Adaptive differential pulse code modulation system and method utilizing whitening filter for updating of predictor coefficients

Abstract

An improved technique for processing digital audio signals is provided wherein adaptation of predictor coefficients in an ADPCM environment is caused to converge in a rapid and computationally efficient manner. The technique employs a whitening filter to generate a filtered reconstructed signal which is utilized to update, or adapt, the prediction coefficients of a pole-based predictor.

Description

BACKGROUND

[0002] 1. Field of Invention

[0003] The present invention relates generally to encoding and decoding of digital audio signals, and more particularly to predictor adaptation in adaptive differential pulse code modulation (ADPCM) systems.

[0004] 2. Description of the Prior Art

[0005] FIG. 1 may be referenced in conjunction with the following discussion. ADPCM is a well-known technique for encoding speech and other audio signals for subsequent transmission over a network. A standard implementation of such a system is described in the International Telecommunication Union (ITU-T) Recommendation G.722, 7 kHz Audio-Coding Within 64 kBit/s, which is incorporated by reference herein.

[0006] As described in U.S. Pat. No. 4,317,208, issued Feb. 23, 1982 to Araseki et al. and incorporated by reference herein, a differential pulse code modulation system is a band compression system in which a prediction of each signal sample at a present time period is based on signal samples at past time periods. Such a process is particularly effective with voice and similar band signals due to their high degree of correlation between successive signal samples. A predicted signal Sj at a time j is typically derived at a transmitter 102 by the general equation:

S j =A 1 S j−1 +A 2 S j−2 + . . . A n S j−n ;

[0007] where A1, A2, . . . An are termed the prediction coefficients. The prediction coefficients are selected to minimize the difference between an input signal Yj and the predicted signal Sj thus minimizing a prediction error Ej which is in turn quantized and transmitted to a receiver 104, thereby requiring significantly less transmission bandwidth than would the input signal. The receiver 104 works in a manner generally the reverse of the transmitter 102, thereby reconstructing the input signal.

[0008] The characteristics of a voice or related audio signal vary with time, consequently the optimum coefficient values also vary. One method of attempting to efficiently derive prediction coefficients is to adapt them with the goal of minimizing the prediction error Ej while such error is being observed, which could generally describe an ADPCM system. A common type of predictor employed in these systems is a pole-based predictor, such as predictors 110 and 126, which utilizes a feedback loop to minimize the energy in the prediction error signal Ej, which is sometimes referred to as the difference or residual signal.

[0009] Due to the reality of frequent transmission errors between the transmitter 102 and the receiver 104, the prediction errors Êj (which have been inverse quantized) produced at the receiver 104, and thus the reconstructed input signal Śj depending thereon, has a tendency to diverge from the real input signal Yj received at the transmitter 102. To gradually eliminate the adverse effect of the transmission errors, the prediction coefficients are typically derived by the general equation:

A i j+1 =A i j (1−δ)+g·F1(S1j−1)·F2(Êj);

[0010] where j=1 to n, δ is a positive value much smaller than 1, g is a proper positive constant, Śj−i is a reconstructed input signal delayed i samples, and F1 and F2 are non-decreasing functions. The receiver 104 prediction coefficient values are tracked, or gradually caused to converge to those of the transmitter 102, by operation of the term (1−δ). The detrimental effect of transmission errors is thus partially overcome.

[0011] Instability or oscillation of the receiver may still occur in pole-based predictor systems due to the feedback loop to the predictor, which uses the prediction error signal Êj and the preceding reconstructed input signal Śj−i to derive the prediction coefficients as described above. Stability checking is often used to ensure that the prediction coefficients remain in desired ranges, but at the expense of increased complexity as the number of poles, i.e., coefficients, increases.

[0012] In U.S. Pat. No. 4,317,208, Araseki et al. describe a system that also employs zero-based predictors, such as predictors 120 and 128, which do not utilize a feedback loop but which are known to provide less predictor gain than pole-based predictors and consequently inhibit or slow down the adaptation process. They propose that such a combination of pole-based and zero-based predictors may overcome the instability issues described above, and gain the advantages of each type of predictor.

[0013] In U.S. Pat. No. 4,593,398, issued Jun. 3, 1986 to Millar and incorporated by reference herein, it is suggested that a pole-based predictor, even coupled with a zero-based predictor, is still vulnerable to mistracking if the input signal contains two tones of equal amplitude but different frequencies. Millar notes that certain input signals may cause the pole-based predictor adaptation driven by the feedback loop to have multiple stable states, thus the receiver 104 may stabilize with its prediction coefficients at values different than the transmitter 102. This in turn is likely to cause a distorted frequency response at the receiver 104 and its associated audio output device.

[0014] The Millar patent proposes to mitigate the problems associated with lower predictor gain in zero-based predictors and mistracking in pole-based predictors. The system described by Millar and depicted in FIG. 1 is such that the predictor means in the transmitter 102 and the receiver 104 derive the prediction coefficients based on an algorithm including a non-linear function having no arguments comprising the value of the reconstructed input signal, such as signals Śj and Śj−i. This coefficient adaptation is depicted by arrows 119 and 127. This is in contrast to the Araseki system wherein the prediction coefficients are partially derived from a reconstructed input signal such as signal Śj−i, which is dependent upon the predicted signal Sj, which is dependent upon all of the immediate past coefficient values.

[0015] It is postulated that the Millar system and method may be computationally expensive to implement. Therefore what is needed is a system and methods in which the convergence to the optimal prediction coefficients, and thus to the predicted signal Sj, occurs more rapidly and efficiently than in prior art systems.

SUMMARY

[0016] An improved adaptive differential pulse code modulation (ADPCM) system and method comprises an encoder and a decoder linked together by a network connection and configured for processing digital audio signals. More particularly, the technique described is related to adaptation of predictor coefficients in an ADPCM environment. The components of the system may be implemented in software form as instructions executable by a processor or in hardware form as digital circuitry. Furthermore, devices implementing the system and method described are preferably configured to include both an encoder and a decoder for bidirectional communication with a similarly situated remote device, or may be configured with solely the encoder or decoder.

[0017] At the encoder, a digitized input signal is applied to a subtractor, which derives a difference signal by subtracting from the input signal a predicted signal generated by a pole-based predictor. After quantizing, transmitting to a decoder, and inverse quantizing, the difference signal is added to the predicted signal by an adder to provide a reconstructed input signal, which is fed back to the predictor and to the subtractor. The encoder is additionally provided with a whitening filter for receiving the reconstructed input signal and applying thereto a filtering algorithm to generate a filtered reconstructed signal. The filtered reconstructed signal is utilized to update, or adapt, the prediction coefficients of the pole-based predictor, thus providing more rapid and computationally efficient convergence to optimal prediction coefficients.

[0018] The decoder operates in an inverse manner to the encoder, receiving the quantized difference signal from an encoder and processing it to reconstruct the input signal for delivery to sound reproducing means. It is noted that devices employing the ADPCM techniques described herein are interoperable with devices employing prior art techniques, for example, those described in ITU-T G.722. It is further noted that the techniques described herein may be adapted for various implementations, one example being the employment of a plurality of encoders and/or decoders for frequency sub-band processing.

[0019] Other embodiments of the invention comprise additional predictors at the encoder and the decoder, operating to maximize the signal-to-noise ratio for certain input signals. The additional predictors are preferably zero-based predictors, the output therefrom being summed with the pole-based predictor output to produce the predicted signal.

BRIEF DESCRIPTION OF THE FIGURES

[0020] In the accompanying drawings:

[0021] FIG. 1 depicts a prior art ADPCM system;

[0022] FIG. 2 depicts an ADPCM system, in accordance with a first embodiment of the invention; and

[0023] FIG. 3 depicts another ADPCM system, in accordance with a second embodiment of the invention.

DESCRIPTION OF PREFERRED EMBODIMENTS

[0024] FIG. 2 depicts a first embodiment of an ADPCM system 200 in accordance with the invention. ADPCM system 200 comprises an encoder 202 and decoder 204 linked in communication by a network connection 206, such as an ISDN line, fractional T1 line, digital satellite link, wireless modems, or like digital transmission service. At encoder 202, a digitized input signal, typically representative of speech, is applied to a conventional subtractor 208. The input signal is represented as Yj, signifying a value at sample period j. Subtractor 208 derives a difference signal Ej by subtracting from input signal Yj a predicted signal Sj generated by a pole-based predictor 210. The difference signal Ej is quantized by a conventional quantizer 212 to obtain a quantized numerical representation Nj for transmission to decoder 204 over the network connection 206. Quantizer 112 is preferably of the adaptive type, but a quantizer utilizing fixed step sizes may also be used.

[0025] Numerical representation Nj is also applied to a conventional inverse quantizer 214, which derives a regenerated difference signal Dj. A conventional adder 216 adds regenerated difference signal Dj to a predicted signal Sj (output by the pole-based predictor 210) to provide a reconstructed input signal Xj. The reconstructed input signal Xj is in turn applied to the pole-based predictor 210, which calculates the predicted signal Sj in accordance with the following equation: 1$S_j=a_1^j{S}_{j-1}+a_2^j{S}_{j-2}+\dots +a_n^j{S}_{j-n}$

[0026] where Sj−1 is a stored value of the predicted signal at sample period j−1, Sj−2 is a stored value of the predicted signal at sample period j−2, and so on, and a1j to anj are the predictor coefficients at sample period j, where n corresponds to the total number of poles (i.e., coefficients) of pole-based predictor 210. In one implementation of ADPCM system 200, the pole-based predictor 210 is limited to two poles, yielding the relation: 2$S_j=a_1^j{S}_{j-1}+a_2^j{S}_{j-2}.$

[0027] The predicted signal Sj generated by predictor 210 is then applied to adder 216, completing the feedback loop.

[0028] Predictor coefficients a1j and a2j are updated in accordance with the generalized equations: 3$a_1^{j+1}=a_1^j\left(1-{\delta }_1\right)+g_1·F_1\left(X_j^f,X_{j-1}^f,X_{j-2}^f\right)$$a_2^{j+1}=a_2^j\left(1-{\delta }_2\right)+g_2·F_2\left(X_j^f,X_{j-1}^f,X_{j-2}^f,X_{j-3}^f,a_1^j\right)$

[0029] where Xfj is a filtered version of reconstructed input signal Xj at sample period j; δ1, δ2, g1 and g2 are proper positive constants, and F1 and F2 are nonlinear functions which may consist of correlations, sign-correlations, or other relationships. Calculation of the filtered reconstructed signal Xfj is discussed below.

[0030] In general, whitening filters modify the spectrum of signals to provide a flatter signal spectrum, so that there is less variation of energy as a function of frequency. It is noted that a perfect white noise signal has equal energy at every frequency. Stochastic gradient adaptive filters generally converge more rapidly with white signals than with non-white signals. Therefore, the use of a whitening filter in the present system and method is preferred at least for its effect on convergence of the adaptive pole-based predictors 210 and 226.

[0031] Referring back to FIG. 2, a whitening filter 218 receives the reconstructed input signal Xj and applies thereto a filtering algorithm to generate a filtered reconstructed signal Xfj. To ensure stable operation of whitening filter 218, the filter coefficients a2fj+1 and a1fj+1 undergo the clamping set forth below at every other time step (i.e., for odd values of j):

[0032] a2fj+1 is clamped to a maximum of 12288 and a minimum of −12288; and

[0033] a1fj+1 is i clamped in magnitude to 15360−a2fj+1.

[0034] Implementation of this clamping routine is exemplified as:

temp=15360−a2fj+1;

[0035] if a1fj+1>temp, then a1fj+1 is set to equal temp;

[0036] if a1fj+1<−temp, then a1fj+1 is set to equal −temp.

[0037] The filtered reconstructed signal Xfj output by whitening filter 218 is utilized to update the predictor coefficients a1j+1 and a2j+1, as described above and indicated on FIG. 2 by arrow 220.

[0038] According to a preferred implementation, whitening filter 218 has two zeroes, yielding the relation:

X j f =X j −a 1 f X j−1 −a 2 f X j−2

[0039] where af1 and af2 are the first and second order filter coefficients. The filter coefficients af1 and af2 are updated at each time step j in accordance with the following equations: 4$\begin{array}{c}{a}_2^{f^{j+1}}=a_2^{f^j}\left(1-\left(\frac{256}{32768}\right)\right)-\left(\frac{1}{32}\right)a_1^{f^j}\mathrm{sgn}\left[X_j^f\right]\mathrm{sgn}\left[X_{j-1}^f\right]+\\ 128*\mathrm{sgn}\left[X_j^f\right]\mathrm{sgn}\left[X_{j-2}^f\right];\mathrm{and}\end{array}$$a_1^{f^{j+1}}=a_1^{f^j}\left(1-\left(\frac{128}{32768}\right)\right)+192*\mathrm{sgn}\left[X_j^f\right]\mathrm{sgn}\left[X_{j-1}^f\right];$

[0040] where sgn [ ] is the sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument.

[0041] In accordance with a computationally economical implementation of ADPCM system 200, the values of the predictor coefficients may be frozen at every other sample interval j. It should be noted that by recalculating predictor coefficients for pole-based predictor 210 only at every other interval, computational resources are conserved. This implementation is described by the following equations:

[0042] for even j:

a2j+1=a2j; and

a1j+1=a1j;

[0043] else for odd j: 5$\begin{array}{c}{a}_2^{j+1}=a_2^{j-1}\left(1-\left(\frac{510}{32768}\right)\right)-\left(\frac{1016}{32768}\right)\lim \left[a_1^{j-1}\right]\mathrm{sgn}\left[X_{j-1}^f\right]\mathrm{sgn}\left[X_{j-2}^f\right]+\\ 127*\mathrm{sgn}\left[X_{j-1}^f\right]\mathrm{sgn}\left[X_{j-3}^f\right]-\left(\frac{1}{32}\right)\lim \left[a_1^{j-1}\right]\mathrm{sgn}\left[X_j^f\right]\mathrm{sgn}\left[X_{j-1}^f\right]+\\ 128*\mathrm{sgn}\left[X_j^f\right]\mathrm{sgn}\left[X_{j-2}^f\right];\end{array}$$\mathrm{and}$$\begin{array}{c}{a}_1^{j+1}=a_1^{j-1}\left(1-\left(\frac{127.5}{32768}\right)\right)+191.25*\mathrm{sgn}\left[X_{j-1}^f\right]\mathrm{sgn}\left[X_{j-2}^f\right]+\\ 192*\mathrm{sgn}\left[X_j^f\right]\mathrm{sgn}\left[X_{j-1}^f\right];\end{array}$

[0044] where sgn [ ] is the sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument, and 6$\lim \left[a_1^{j-1}\right]=a_1^{j-1}\mathrm{for}-8192\le a_1^{j-1}\le 8191;$$\lim \left[a_1^{j-1}\right]=-8192\mathrm{for}{a}_1^{j-1}<-8192;\text{and}$$\lim \left[a_1^{j-1}\right]=8191\mathrm{for}{a}_1^{j-1}>8191.$

[0045] To ensure stability, a2j+1 and a1j+1 are clamped similarly to a1fj+1 and a1fj+1 as described above. That is:

[0046] a2j+1 is clamped to a maximum of 12288 and a minimum of −12288; and

[0047] a1j+1 is clamped in magnitude to 15360−α2j+1.

[0048] Implementation of this clamping routine is exemplified as:

temp=15360−a2j+1;

[0049] if a1j+1>temp, then a1j+1 is set to equal temp;

[0050] if a1j+1<−temp, then a1j+1 is set to equal −temp.

[0051] Decoder 204 operates in an inverse manner to encoder 202. Inverse quantizer 222 receives the numerical representation Nj over network connection 206 and derives the regenerated difference signal Dj. Adder 224 sums the regenerated difference signal Dj with the predicted signal Sj generated by pole-based predictor 226 to produce the reconstructed input signal Xj. The reconstructed input signal Xj is then delivered to sound-reproducing means (which will typically include a D/A converter and loudspeaker) for reproduction of the speech represented by the input signal Yj.

[0052] At the decoder 204, the reconstructed input signal Xj is additionally applied to whitening filter 230 and pole-based predictor 226. Pole-based predictor 226 operates in a substantially identical manner to pole-based predictor 210 of encoder 202 and generates as output predicted signal Sj, which is applied to adder 224 to complete the feedback loop. Whitening filter 230, which operates in a substantially identical manner to whitening filter 218 of encoder 202, provides as output a filtered reconstructed signal Xfj for use by pole-based predictor 226 in updating the predictor coefficients, as discussed above and indicated on FIG. 2 by arrow 228.

[0053] Those skilled in the art will recognize that the various components of encoder 202 and decoder 204 will typically be implemented in software form as program instructions executable by a general purpose processor. Alternatively, one or more components of encoder 202 and/or decoder 204 may be implemented in hardware form as digital circuitry.

[0054] Additionally, those skilled in the art will recognize that, although the pole-based predictors 210 and 226 are described above in terms of a two-pole implementation, the invention is not limited thereto and may be implemented in connection with pole-based predictors having any number of poles.

[0055] It is additionally noted that the ADPCM technique embodied in the invention may be adapted in various well-known ways in order to improve the speed and performance of the encoding and decoding processes. For example, a transmitting entity may break the input signal into a plurality of frequency-limited sub-bands, wherein each sub-band is applied to a separate encoder operating in a substantially identical manner to encoder 202. The sub-banded encoded signals are then multiplexed for transmission to a receiving entity over the network connection. The receiving entity then demultiplexes the received signal into a plurality of sub-banded signals and directs each sub-banded signal to a separate decoder operating in a manner substantially identical to decoder 204. The sub-banded reconstructed signals are thereafter combined and conveyed to sound-reproducing means.

[0056] In other embodiments of the invention, additional predictors may be combined with the pole-based predictors to maximize the signal-to-noise ratio for certain input signals. Referring now to the FIG. 3 embodiment of an ADPCM system 300, encoder 302 differs from encoder 202 of the FIG. 2 embodiment by the addition of a conventional zero-based predictor 306. Zero-based predictor 306 receives the regenerated difference signal Dj and produces a zero-based partial predicted signal Sjz, which is added to the partial pole-based predicted signal Sjp (equal to Sj in the FIG. 2 embodiment) by adder 308 to provide predicted signal Sj. Predicted signal Sj is in turn applied to the feedback loop of pole-based predictor 210 and to subtractor 208. It is noted that zero-based predictor 306 does not have a feedback loop, and its predictor coefficients are conventionally updated with dependence on regenerated difference signal Dj.

[0057] Similarly, decoder 304 differs from decoder 204 of the FIG. 2 embodiment by the inclusion of zero-based predictor 310. The regenerated difference signal Dj is applied to zero-based predictor 310, which generates as output a zero-based partial predicted signal Sjz. Adder 312 combines the zero-based partial predicted signal Sjz with pole-based partial predicted signal Sjp provided by pole-based predictor 226 to produce the predicted signal Sj.

[0058] Another embodiment of the invention utilizes at least one look-up table in determining the proper coefficients for the predictors, i.e., pole-based predictors 210 and 226 of FIGS. 1 and 2, and/or zero-based predictors 306 and 310 of FIG. 3. For example, the first pole-based predictor coefficient is a function of three quantities: its former value, the sign of the current value of the sum of the quantized prediction error plus the all-zero predictor, and the sign of the past value of the sum of the quantized prediction error plus the all-zero predictor. In this embodiment, no arithmetic is necessary in determining a prediction coefficient value, however, identical input-output characteristics of the predictors are preserved.

[0059] It should be appreciated that devices utilizing the above-described ADPCM techniques, such as audioconferencing or videoconferencing endpoints, will typically be equipped for bi-directional communications over the network connection, and so will be provided with both an encoder (such as encoder 202 or 302) for encoding local audio for transmission to a remote endpoint as well as a decoder (such as decoder 204 or 304) for decoding audio signals received from the remote endpoint.

[0060] It is further noted that devices employing the above-described ADPCM techniques of the invention are advantageously interoperable with devices employing some prior art ADPCM techniques, such as those described in the aforementioned Millar reference and the ITU-T G.722 reference.

[0061] Finally, it is generally noted that while the invention has been particularly shown and described with reference to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention.

Claims

1. An adaptive differential pulse code modulation system comprising: an encoder including: a subtractor configured for deriving a difference signal Ej, the difference signal Ej being the difference between an input signal Yj and a predicted signal Sj, j representing a sample period; a quantizer configured for quantizing the difference signal Ej to obtain a numerical representation Nj for transmission to an encoder inverse quantizer for deriving a regenerated difference signal Dj, and to a decoder inverse quantizer coupled to the quantizer through a network for deriving the regenerated difference signal Dj, an encoder adder configured for deriving a reconstructed input signal Xj, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; an encoder whitening filter Fe configured for receiving the reconstructed input signal Xj and for generating a filtered reconstructed signal Xfj, the Xjf=Xj−a1fXj−1a2fXj−2− . . . anfXj−nfiltered reconstructed signal Xfj being generated according to the equation: Xj, being a value of reconstructed input signal Xj at sample period j−n, and; n being a number of filter tap coefficients afn corresponding to the whitening filter Fe; an encoder predictor Pep configured for receiving the reconstructed input signal Xj and for generating a predicted signal Sjp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: 7Sjp=a1j⁢Sj-1+a2j⁢Sj-2⁢ ⁢…⁢ ⁢anpj⁢Sj-npSj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of predictor coefficients ajnp corresponding to the predictor Pep; and an encoder feedback loop configured for applying the predicted signal Sj to the adder; transmission means configured for transmitting the numerical representation Nj from the encoder to a decoder; and the decoder including: the decoder inverse quantizer coupled to the quantizer through a network and configured for receiving the numerical representation Nj and for deriving the regenerated difference signal Dj therefrom, a decoder adder configured for deriving the reconstructed input signal Xj, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; a decoder whitening filter Fd configured for receiving the reconstructed input signal Xj and for generating the filtered reconstructed signal Xfj, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xjaf1Xj−1−af2Xj−2− . . . afnXj−n Xj−n being a value of reconstructed signal Xj at sample period j−n, and n being the number of filter tap coefficients afn corresponding to the whitening filter Fd; a decoder predictor Pdp configured for receiving the reconstructed input signal Xj and for generating a predicted signal Sjp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=a1jSj−1+a2jSj−2 . . . ajnpSj−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being the number of predictor coefficients ajnp corresponding to the predictor Pdp; and a decoder feedback loop configured for applying the predicted signal Sj to the decoder adder.

2. The system of claim 1, further comprising: a second encoder predictor Pez configured for receiving the regenerated difference signal Dj and for generating a predicted signal Sjx; a second encoder adder configured for deriving the predicted signal Sj at the encoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz; a second decoder predictor Pdz configured for receiving the regenerated difference signal Dj and for generating a predicted signal Sjz; and a second decoder adder configured for deriving the predicted signal Sj at the decoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz.

3. The system of claim 1 wherein: np is 2; the predictor coefficient a1j is updated according to the equation: a1j+1=a1j(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and the predictor coefficient a2J is updated according to the equation: a2j+1=a2j(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1j); δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

4. The system of claim 1 wherein: n is 2; the filter tap coefficient a1f is updated at each sample period j according to the generalized equation: a1fj+1=a1fj(1−δ1)+g1·F1(Xjf, Xfj−1, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and the filter tap coefficients a2f is updated at each sample period j according to the generalized equation: a2fj+1=a2fj(1−δ2)+g2·F2(Xjf, Xj−1f, a1fj) δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

5. The system of claim 4 wherein: the filter tap coefficient a1fj is updated according to the equation: 8a1fj+1=a1fj⁡(1-(12832768))+192*sgn⁡[Xjf]⁢sgn⁢ [Xj-1f]⁢ ; ⁢andthe filter tap coefficient a2fj is updated according to the equation: 9a2fj+1= ⁢a2fj⁡(1-(25632768))-(132)⁢a1fj⁢sgn⁢ [Xjf]⁢sgn⁡[Xj-1f]+ ⁢128*sgn⁢ [Xjf]⁢sgn⁢ [Xj-2f];sgn[ ] being a sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument.

6. The system of claim 5 wherein at every other sample period j, the filter tap coefficient afj+12 is maintained in a range −12288≦afj+12≦12288; and the filter tap coefficient afj+11 is maintained in a range −(15360−afj+12)≦afj+11≦(15360−afj+12); whereby afj+11 is set equal to (15360−afj+12) when afj+11>15360−afj+12; and whereby afj+11 is set equal to −(15360−afj+12) when afj+1121 −(15360−afj+12).

7. The system of claim 5, further comprising: a second encoder predictor Pez configured for receiving the regenerated difference signal Dj and for generating a predicted signal Sjz; a second encoder adder configured for deriving the predicted signal Sj at the encoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz; a second decoder predictor Pdz configured for receiving the regenerated difference signal Dj and for generating a predicted signal Sjz; and a second decoder adder configured for deriving the predicted signal Sj at the decoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz.

8. The system of claim 1 wherein at every other sample period j, the predictor coefficient ajnp corresponding to the predictors Pep and Pdp is maintained unchanged.

9. The system of claim 8, such that if for even j:

10. An encoder for encoding digital audio signals, comprising: a subtractor configured for deriving a difference signal Ej, the difference signal Ej being the difference between an input signal Yj and a predicted signal Sj, j representing a sample period; a quantizer configured for quantizing the difference signal Ej to obtain a numerical representation Nj for transmission to an encoder inverse quantizer for deriving a regenerated difference signal Dj, and to a decoder inverse quantizer coupled to the quantizer for deriving the regenerated difference signal Dj; an adder configured for deriving a reconstructed input signal Xj, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; a whitening filter configured for receiving the reconstructed input signal Xj and for generating a filtered reconstructed signal Xfj, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−nXfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of filter tap coefficients afn corresponding to the whitening filter; a predictor configured for receiving the reconstructed input signal Xj and for generating a predicted signal Sjp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=aj1Sj−1−aj2Sj−2− . . . ajnpSn−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of predictor coefficients ajnp corresponding to the predictor; and a feedback loop configured for applying the predicted signal Sj to the adder.

11. The system of claim 10, the encoder further comprising: a second predictor configured for receiving the regenerated difference signal Dj and for generating a predicted signal Sjz, the predicted signal Sjz being at least constituent to predicted signal Sj; and a second adder configured for deriving the predicted signal Sj, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz.

12. The system of claim 10 wherein: n is 2; the filter tap coefficient a1f is updated at each sample period j according to the generalized equation: a1fj+1=a1fj(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; the filter tap coefficients a2f is updated at each sample period j according to the generalized equation: a2fj+1=a2fj(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1fj) δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

13. The system of claim 12 wherein: the filter tap coefficient a1f is updated according to the equation: 11a1fj+1=a1fj⁡(1-(12832768))+192*sgn⁡[Xjf]⁢sgn⁢ [Xj-1f]⁢ ⁢andthe filter tap coefficient a2f is updated according to the equation: 12a2fj+1= ⁢a2fj⁡(1-(25632768))-(132)⁢a1fj⁢sgn⁢ [Xjf]⁢sgn⁡[Xj-1f]+ ⁢128*sgn⁢ [Xjf]⁢sgn⁢ [Xj-2f],sgn[ ] being a sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument.

14. The system of claim 13 wherein at every other sample period j, the filter tap coefficient afj+12 is maintained in a range −12288≦afj+12≦12288; and the filter tap coefficient afj+11 is maintained in a range −(15360−afj+12)≦afj+11≦(15360- afj+12); whereby afj+11 is set equal to (15360−afj+12) when afj+11>15360−afj+12; and whereby afj+11 is set equal to −(15360−afj+12) when afj+11<−(15360−afj+12).

15. The system of claim 10 wherein at every other sample period j, the predictor coefficient ajnp corresponding to the predictor is maintained unchanged.

16. The system of claim 10, wherein the encoder is constituent to or coupled to a videoconferencing device or application.

17. A decoder for decoding digital audio signals encoded by a properly associated encoder, comprising: an inverse quantizer coupled to the encoder and configured for receiving a numerical representation Nj and for deriving a regenerated difference signal Dj therefrom, the numerical representation Nj being a quantized representation of a difference signal Ej, the difference signal Ej being the difference between an input signal Yj and a predicted signal Sj, j representing a sample period; an adder configured for deriving a reconstructed input signal Xj, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; a whitening filter configured for receiving the reconstructed input signal Xj and for generating a filtered reconstructed signal Xfj, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfn−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of filter tap coefficients afn corresponding to the whitening filter; a predictor configured for receiving the reconstructed input signal Xj and for generating a predicted signal Sjp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=aj1Sj−1−aj2Sj−2 . . . ajnpSj−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of predictor coefficients ajnp corresponding to the predictor; and a feedback loop configured for applying the predicted signal Sj to the adder.

18. The system of claim 17, the decoder further comprising: a second predictor configured for receiving the regenerated difference signal Dj and for generating a predicted signal Sjz, the predicted signal Sjz being at least constituent to predicted signal Sj; and a second adder configured for deriving the predicted signal Sj, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz.

19. The system of claim 17 wherein: n is 2; the filter tap coefficient a1f is updated at each sample period j according to the generalized equation: a1fj+1=a1fj(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; the filter tap coefficients a2f is updated at each sample period j according to the generalized equation: a2fj+1=a2fj(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1fj) δ2 and g2 being proper positive constants, and; F2 being a nonlinear function.

20. The system of claim 19 wherein: the filter tap coefficient a1f is updated according to the equation: 13a1fj+1=a1fj⁡(1-(12832768))+192*sgn⁡[Xjf]⁢sgn⁢ [Xj-1f]⁢ ⁢andthe filter tap coefficient a2f is updated according to the equation: 14a2fj+1= ⁢a2fj⁡(1-(25632768))-(132)⁢a1fj⁢sgn⁢ [Xjf]⁢sgn⁡[Xj-1f]+ ⁢128*sgn⁢ [Xjf]⁢sgn⁢ [Xj-2f]sgn[ ] being a sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument.

21. The system of claim 20 wherein at every other sample period j, the filter tap coefficient afj+12 is maintained in a range −12288≦afj+12≦12288; and the filter tap coefficient afj+11 is maintained in a range −(15360−afj+12)≦afj+11≦(15360−afj+12); whereby afj+11 is set equal to (15360−afj+12) when afj+11>15360−afj+12; and whereby afj+11 is set equal to −(15360−afj+12) when afj+11<−(15360−afj+12).

22. The system of claim 17 wherein at every other sample period j, the predictor coefficient ajnp corresponding to the predictor is maintained unchanged.

23. The system of claim 17, wherein the decoder is constituent to or coupled to a videoconferencing device or application.

24. A method for encoding and decoding digital audio signals, comprising the steps of: deriving a difference signal Ej at an encoder, the difference signal Ej being the difference between an input signal Yj and a predicted signal Sj, j representing a sample period; quantizing the difference signal Ej to obtain a numerical representation Nj for transmitting to an encoder inverse quantizer for deriving a regenerated difference signal Dj, and to a decoder inverse quantizer coupled to the quantizer through a network for deriving the regenerated difference signal Dj; deriving a reconstructed input signal Xj at a first adder, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; receiving the reconstructed input signal Xj at a whitening filter Fe; generating a filtered reconstructed signal Xfj by the whitening filter Fe, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of filter tap coefficients afn corresponding to the whitening filter Fe; receiving the reconstructed input signal Xj at a predictor Pep; generating a predicted signal Sjp by the predictor Pep, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=aj1Sj−1−aj2Sj−2− . . . ajnpSj−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of predictor coefficients ajnp corresponding to the predictor Pep; applying the predicted signal Sj to the first adder to provide feedback; receiving the numerical representation Nj at a decoder; deriving the regenerated difference signal Dj from the numerical representation Nj, deriving the reconstructed input signal Xj at a second adder, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; receiving the reconstructed input signal Xj at a whitening filter Fd; generating a filtered reconstructed signal Xfj by the whitening filter Fd, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n; n being a number of filter tap coefficients afn corresponding to the whitening filter Fd; receiving the reconstructed input signal Xj at a predictor Pdp; generating a predicted signal Sjp by the predictor Pdp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=aj1Sj−1−aj2Sj−2− . . . ajnpSj−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of predictor coefficients ajnp corresponding to the predictor Pdp; and applying the predicted signal Sj to the second adder to provide feedback.

25. The method of claim 24, further comprising the steps of: receiving the regenerated difference signal Dj at a predictor Pez at the encoder; generating a predicted signal Sjz by the predictor Pez; deriving the predicted signal Sj at the encoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz; receiving the regenerated difference signal Dj at a predictor Pdz at the decoder; generating the predicted signal Sjz by the predictor Pdz; and deriving the predicted signal Sj at the decoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz.

26. The method of claim 24 wherein np is 2, further comprising the steps of: updating the predictor coefficient a1j according to the equation: a1j+1=a1j(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating the predictor coefficient a2j according to the equation: a2j+1=a2j(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1j) δ2 and g2 being proper positive constants, and; F2 being a nonlinear function.

27. The method of claim 24 wherein n is 2, further comprising the steps of: updating the filter tap coefficient a1f at each sample period j according to the generalized equation: a1fj+1=a1fj(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating the filter tap coefficients a2f at each sample period j according to the generalized equation: a2fj+1=a2fj(1−δ2)+g2·F2(Xjf, Xj−1f, a1fj) δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

28. The method of claim 27 wherein: the filter tap coefficient a1f is updated according to the equation: 15a1fj+1=a1fj⁡(1-(12832768))+192*sgn⁡[Xjf]⁢sgn⁢ [Xj-1f]⁢ , ⁢andthe filter tap coefficient a2f is updated according to the equation: 16a2fj+1= ⁢a2fj⁡(1-(25632768))-(132)⁢a1fj⁢sgn⁢ [Xjf]⁢sgn⁡[Xj-1f]+ ⁢128*sgn⁢ [Xjf]⁢sgn⁢ [Xj-2f]sgn[ ] being a sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument.

29. The method of claim 28 wherein at every other sample period j, the filter tap coefficient afj+12 is maintained in a range −12288≦afj+12≦12288; and the filter tap coefficient afj+11 is maintained in a range −(15360−afj+12)≦afj+11≦(15360−afj+12); whereby afj+11 is set equal to (15360−afj+12) when afj+1122 15360−afj+12; and whereby afj+11 is set equal to −(15360−afj+12) when afj+11<−(15360−afj+12).

30. The method of claim 28, further comprising the steps of: receiving the regenerated difference signal Dj at a predictor Pez at the encoder; generating a predicted signal Sjz by the predictor Pdz; deriving the predicted signal Sj at the encoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz; receiving the regenerated difference signal Dj at a predictor Pdz at the decoder; generating the predicted signal Sjz by the predictor Pdz; and deriving the predicted signal Sj at the decoder, the predicted signal Sj being the sum of the predicted signal Sjp and the predicted signal Sjz.

31. The method of claim 28 wherein np is 2, further comprising the steps of: updating the predictor coefficient a1j according to the equation: a1j+1=a1j(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating the predictor coefficient a2j according to the equation: a2j+1=a2j(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1j) δ2 and g2 being proper positive constants, and; F2 being a nonlinear function.

32. A method for adapting coefficients in a two pole predictor in an adaptive differential pulse code modulation system, comprising the steps of: generating a filtered reconstructed signal Xfj by a whitening filter Fe, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfn−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of filter tap coefficients afn corresponding to the whitening filter Fe; updating a predictor coefficient a1f according to the equation: a1j+1=a1j(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating a predictor coefficient a2j according to the equation: 17a2j+1=a2J⁢(1-δ2)+g2·F2⁢(Xjf,Xj-1f,Xj-2f,a1j)δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

33. The method of claim 32, further comprising the steps of: updating the filter tap coefficient a1f at each sample period j according to the generalized equation: a1fj+1=a1fj(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating the filter tap coefficients a2f at each sample period j according to the generalized equation: a2fj+1=a2fj(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1fj) δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

34. The method of claim 32 wherein: the filter tap coefficient a1f is updated according to the equation: 18a1fj+1=a1fj⁡(1-(12832768))+192*sgn⁡[Xjf]⁢sgn⁢ [Xj-1f]⁢ ⁢andthe filter tap coefficient a2f is updated according to the equation: 19a2fj+1= ⁢a2fj⁡(1-(25632768))-(132)⁢a1fj⁢sgn⁢ [Xjf]⁢sgn⁡[Xj-1f]+ ⁢128*sgn⁢ [Xjf]⁢sgn⁢ [Xj-2f]sgn[ ] being a sign function that returns a value of 1 for a nonnegative argument and a value of −1 for a negative argument.

35. The method of claim 34 wherein at every other sample period j, the filter tap coefficient afj+12 is maintained in a range −12288≦afj+12≦12288; and the filter tap coefficient afj+11 is maintained in a range −(15360−afj+12)≦afj+11≦(15360−afj+12); whereby afj+11 is set equal to (15360−afj+12) when afj+11>15360−afj+12; and whereby afj+11 is set equal to −(15360−afj+12) when afj+11<−(15360−afj+12).

36. A machine readable medium embodying instructions executable by a machine to perform a method for adapting coefficients in a two pole predictor in an adaptive differential pulse code modulation system, the method steps comprising: generating a filtered reconstructed signal Xfj by a whitening filter, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of filter tap coefficients afn corresponding to the whitening filter; updating a predictor coefficient a1j according to the equation: a1j+1=a1j(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating a predictor coefficient a2j according to the equation: a2j+1=a2j(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1j) δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

37. A digital circuit embodying instructions to perform a method for adapting coefficients in a two pole predictor in an adaptive differential pulse code modulation system, the method steps comprising: generating a filtered reconstructed signal Xfj by a whitening filter, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period i−n, and n being a number of filter tap coefficients afn corresponding to the whitening filter; updating a predictor coefficient a1j according to the equation: a1j+1=a1j(1−δ1)+g1·F1(Xjf, Xj−1f, Xj−2f) δ1 and g1 being proper positive constants, and F1 being a nonlinear function; and updating a predictor coefficient a2j according to the equation: a2j+1=a2j(1−δ2)+g2·F2(Xjf, Xj−1f, Xj−2f, a1j) δ2 and g2 being proper positive constants, and F2 being a nonlinear function.

38. An adaptive differential pulse code modulation system comprising: at a first instance: means for deriving a difference signal Ej, the difference signal Ej being the difference between an input signal Yj and a predicted signal Sj, j representing a sample period; means for quantizing the difference signal Ej to obtain a numerical representation Nj; means for deriving a regenerated difference signal Dj based on the numerical representation Nj; means for transmitting the numerical representation Nj to an inverse quantizing means coupled to the quantizing means through a network; means for deriving a reconstructed input signal Xj, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; means for generating a filtered reconstructed signal Xfj, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of coefficients afn corresponding to the means for generating a filtered reconstructed signal; means for generating a predicted signal Sjp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=aj1Sj−1−aj2Sj−2− . . . ajnpSj−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of predictor coefficients ajnp corresponding to the means for generating a predicted signal; and feedback means for applying the predicted signal Sj to the means for deriving a reconstructed input signal Xj; at a second instance: the inverse quantizing means for deriving the regenerated difference signal Dj from the numerical representation Nj; second means for deriving a reconstructed input signal Xj, the reconstructed input signal Xj being the sum of the regenerated difference signal Dj and the predicted signal Sj; second means for generating a filtered reconstructed signal Xfj, the filtered reconstructed signal Xfj being generated according to the equation: Xfj=Xj−af1Xj−1−af2Xj−2− . . . afnXfj−n Xfj−n being a value of filtered reconstructed signal Xfj at sample period j−n, and n being a number of coefficients afn corresponding to the second means for generating a filtered reconstructed signal; second means for generating a predicted signal Sjp, the predicted signal Sjp being at least constituent to predicted signal Sj and being generated according to the equation: Sjp=aj1Sj−1−aj2Sj−2− . . . ajnpSj−np Sj−np being a value of the predicted signal Sj at sample period j−np, and np being a number of coefficients ajnp corresponding to the means for generating a predicted signal; and feedback means for applying the predicted signal Sj to the means for deriving a reconstructed input signal Xj.