JPH05232995A - Method and device for encoding analyzed speech through generalized synthesis - Google Patents

Abstract

PURPOSE: To reduce the number of band requests from an analyzed speech encoding system by synthesis by selecting the parameter of a trial original signal having an error satisfying an error evaluating process as the encoded expression of the original signal. CONSTITUTION: A trial original signal generator 10 generates a trial original signal s(i) which can be heard like as an original signal s(i) to be encoded. A speech encoder/synthesizer 15 decides the encoded expression of the signal s(i) and generates a reconstituted speech signal s(i) based on the encoded expression. A subtracting circuit 17 forms the error signal E(i) between the signals s(i) and s(i). The signal E(i) is fed back to the generator 10 so as to reduce the signal F(i) by selecting another trial original signal. Thus the trial original signal smin (i) which generates the minimum error Emin within a certain limit is discriminated. The parameter used by the encoder/synthesizer 15 is used as the encoded expression of the trial original signal s(i).

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD This invention relates generally to speech coding systems, and more particularly to reducing the bandwidth requirements of analytic speech coding systems by synthesis.

[0002]

BACKGROUND OF THE INVENTION Speech coding systems provide a codeword representation of a speech signal for communication over a channel or network to one or more system receivers. Each system receiver reconstructs a speech signal from the received codeword. The amount of codeword information communicated by the system in a given time defines the bandwidth of the system and affects the quality of speech reproduced by the system receiver.

Voice coding system designers seek high quality voice reproduction capabilities by utilizing as little bandwidth as possible. However, the requirement for high quality voice and the requirement for low bandwidth conflict with each other and thus require trade-offs in the design process. However, the present speech coding scheme was developed to provide acceptable speech quality with reduced channel bandwidth. The analysis voice coding method by synthesis is included therein.

In the analytic speech coding technique by synthesis, the speech signal is coded by a waveform matching procedure. The candidate speech signal is compared to the original speech signal to be combined and encoded from one or more parameters. Different synthesized candidate speech signals are determined by changing the parameters. The parameters of the best matching candidate speech signal are used to represent the original speech signal.

Many synthetic analysis encoders, such as Code Excited Linear Prediction (CELP) encoders, use a long-term predictor (LTP) to model the long-term correlation of a speech signal (herein referred to as the "speech signal"). The term refers to the excitation signal of either the actual speech or synthetic coder of analysis. In general, correlation allows the past speech signal to be used as an approximation of the current speech signal. The LPT allows some past speech signals (which have already been encoded) to be compared with the current (original) speech signal. By such comparison, the LTP determines which of the past signals best matches the original signal. The voice signal in the past can be identified by the delay, which indicates how long in the past the signal was. LTP
The coder that uses ∑ produces a scaled version of the best matching past speech signal (ie, the best approximation) with a long-term correlation with the subtraction signal (which is called the residue or excitation) subtracted from the current speech signal. This signal is then typically encoded with a fixed statistical codebook (FSCB). The FSCB index and LTP delay are sent along with other information to the CELP decoder, which recovers the original audio signal's specified values from these parameters.

By modeling the long-term correlation of speech, the quality of the reproduced speech at the decoder can be improved. However, this improvement cannot be achieved without a significant increase in bandwidth. For example, in order to model the long-term correlation of speech, a conventional CELP coder transmits 8 bits of delay information every 5 ms or 7.5 ms (this is called a subframe). Delay parameters that change over time are, for example, 1 kb / s to 2 kb / s in the band
Will be increased. Changes in LTP delay may be unpredictable in time (ie, the sequence of LTP delay values is statistical in nature), so it is difficult to reduce the demand for additional bandwidth by encoding the delay parameters. It may be.

One way to reduce the need for extra bandwidth by LTP-preserving analytic encoders is as follows:
The LTP delay value is transmitted less frequently and the intermediate LTP delay is determined by interpolation. However, if the interpolation is performed, the sub-optimal delay value is used by LTP in each subframe of the audio signal. For example, if the delay is suboptimal, the LTP will suboptimally map the past audio signal to the current audio signal. As a result, the residual excitation signal becomes larger than in other cases. In this case, the FSCB must operate to correct the effect of its suboptimal time shift, rather than performing its normal function of trimming the waveform. Without such correction, there will be significant audible distortion.

[0008]

SUMMARY OF THE INVENTION The present invention provides a method and apparatus for reducing bandwidth requirements in an analytical speech coding system with synthesis. The present invention is the actual original signal to be encoded.
signal) based on the number of trial origin signals (trial origin).
al signal). These trial original signals are constrained to have listening characteristics similar to the actual original signal, and are used instead of or as an adjunct to the actual original signal in encoding. The original signal, and thus the trial original signal, may take the form of the actual speech signal, or it may take the form of the excitation signal present in the synthetic analysis encoder. The present invention allows for analytic coding by generalized synthesis by allowing changes in the original speech signal to reduce coding errors and bit frequencies. The invention is applicable, among other applications, to networks for voice communication, such as cellular or regular telephone networks.

In one embodiment of the present invention, the trial original signal is used in the encoding and combining processes to produce a reconstructed original signal. An error signal is formed between the trial original signal and the reconstructed signal. The trial original signal determined to produce the smallest error is used as the basis of the coding and coding for communication to the receiver. By reducing the error in this way, the coding process is modified to reduce the desired system bandwidth.

In another illustrated embodiment for the CELP encoder, one or more trial original signals are provided by applying a time warp codebook to the actual original signal. In the LTP procedure of the CELP encoder, the trial original signal is compared with past speech signal candidates provided by the adaptive codebook. The trial original signal closest to the candidate is identified. As part of the LTP process, the candidates are subtracted from the identified original trial signals to form a residue. This residue is then encoded by applying a fixed statistical codebook. As a result of using a large number of trial original signals in the LTP procedure, one embodiment of the present invention improves the mapping of past signals to current signals, resulting in lower residual error. By thus reducing the residual error, the transmission frequency of the LTP delay information can be reduced, and delay interpolation can be performed without deterioration of the reconstructed voice or with a slight deterioration.

In another embodiment of the present invention, multiple trial original signals are provided by a time shift technique.

[0012]

DETAILED DESCRIPTION INTRODUCTION FIG. 1 illustrates one embodiment of the present invention. The original signal s (i) to be encoded is provided to the trial original signal generator 10.

[0013]

[Outer 1]

By allowing the original signal to change so as to reduce the error in the encoding process, the present invention generalizes the conventional synthetic analysis encoder.
The encoder / combiner 15 may thus be any conventional combining analysis encoder such as conventional CELP.

Conventional CELP FIG. 2 illustrates a conventional synthetic analysis CELP encoder. The sampled speech signal s (i) (where i is the sample index) is provided to an optimized Nth order short-term linear prediction filter (STP) 20 for the current speech segment.
The signal x (i) is the excitation after filtering by STP.

[0016]

[Equation 1] Here, the parameter a _n is given by the linear prediction analyzer 10. N is usually about 10 samples (8kH
(at the sampling frequency of z), the excitation signal x (i) preserves the long-term periodicity of the original signal s (i). LTP 30 is provided to remove this redundancy.

The value of x (i) is determined for each block.
Each block is called a subframe. Linear prediction coefficient a
_n is determined by the analyzer 10 on a frame-by-frame basis. The frame has a fixed length, which is an integral multiple of the subframe length, and usually has a length of 20-30 milliseconds. The value of the sub-frame of a _n are usually determined by interpolation.

The LTP determines the gain λ (i) and delay d (i) to be used as follows.

[0019]

[Equation 2]

[0020]

[Outside 2]

The data representation of each subframe of speech, namely the LTP parameters λ (i) and d (i) and the FSCB index, are collected for the number of subframes equal to the frame (typically 2, 4 or 6). With coefficients a _n, this frame of data is notified to the CELP decoder, wherein utilizing this reconstruction of the speech is performed.

[0022]

[Outside 3]

[0023]

[Outside 4] Perceptually related error conditions may be used to select vectors from this codebook. This can be done by utilizing the spectral masks present in human hearing. Thus, instead of using the difference between the original speech signal and the reconstructed speech signal, this error condition will use a perceptually weighted difference.

Perceptual weighting of the signal will give de-emphasis to the formats present in the speech. In this example, the format can be described as an all-pole filter that achieves spectral de-emphasis by moving the poles inward. This corresponds to replacing the filter having the prediction coefficients a ₁ , a ₂ , ..., A _N with the filter having the coefficients γ a ₁ , γ ² a ₂ , ..., γ ^N a _N. γ is a perceptual weighting coefficient. (Usually a value of about 0.8.)

The sampled error signal g (i) in the perceptually weighted region is

[0026]

[Equation 3] Is. The error condition of the analytic encoder by combining is formed for each subframe. For L sample length subframes, the commonly used conditions are

[0027]

[Equation 4]

[0028]

[Outside 5]

[0029]

[Outside 6]

In the time domain, the de-emphasis of the spectrum by the coefficient γ causes the impulse response of the all-pole filter to decay quickly. Actually, 8kH
When the sampling frequency of z is γ = 0.8, the impulse response has no significant energy 20 samples ahead.

By thus rapidly attenuating, the impulse response of the all-pole filter 1 / (1-γa ₁ z ^-1
... γ ^N a _N z ^-N ) can be approximated by a finite impulse response filter. Let the impulse response of this filter be represented by h ₀ , h ₁ , ..., H _R-1 . As a result, the operation of the error condition for the perceptually weighted voice can be displayed as a vector. Since the encoder operates on a subframe-by-subframe basis, it is convenient to define the vector at its sample L to match the length of the subframe. For example, for the excitation signal

[0032]

[Equation 5] Becomes Further, the spectrum weighting matrix H is defined as follows.

[0033]

[Outside 7] If the matrix H is partitioned into L × L rectangles, then equation (8) approximates equation (4), which approaches the general covariance condition used in the original CELP. .

Illustrative Embodiment of CELP Encoding FIG. 3 illustrates the application of one embodiment of the present invention to CELP encoding. A sampled speech signal s (i) is provided for coding. Signal s (i) is given to the linear prediction analyzer 100 which produces linear predictive coefficients a _n. Signal s (i)
Is also provided to STP 120, which operates according to the process shown in equation (1). The signal s (i) is the delay estimator 1
Also given to 40.

The delay estimator 140 searches the recent history of s (i) (eg, between the past 20 and 160 samples) and the subframe s (i) of the current speech to be encoded.
Determine the set of consecutive past samples (of length equal to the subframe length) that best match The delay estimator 140 uses the current subframe and i-160 < i < i-2.
The determination may be made through a correlation procedure with a continuous set of values of zero past samples s (i). As an example of the correlation technique, the technique used by the conventional open loop LTP used in CELP encoders may be used (the term open loop here refers to the original signal rather than the reconstructed past speech signal). The LTP delay estimation process is described, and the delay estimation process using the reproduced audio signal is referred to as closed loop.). The delay estimator 140 determines the delay estimation value once per frame according to the procedure described above. Delay estimator 140
Calculates the delay value M of each sample frame by interpolating the delay value determined at the frame boundary.

[0036]

[Outside 8]

[0037]

[Outside 9]

As mentioned above, the LTP process operates to identify the past speech signal that best matches the current speech signal in order to reduce the long-term correlation of the coded speech. In the embodiment of FIG. 3, multiple trial original signals are provided to the LTP process. A number of such trial original signals are provided by the time warp function 130.

The time warp function 130 shown in FIG. 4 provides a codebook 133 for time warp (TWCB) to apply to the original signal to produce a number of trial original signals. In principle, the codebook 133 of the time warp function 130 is an arbitrary time warp.

[0040]

[Equation 6] , Which does not change the perceptual quality of the original signal.

[0041]

[Equation 7] Where t _j and τ _j include the start of subframe _j in the original and warp regions.

In order to increase the stability of the warp process, it is desirable that the main pitch pulse falls near the right border of the subframe. This is accomplished by defining the boundaries of such subframes to fall just to the right of such pulses using well known techniques. Assuming that the pitch pulse of the speech signal to be coded is at the boundary point, it is desirable that the warping function satisfies

[0043]

[Equation 8] If the pitch pulse is somewhere before the subframe boundary, ζ (t) should keep its ending value near the subframe boundary. If the expression (10) is not satisfied, the warp will be oscillating. The following family of time warp functions may be used to implement the time warp codebook.

[0044]

[Equation 9] Here, A, B, C, σ _B and σ _C are constants. The time warp function converges to A as t increases. t
_{In j} , the value of the warping function is A + B. The value of C is used to exactly satisfy equation (10). A continuous time warp codebook selects 1) the value of A (typically between 0.95 and 1.05), 2) the values of σ _B and σ _C (typically 2. 5 ms), 3) Use B to satisfy the boundary condition of t _j (where ζ (t _j ) = A
+ B), 4) by selecting C to satisfy the boundary condition of equation (10). Please note that no information about the warping codebook is transmitted. Its size is determined only by computational requirements.

[0045]

[Outside 10]

[0046]

[Equation 10] Equation (12) is similar to equation (8), but unlike equation (8), equation (12) is normalized so that the least squares error process is only sensitive to shape differences. It has become.

[0047]

[Outside 11]

[0048]

[Outside 12]

[0049]

[Outside 13]

[0050]

[Outside 14]

Like the conventional speech encoder, the LTP delay,
The scale factor values λ and M, the FSCB index and the linear prediction coefficient a _n are provided to the decoder over the line and reconstructed by a conventional CELP receiver. However, because of the error reduction achieved in the illustrated embodiment of the present invention (in the encoding process), the LTP delay information need only be sent once per frame, rather than once per subframe. The M subframe values are provided at the receiver by interpolating the delay values in a manner similar to that performed by the delay estimator 140 at the transmitter.

By transmitting the LTP delay information M on a frame-by-frame basis rather than on a subframe-by-frame basis, the delay-related bandwidth requirements are significantly reduced.

LTP with continuous delay contour In the conventional LTP, the delay is constant in each subframe and changes discontinuously at the subframe boundary. This discontinuous behavior is called a step delay contour. With step delay contours, the discontinuous change in delay between subframes corresponds to the discontinuity of the LTP mapping of the past excitation to the present. Such discontinuities can be modified by interpolation so that they do not interfere with the reproduction of signals with smoothly varying pitch cycle waveforms. It can be said that it is advantageous to provide LTPs with continuous delay contours for convenience of interpolation, since in the embodiments described above, interpolation of delay values is required. This reconstructed LTP gives a delay contour with no discontinuities, so this is the LTP of the continuous delay contour.
Called.

The process of providing delay values for continuous delay contours that provides an adaptive codebook replaces the delay estimator described above. The best set of possible contours of the current subframe is selected to give the LTP continuous delay contour.
Each contour starts at the end value of the delay contour of the previous subframe d (t _j ). In the current embodiment, the delay contour of each of the sets is chosen to be linear within the subframe. Therefore, in the current N-sample subframe j (with a sampling interval T), t _j <t
Although within < t _{j + 1} , the instantaneous delay d (t) has the form

[0055]

[Equation 11] Here, α is constant. Given d (t), L of the past speech signal (not scaled by LTP gain)
The mapping to the present by TP is

[0056]

[Equation 12]

[0057]

[Outside 15]

When using LTP with a continuous delay contour to obtain a time scaled version of the past signal,
It is desirable to set the gradient of the delay contour to be smaller than 1 and d (t) <1. If this condition is violated, time reversal of the mapped waveform may occur. Further, the pitch doubling cannot be accurately described by the continuous delay contour. To model pitch doubling, the delay contour must be discontinuous. Consider again the delay contour of equation (14). Since each pitch period is normally dominated by one large center of energy (pitch pulse), it is desirable for the delay contour to have one degree of freedom per pitch cycle. Therefore, the illustrated continuous delay contour LTP provides a subframe with an adaptive length of approximately one pitch cycle. This adaptive length is used to provide a subframe boundary placed immediately after the pitch pulse. By doing so, an oscillating delay profile is prevented from occurring. Since the LTP parameters are transmitted at fixed time intervals, the subframe size does not affect the bit frequency. In the illustrated embodiment, well known techniques for locating pitch pulses or delayed frame boundaries are suitable. These techniques are applied as part of the adaptive codebook process 150.

Implementation of CELP encoding with time shift
Examples In addition to the time warping embodiments described above, the time shift embodiments of the present invention may also be used. As an illustrative example, a time shift embodiment is shown in FIG. This is similar to FIG. 3, but the time warp function 130 has been replaced by a time shift function 200.

Similar to the time warp function 130, the time shift function 200 provides a number of trial original signals that are aurally similar to the original signal to be encoded. Similar to the time warp function 130, the time shift function 200 determines which trial original signal has the closest shape to the identified past speech signal. But unlike the time warp function 130,
The time-shift function is an original audio signal, preferably an original audio signal that causes a minimum error when compared with a past audio signal by temporally shifting the excitation signal by a time θ in the range of θ _min < θ < θ _max. Operates to determine the position of the signal (typically | θ _min | = | θ _max | = 2.5 samples, achieved when upsampling is performed). Moving the shift of the original audio signal to the right by θ (ie, delaying it in time) repeats the last interval of the previous subframe length θ, thereby padding to the left edge of the original audio subframe. Executed by The operation of shifting the original audio signal to the left by θ is simply performed here by removing (ie omitting) the signal of length equal to θ from the left edge of the subframe.

Note that the subframe size does not have to be a function of the pitch period. However, it is desirable that the size of the subframe is always smaller than the pitch period. By doing so, the position of each pitch pulse can be independently determined. Subframes as large as 2.5 ms can also be used. Since the LTP parameter is transmitted at fixed time intervals, the subframe size does not affect the bit frequency. To prevent the sub-frame from entering between pitch pulses, the shift change should be appropriately constrained (on the order of 0.25 ms for a 2.5 ms sub-frame). Alternatively, the delay can be kept constant for subframes whose energy is significantly smaller than the surrounding subframes.

An example of the time shift function 200 is shown in FIG. The function 200 is similar to the time warp function 130 described above, or a pad / omit process 232 is provided in place of the codebook 133 associated with the warping process 132. The shift procedure performed by function 200 is

[0063]

[Equation 13] Is. Where t _j is the start of the current frame j of the original signal. The closed loop fitting procedure searches for a value of θ _min < θ < θ _max that minimizes the error condition similar to equation (12).

[0064]

[Equation 14] This procedure consists of process 234 (which determines ε'according to equation (17)) and error estimator 135 (which is ε ').
Determine _min ).

The optimum value θ of subframe j is that θ corresponding to ε ′ _min , and is denoted as θ _j . For subframe length L _subframe , _subframe j + 1 of the original voice
The start of is determined by the following equation.

[0066]

[Equation 15] For the reconstructed signal, the time τ _{j + 1} is simply

[0067]

[Equation 16] Like the illustrated embodiment described above, this embodiment of the present invention uses scaling and delay information, linear prediction coefficients and regular C
The fixed statistical codebook index of the ELP receiver is given. Also in this case, the delay information is transmitted not in each subframe but in each frame due to the reduction of the coding error according to the present invention. The receiver interpolates the delay information and determines a delay value for each subframe as it did in the delay estimator 140 of the receiver.

The interpolation for the staircase delay contour is performed as follows. At t _A and t _B represent the beginning and end of the current interpolation period for the original signal. Further, intex j _A represents the first LTP subframe of the current interpolation period, and j _B represents the first LTP subframe of the next interpolation period. First, at the end of the current interpolation interval d _B , an open-loop estimate of the delay is obtained, for example, by the cross-correlation process of the past and the current speech signal (in fact used for t _B for this purpose The value to do is an estimate, because its final value is obtained at the end of the interpolation.) The delay at the end of the previous interpolation period is represented by d _A. At this time, the delay of subframe j is simply

[0069]

[Equation 17] Given in. The unscaled contribution of LTP to the excitation is given by:

[0070]

[Equation 18] Where τ _j is the beginning of subframe j for the reconstructed signal.

In the analytic encoder based on the doubling of the delay pitch and the halving synthesis, if the consecutive pitch cycles are similar, the phenomenon of doubling the delay and halving is likely to occur. However, with respect to the present invention, delay doubling and halving are dealt with as follows. As a first step, the open loop delay estimate at the end of the current interpolation period is compared to the final delay of the previous interpolation interval. When this is close to a multiple or divisor of the value at the end of the previous interpolation period, it is considered that delay multiple or divisor has occurred. Next, we will explain the doubling and halving of the delay.
Other multiplications can be handled in the same way.

To explain the doubling of the delay, the open-loop estimated value of the delay of the termination value is d ₂ (τ _B ). The subscript 2 here represents the delay corresponding to two pitch cycles.
d ₁ (τ _A ) represents the delay corresponding to one pitch cycle. In general, there is the following relationship between the doubled delay and the standard delay.

[0073]

[Formula 19] Expression (22) shows two sequential mappings by LTP. If the pitch period is not constant, simply doubling the delay will not give the correct mapping. Next, consider the case where d ₁ (τ) is linear in the current interpolation period.

[0074]

[Equation 20] Combining equations (22) and (23)

[0075]

[Equation 21] Equation (24) shows that d ₂ (τ) is linear within a limited range. However, in general, τ _A <τ <τ _A
In the range of + d ₁ (τ), d ₂ (τ) is not linear. The following procedure can be used to double the delay. _First , d ₁ (τ _A ) and d ₂ (τ _B ) are known. Β is obtained by using τ = τ _B in the equation (24).

[0076]

[Equation 22] Next, d ₁ (τ) and d ₂ (τ) within the interpolation period are known. The standard delay d ₁ (τ) satisfies equation (23) during the entire interpolation period. Note that for d ₂ (τ), equation (22) is valid within the entire interpolation period, while equation (24) is valid only in the restricted portion.

The actual LTP excitation contribution to the interpolation period is now obtained by the smoothed change from the standard delay to the doubling delay.

[0078]

[Equation 23] Here Ψ (τ) has a flat function that is increased from 0 to 1 over the indicated interpolation period, which linearizes the current interpolation period. This procedure assumes that the interpolation period is essentially larger than the doubled delay.

For delay halving, the same procedure is used in the opposite direction. Boundary conditions d ₂ (τ _A ) and d ₁ (τ _A ) are assumed. In order to be able to use the equation (22) for τ _A <τ < τ _B , d ₁ (τ _A ) is τ _A −d ₁ (τ _A ) <τ < τ
Must be defined in the _A range. Voice quality is maintained by proper definition. Since the doubling delay is linear in the previous interpolation period, equation (24) can be used to obtain a suitable definition of d ₁ (τ) in this range. In the case of a linear delay contour, d ₂ (τ) satisfies

[0080]

[Equation 24] Where ′ refers to the value of the previous interpolation period (τ _B ′ =
Note that τ _A ), and η ′ is a definition. Comparing this with equation (24), d ₁ (τ) in the last part of the previous interpolation period is

[0081]

[Equation 25] Is. Equation (28) is also the boundary value d of the current interpolation period.
Give ₁ (τ _A ). From this value and d ₁ (τ _B ), equation (2
The value of β in 3) can be obtained. Again, equation (22) can be used to calculate d ₂ (τ) for the current interpolation period.
The change from d ₂ (τ) to d ₁ (τ) is again performed by the equation (22). However, in this case Ψ (τ) decreases from 1 to 0 within the interpolation period.

[Brief description of drawings]

FIG. 1 is a diagram of an embodiment of the present invention.

FIG. 2 is a diagram of a conventional CELP encoder.

FIG. 3 is a diagram of an embodiment of the present invention.

4 is a diagram of a time warp function of the embodiment shown in FIG.

FIG. 5 is a diagram of an embodiment of the present invention related to time shifting.

FIG. 6 is a diagram of a time shift function of the embodiment shown in FIG.

[Explanation of symbols]

 10 means for generating a plurality of trial signals 15 means for generating parameters 17 means for determining an error

Claims (23)

[Claims] 1. A method of encoding an original signal, the method generating a plurality of trial original signals based on the original signals and encoding the trial original signals to produce one or more parameters representative thereof. Then, an estimate of the original trial signal is generated from one or more parameters, the error between the original trial signal and the combined estimate of the original trial signal is determined, and the trial with the error satisfying the error evaluation process is determined. A method of encoding an original signal, comprising the step of selecting one or more parameters of the original signal as an encoded representation of the original signal. 2. The method of claim 1, wherein generating a plurality of trial original signals comprises applying one or more time warps to the original signals. How to make. 3. The method of claim 1, wherein generating a plurality of trial original signals comprises performing one or more time shifts of the original signals. How to make. 4. The method of claim 1, wherein the step of encoding the trial original signal comprises the step of performing an analytical encoding by combining. 5. The method of claim 4, wherein performing analysis-by-synthesis coding comprises performing code-excited linear predictive coding. 6. The method of claim 1, wherein the step of determining an error comprises the step of determining the sum of squared samples of the difference between the filtered trial original signal and its filtered composite signal. A method for encoding an original signal characterized by. 7. The method of claim 6, wherein the error estimation process comprises the step of determining a minimum sum of squares of the samples from a plurality of sums of squares of the samples. . 8. The method of claim 1, wherein the step of determining the error is the square of the difference sample between the perceptually weighted trial original signal and its perceptually weighted synthesized estimate. A method for encoding an original signal, comprising the step of determining a sum. 9. The method of claim 8, wherein the error estimation process comprises the step of determining a minimum sum of squares of the samples from a plurality of sums of squares of the samples. how to. 10. The method of claim 1, wherein the step of selecting an encoded representation of the original signal comprises the step of determining a trial original signal with a minimum associated error. A method of encoding a signal. 11. An apparatus for encoding an original signal, said apparatus comprising: means for generating a plurality of trial original signals based on the original signals; and means for coupling the trial original signals to encode it Means for generating one or more parameters to be expressed, means for synthesizing an estimated value of the trial original signal from the one or more parameters, coupled to the encoding means; the encoding means and the generating means And a means for determining the error between the trial original signal and the combined estimate of the trial original signal, and as a coded representation of the original signal, the trial original signal whose error satisfies the error estimation process. An apparatus for encoding an original signal, characterized by comprising a means for selecting one or more parameters of 12. The apparatus of claim 11, wherein the means for generating a plurality of trial original signals includes means for applying one or more time warps to the original signals. The encoding device. 13. The apparatus of claim 11, wherein the means for generating a plurality of trial original signals comprises a time warp codebook. 14. The apparatus of claim 11 wherein the apparatus for generating a plurality of trial original signals comprises means for performing one or more time shifts of the original signals. apparatus. 15. The apparatus for encoding an original signal according to claim 11, wherein the means for encoding the trial original signal comprises means for performing analysis encoding by combining. 16. The apparatus for encoding an original signal according to claim 15, wherein the means for performing analysis encoding by combining comprises a code-excited linear predictive encoder. 17. The apparatus for encoding an original signal according to claim 11, wherein the means for combining the estimated values of the trial original signal comprises a fixed statistical codebook. 18. The apparatus of claim 17, wherein the means for synthesizing an estimate of the trial original signal further comprises an adaptive codebook. 19. The apparatus according to claim 11, wherein the means for determining an error comprises means for determining a sum of squares of a sample of a difference between the trial original signal and its combined estimated value. A device for encoding an original signal to be reproduced. 20. The apparatus according to claim 19, wherein the error evaluation process determines which of the plurality of sums of squares of the sample has the smallest sum of squares of the sample. A device to be converted. 21. The apparatus according to claim 19, wherein the difference between the original signal and its combined estimate is perceptually weighted. 22. The apparatus of claim 11, wherein the means for selecting an encoded representation of the original signal comprises means for determining a trial original signal having a minimum associated error. A device that encodes a signal. 23. In a network for communicating an original signal, the network is a communication line, a transmitter connected to the communication line for transmitting an encoded representation of the original signal, the plurality of which are based on the original signal. Means for generating a trial original signal of, and an encoding means coupled to the generating means for producing one or more parameters for encoding the trial original signal to express it, and coupled to the encoding means, Means for synthesizing an estimate of the trial original signal from one or more parameters; and a means connected between the encoding means and the generating means, between the trial original signal and the synthesized estimate of the trial original signal. Means for determining the error, and means for selecting as the encoded representation of the original signal one or more parameters of the trial original signal with associated errors that satisfy the error evaluation process. And a receiver connected to the communication line for decoding the encoded representation of the original signal received from the transmitter.