 or 51 (immediately In other words, it can be taken out (as an intermediate output of the entire filter H(Z)). this place In this case, this filter element has a feedback circuit between its output and input. It can be thought of as an infinite response filter. Here, the excitation signal is Extracted only from output. Also, the input to the A(Z) filter is, for example, °Self-excitation by Rose and Barnwell I ``Vocoder: Another approach to long distance calls using 4800 baud'', ( 453-45 of ICASSP's IEEE PROC, April 1986 issue. This is equivalent to the data generated in the “self-exciting vocoder” shown in page 6). be.

The parameter optimization method described above can be applied similarly to cases with parallel filter structures. Applicable.

The graph shown in FIG. A typical signal waveform is shown. This signal is a PDS signal and is passed through the all-zero filter 31. It is filtered and has a very similar waveform to the voice waveform. Second PDS The audio approximation of the excitation is completely different from the PES excitation waveform shown in Figure 7b. It is. These PES excitation waveforms are compatible with conventional systems such as MPL This is very similar to the excitation waveform used in the system.

These excitations are unique to the glottal pulse train formed by the human vocal chord. This is a similar pulse train.

In this way, PDS excitation according to the present invention is more effective than when using conventional PES excitation. gives much better results. Therefore, the PDS signal undergoes less filtering. However, it is an “audio approximation” signal and cannot be used with conventional audio sources/filters. Those who have to resynthesize each audio frame from random waveforms in the model For the method, in order to see the changes between audio frames, we simply use the preceding sound in the filter. All you have to do is transform the voice frame.

In each example, it was found that the SNR decreased as the frame size increased. Ta. In particular, in embodiments using PDS with a single excitation sequence, the frequency High 5NR4' in a small room size! This is transmitted This means improving the clarity of speech. Furthermore, using single excitation type PDS implementations can be configured with smaller frame sizes, resulting in code This has the effect of reducing delays caused by conversion.

Instead of evaluating filter parameters individually, combine several parameters It may be advantageous to evaluate the

It is possible to increase the number of coefficients of the filter or increase the maximum value of d. It was found that the SNR during encoding and decoding can be increased by It was. In order to improve SNR, it is necessary to increase the transmitted bits per frame. Even if necessary, it is better to allocate the number of bits to the coefficients of the new filter. that it is advantageous to allocate a special bit for the delay parameter is clear.

international search report

Claims (25)

[Claims] (1) The filter parameters are periodically derived from the input audio signal in the coder. sent to the decoder to update the decoder filter response. and the excitation input to the decoder filter is the audio of the decoder filter. A method of voice communication, characterized in that the output is derived in the decoder. . (2) Deriving the filter parameters includes combining the input audio signal and the data to be performed, so as to reduce the actual error between the coder's output and the coder's output. The method of claim 1, characterized in that: (3) The decoder includes a composite filter comprising a first and a second filter in series. the first filter has a variable delay parameter and the second filter has a variable delay parameter; router is an infinite impulse response filter, and the filter parameters the input audio signal and the output of the decoder by changing the delay parameter; Derive the corresponding evaluation of said actual error between the forces and then this estimated and selecting said delay parameter values that reduce the actual error. The coder is periodically derived by a process. The method described in claim 2. (4) said coder includes such a synthesis filter, and said coder includes an evaluation of said actual error; generating a synthesized speech output for each value of said delay parameter; in the coder by comparing the output with the input audio signal. 4. A method according to claim 3, characterized in that: (5) For each of the delay parameter values, other parameters of the first filter are determined. The values for the meter and for the parameters of the second filter are together calculated and used to derive said estimated actual error. 5. The method according to claim 3 or 4. (6) The values for the parameters of the second filter are determined outside the iterative process. The values for the other parameters of the first filter are initially evaluated and the values for the other parameters of the first filter are Deriving the estimated actual error calculated for each of the extended parameter values 5. The method according to claim 3, wherein the method is used for: (7) The filter parameters are periodically derived from the input audio signal in the coder. sent to the decoder to update the decoder filter response. The decoder includes first and second filters in series, and the first and second filters are arranged in series. The filter has a variable delay parameter, and the second filter has an infinite impulse response. the excitation input to the decoder filter is the excitation input to the decoder filter. derived in the decoder from one or more intermediate outputs of the filter; The filter parameters vary the delay parameters and the filter parameters vary between the first and second filters. calculate other parameter values of the input audio signal and the output of the decoder; and then use this value to derive an estimate of the actual error between selecting the delay parameter value to reduce the estimated actual error of is derived in said coder by an iterative process including A method of voice communication. (8) For each of the delay parameter values, other parameters of the first filter are determined. 5. The meter value is calculated such that the prediction error is low. The method described in any one of (7) to (7) above. (9) For each of the delay parameter values, other parameters of the first filter are determined. 5. The meter value is calculated in such a way that the actual error is low. 7. The method according to any one of 7. (10) After selecting the delay parameter, other parameters of the first filter The value of is recalculated so that the actual error is lowered. 7. The method described in any one of 7. (11) The actual error depends on the spectral region that is perceived to be less important. Claims 1 to 10 are weighted to reduce the The method described in Shifting. (12) When excited by a signal derived from a previous synthesized speech output, they The difference between the input audio signal and the synthesized audio output generated by the filter with the characteristics of filter characteristics from said input audio signal so as to reduce the estimation of the actual error between A voice code, characterized in that it comprises means configured to periodically derive the gender. -der. (13) comprising a synthesis filter adapted to generate the synthesized speech output; The actual error evaluation is based on the input audio signal and the synthesized audio output of the synthesis filter. 13. The coder according to claim 12, wherein the error is an error formed between the force and the force. (14) configured to periodically derive filter characteristics from the input audio signal means, the filter characteristic comprising first and second filters in series. an output produced by a synthesis filter comprising a synthesis filter and said input the first filter is derived to reduce the actual error between the audio signal and the audio signal; comprises a plurality of parallel feedforward filters, and the plurality of parallel feedforward filters At least one of the feedforward filters includes the parallel feedforward filter. Receive one of several different excitation sequences derived from the combination of router outputs and the second filter is an infinite impulse response filter. , said infinite impulse response filters change their characteristics when excited A voice coder comprising: (15) Changing the delay parameter of a certain first parallel feedforward filter , derive such an estimate of said actual error for such a value, and calculate said actual error Then select that delay parameter value to decrease the parallel feedforward repeating the above steps for the delay parameters of each of the code filters; configured to derive the filter characteristics by a method comprising: and 15. A coder according to claim 14, characterized in that it comprises means. (16) further comprising a local decoder adapted to generate synthesized speech output; The coder according to any one of claims 12 to 14. (17) comprising a filter, and upon receiving the coded audio signal, the filter The filter is configured to update the characteristics of the filter, and the filter is configured to update the characteristics of the filter. an audio output connected to a A decoder for coded speech, characterized in that it is excited when used. (18) comprising a filter, and upon receiving the coded audio signal, the filter The filter is configured to update the characteristics of the filter, and the filter includes a first and a second , the first filter includes a plurality of parallel feed forward filters. filter, at least one of the plurality of parallel feedforward filters One is derived from the combination of the outputs of the parallel feedforward filters. said second fibre, connected to receive one of a plurality of different excitation sequences; A router is an infinite impulse response filter for coded speech. decoder for. (19) At least one of the parallel feedforward filters includes the second connected to receive the excitation sequence derived from the output of the filter of 19. The decoder according to claim 18, wherein: (20) A receiver comprising the decoder according to any one of claims 17 to 19. (21) A receiver substantially equivalent to that described with reference to Figures 2 and 2a. . (22) A transmitter comprising the coder according to any one of claims 12 to 16. (23) A transmitter substantially equivalent to that described with reference to FIG. 1 and FIG. 1a. . (24) An audio communication device comprising the transmitter and receiver according to any one of claims 1 to 23. trust system. (25) A voice communication method almost equivalent to that described with reference to the drawings.