The invention relates to differential pulse code modulation systems and in particular such systems having adaptive prediction coefficients.
The invention is particularly concerned with adaptive differential pulse code modulation (ADPCM) systems of the kind exemplified as prior art in U.S. Pat. No. 4,317,208, issued Feb. 23, 1982 to Takashi Araseki, which is incorporated herein by reference. Generally such systems include a transmitter in which a subtractor provides the difference between the instant signal sample and a prediction signal derived from one or more earlier samples. The difference signal is then quantized and transmitted. The receiver includes an inverse quantizer and a predictor which reconstruct the signal from the received difference or residual signal. Usually the quantizers will be adaptive so as to vary the step size, or transfer function slope, according to the magnitude of the input difference signal. This better utilizes the dynamic range of the quantizer and improves response to low amplitude signals.
Additionally, each predictor may be adaptive, i.e. its coefficients change with time, to better follow the variations with time of the signal to be predicted, and to optimize performance with different types of signal, for example voice, voiceband data. In effect the predictor transfer function is adapted to the time varying input signal so that, ideally, the energy in the difference or residual signal is minimized at all times. In ADPCM systems, the values of the predictor coefficients are not transmitted explicitly to the receiver, but are derived from the quantized difference signal in an identical manner in both the transmitter and the receiver.
One type of predictor, known as "pole-based", uses a feedback loop and derives its coefficients according to the equation: ##EQU2## where A.sub.i.sup.j is the ith predictor coefficient at sample time j. X'.sub.j-i is the reconstructed signal delayed i samples, E.sub.j is the quantized difference between the input signal and the predicted value, g is a small positive value and F.sub.1 and F.sub.2 are non-decreasing functions. As discussed in U.S. Pat. No. 4,317,208, in such systems the coefficients of the receiver differ from those of the transmitter if transmission errors occur. This is because the prediction coefficients in the receiver are derived from the received difference signal. Errors in this signal cause the receiver prediction coefficients to depart from those in the transmitter. The difference or mistracking may persist even when the errors have ceased.
It has been proposed to alleviate this problem by deriving the prediction coefficients A.sub.i according to the equation: ##EQU3## where i=1 to N, .delta. is a positive value much smaller than 1, g is a proper positive constant, X.sub.j-i is the reconstructed signal delayed i samples, E.sub.j is the quantized difference between the input signal and the prediction signal and F.sub.1 and F.sub.2 are non-decreasing functions.
Inclusion of the term (1-.delta.) is intended to cause the receiver's predictor coefficient values to gradually converge to those of the transmitter predictor at a rate determined by .delta.. This desirable property is termed tracking of the receiver predictor coefficients.
Even so, instability or oscillation of the receiver may still occur because of the feedback loop in the predictor which uses both the difference signal E.sub.j and the preceding reconstructed signal X.sub.j-i to derive the predictor coefficients. Usually stability checking is used to ensure that the predictor coefficients remain within prescribed ranges. A drawback of such stability checking is the increased complexity as the number of poles (coefficients) increases.
In U.S. Pat. No. 4,317,208, Araseki proposes overcoming the stability problem by using a zero-based predictor, i.e. which does not have a feedback loop. However, whilst such zero-based predictors are not susceptible to instability, they do suffer from the disadvantage that they provide less prediction gain for speech and like signals than pole-based predictors. It is possible to use both a pole-based predictor and a zero-based predictor, as suggested by Araseki, to gain the advantages of each. However, it has been found that, whether combined with a zero-based predictor or not, a pole-based predictor is still vulnerable to mistracking if the input signal contains two tones of equal amplitude but different frequency. A particular problem arises with the tones used for dual tone multiple frequency (DTMF) signalling in the telephone network, but a problem may also arise with some modems which use tones differing by more than about 300 Hz. When such signals the predictor adaptation driven via the feedback loop by the predictor output signal, may have multiple stable states. Thus, once transmission errors have produced mistracking, the receiver may stabilize with its predictor coefficients at values different from those of the transmitter. Its transfer function, which is normally the inverse of that of the transmitter, will have a distorted frequency response, so one tone will be attenuated and the other amplified, possibly to an extent that the inequality is unacceptable.
In summary, zero-based predictors overcome the problems of instability and mistracking, but suffer from lower predictor gain. Pole-based predictors can be made stable by applying a stability check, but hitherto have suffered from mistracking.
The present invention seeks to mitigate this problem and to this end according to one aspect provides an adaptive differential pulse code modulation system comprising:
a transmitter including; a subtractor for deriving the difference (E.sub.j) between an input signal (X.sub.j) and a transmitter predicted value (X.sub.j), a quantizer for quantizing the difference signals E.sub.j from said subtractor to obtain a numeric representation (N.sub.j) thereof;
an inverse quantizer for regenerating the (quantized) difference signal E.sub.j from said numeric representation (N.sub.j);
summing means for summing the difference signal E.sub.j and the transmitter predicted value X.sub.j to provide a transmitter reconstructed signal X.sub.j ;
predictor means having variable prediction coefficients for receiving the output of said summing means and generating therefrom said predicted value;
a feedback loop for applying said predicted value to said summing means;
a receiver comprising an inverse quantizer for recovering the receiver quantized difference signal (E.sub.j) from the received numeric representation;
an adder for combining a receiver predicted value and a receiver quantized difference signal E.sub.j to provide a receiver reconstructed signal X.sub.j ; and
receiver predictor means having variable prediction coefficients, for receiving said reconstructed signal and providing therefrom said receiver predicted value.
Each predictor coefficient is derived as constant less than one times its immediate past value plus a non-linear function dependent upon a finite number of past values of the quantized difference signal (E.sub.j).
Where the predictor has a plurality of coefficients, corresponding to past values of the reconstructed signal occurring at different times, the corresponding non-linear function may be dependent also upon the immediate past coefficient values of one or more lower numbered predictor coefficients. It may not derive from any other, so precluding the function being derived from the predicted value or from the reconstructed signal. The numbering of coefficients for this purpose is arbitrary, but they are typically ordered such that higher numbered coefficients correspond to earlier past values of the reconstructed signal. Thus the first or lowest numbered predictor coefficient will not be a non-linear function of any past coefficient value. This is in contrast to prior implementations, such as that by Araseki, wherein this function is derived in part from the reconstructed signal X.sub.j, which derives in turn from the predicted value X.sub.j, which derives from all of the immediate past coefficient values.
More particularly the predicted values X.sub.j are derived in accordance with the equation:
where A.sub.1 to A.sub.n are the individual predictor coefficients A.sub.i derived in accordance with the equation: ##EQU4## where i=1 . . . n, the number of pole-predictor coefficients.
.delta..sub.i is a positive constant much smaller than one
g is a proper positive constant
E.sub.j is the value of the (quantized) difference signal at time point
j and
A.sub.i.sup.j is the respective predictor coefficient at time point j, and
F.sub.i is a non-linear function X'.sub.j is the sum of the difference signal E.sub.j and the predicted value X.sub.j.
It should be noted that for the case i=1, i.e. the first or only pole, there are no A arguments. Some, but not all, of the E.sub.j or A.sub.i.sup.j terms may be omitted.
In a preferred embodiment having a two pole predictor the two predictor coefficients A.sub.1 and A.sub.2 are derived in accordance with the equations: ##EQU5## where j is a particular sample period,
.delta..sub.1 and .delta..sub.2 are positive values much smaller than 1 (e.g. 1/256 and 1/128, respectively);
d is a small positive constant;
g.sub.1 and g.sub.2 are proper positive constants, for example each 1/32; and
K=Max (d, E.sub.j.sup.2, E.sub.j-1.sup.2, E.sub.j-2.sup.2)
The aforementioned embodiments of the invention, (with a pole-based predictor) may be used alone or with an additional predictor not employing feedback (zero-based). When such an additional predictor is provided, it may also be preferable to derive the prediction coefficients for the pole-based predictor not only from the difference signal, but also from the output of the zero-based predictor, i.e. from the partially reconstructed input signal.
The invention will be readily understood from the following description taken in conjunction with the accompanying drawings, in which:
FIG. 1 is a schematic representation of an ADPCM system according to the PRIOR ART;
FIG. 2 is a schematic representation of a first exemplary embodiment of the invention;
FIG. 3 is a schematic representation of a second exemplary embodiment of the invention; and
FIG. 4 is a schematic representation of a third exemplary embodiment of the invention.
Referring to FIG. 1, a conventional adaptive differential pulse code modulation system (ADPCM) with adaptive prediction comprises a transmitter 10 and a receiver 12. A digital signal to be transmitted is applied to an input terminal 14 of the transmitter 10. The signal is represented as X.sub.j, signifying it is applied at time point or sample period j. The input terminal 14 is connected to a subtractor 16, which provides a difference signal E.sub.j obtained by subtracting from the input signal X.sub.j the output X.sub.j of a pole-based predictor 18. The difference signal E.sub.j is quantized by a quantizer 20 to provide a corresponding numeric representation N.sub.j at transmitter output terminal 22 for transmission to the receiver 12.
Generally the quantizer 20 will be adaptive i.e. its step size or transfer function will vary according to the input signal magnitude. Such quantizers are known and so will not be described in detail here. It should be noted that although an adaptive quantizer is preferred, a fixed quantizer might be used instead.
The numerical representation N.sub.j is also applied to an inverse quantizer 24 which regenerates the difference signal E.sub.j. Naturally, the characteristics of the inverse quantizer 24 must match those of the quantizer 20, and so will be adaptive if quantizer 20 is adaptive.
An adder 26 sums the regenerated difference signal E.sub.j with the predictor output signal or predicted value X.sub.j to provide a reconstructed input signal X.sub.j at the input of predictor 18. The pole-based predictor 18 has a feedback loop 28 which applies the predictor output X.sub.j to the adder 26. The predictor 18 derives the signal X.sub.j using past input signal values in accordance with the equation:
where A.sub.1 -A.sub.n are prediction coefficients.
The predictor coefficients are adaptively corrected, as signified by arrow 30, in dependence upon the regenerated difference signal E.sub.j, as signified by the broken line 32, and upon the previously reconstructed input signal X.sub.j, as signified by the broken line 34. More specifically, the coefficients A.sub.i.sup.j are adaptively corrected in accordance with the equation: ##EQU6## where g is a positive small value and F.sub.1 and F.sub.2 are non-decreasing functions.
The receiver 12 comprises an inverse quantizer 124, and a pole-based predictor 118, corresponding to inverse quantizer 24 and predictor 18 in the transmitter 10. The receiver inverse quantizer 124 receives the numerical representation N.sub.j from the transmitter 10 and produces therefrom the regenerated difference signal E.sub.j. An adder 126 sums the output signal X.sub.j from the predictor 118 with the difference signal E.sub.j to produce the reconstructed input signal X.sub.j at the output terminal 122 of the receiver 12. This signal X.sub.j is also applied to the input of predictor 118.
The coefficients of receiver predictor 118 are adaptively corrected in like manner to those of the transmitter predictor 18 as indicated by corresponding broken lines 132 and 134.
The receiver 12 operates in the inverse manner to the transmitter 10 and will faithfully reconstruct the original signal so long as the predictor coefficients are the same in both predictors 18 and 118 at any instant in time. The receiver and transmitter are then said to be "tracking". As mentioned previously, however, in practice errors will occur in the transmission between the transmitter and the receiver. These errors will result in differences between the prediction coefficients of the predictors 18 and 118 so the receiver output will no longer faithfully reproduce the original signal. In most cases, once the errors have ceased, the coefficients in the receiver will realign with those in the transmitter. The mechanisms whereby this is achieved are stability checks which restrict the range of the predictor and the leakage factor (1-.delta.) so that they will converge. There is a limit to the extent to which these mechanisms can be applied whilst assuring adequate predictor performance.
It has been found that "mistracking" can occur, however, when the input signal comprises two tones of different frequency. "Mistracking" is a situation arising when the transmission errors have ceased and the prediction coefficients in the transmitter and receiver have stabilized, but are not the same. The effect then is to amplify one tone and attenuate the other.
In embodiments of the present invention the problem is overcome by not using the reconstructed input signal X.sub.j to adjust the predictor coefficients. Thus, referring to FIG. 2, which illustrates a first embodiment of the invention, the component parts of the transmitter 10A and receiver 12A are the same as those illustrated in FIG. 1 and so for ease of description corresponding parts are identified by the same reference numeral. It should be noted, however, that in FIG. 2 there are no broken lines corresponding to lines 34 and 134 in FIG. 1. This is because the prediction coefficients are no longer dependent upon X.sub.j, the reconstructed input signal.
In the embodiment shown in FIG. 2, the predictor coefficients are derived in accordance with the general equation: ##EQU7## where i is the number of the coefficient from 1 to n, the higher numbers corresponding to earlier time values.
.delta..sub.i is a positive constant much smaller than one;
g is a proper positive constant;
E.sub.j is the value of the difference signal at time point j;
A.sub.i.sup.j is the respective preditor coefficient at time point j; and
F.sub.i is a non-linear function.
It should be noted that when i=1, the function F.sub.i will not have any arguments A.sub.1 -A.sub.i-1.
This approach to adapting the predictor coefficients avoids any dependence upon the output at the predictor X.sub.j in the receiver so that mistracking due to dual tones is avoided. The main features guaranteeing tracking are:
(a) F.sub.i depends on a finite number n of past values of the difference signal E.sub.j.
(b) F.sub.i depends on A.sub.k.sup.j only up to k=i-1 (Thus A.sub.1.sup.j does not depend on any A.sub.k).
In these equations for coefficients A.sub.i, the first term is a linear decay term to allow effects of transmission errors to die away, and the second term is the adaptation term. Whilst a system for transmitting signals, such as speech, over telecommunications networks will usually require the first term, it should be appreciated that the .delta..sub.i constant might be omitted in some ADPCM system applications.
Although the system illustrated in FIG. 2 can be used with only one pole, or several poles, it is preferred to use two poles in the predictors 18 and 118.
In such a case, the predictor coefficients A.sub.1 and A.sub.2 are derived according to the equations: ##EQU8## where K.sub.j =Max (d, E.sub.j.sup.2, E.sub.j-1.sup.2, E.sub.j-2.sup.2)
.delta..sub.1 and .delta..sub.2 are positive values much smaller than 1 (for example 1/256 and 1/128, respectively).
g.sub.1, g.sub.2 and d are proper positive constants, e.g. 1/32, 1/32 and 10.sup.-6, respectively.
The values for g.sub.1 and g.sub.2 are chosen depending upon the characteristics of the signal and those specified are typically suitable for speech. Other values may be used providing the ratios of .delta..sub.1 :g.sub.1 and .delta..sub.2 :g.sub.2 are maintained about 1:8 and 1:4, respectively.
A particularly economic implementation of the invention can be achieved by approximating the coefficient equations as follows: ##EQU9## and .delta..sub.1 and .delta..sub.2 are about 1/181 and 1/90, respectively, and g.sub.1 and g.sub.2 are 1/64 and 1/90, respectively.
These values for g.sub.1 and g.sub.2 are typically suitable for speech signals. Other values may be used depending upon the characteristics of the signal, providing the ratios of .delta..sub.1 :g.sub.1 and .delta..sub.2 :g.sub.2 are maintained at about 1:2.sqroot.2 and unity, respectively.
In many applications satisfactory results will be obtained using the embodiment shown in FIG. 2. However, in some cases, to maximize the signal to noise ratio for certain input signals, an additional predictor, not using feedback, may be added. Such an embodiment is illustrated in FIG. 3, in which parts corresponding to those shown in FIG. 2, have the same reference numeral.
The transmitter 10B differs from that in FIG. 2 by the addition of a zero-based predictor 40 (having no feedback) which produces from the regenerated difference signal E.sub.j a partial predicted value x.sub.jo. A second adder 42 sums the outputs of the zero-based predictor 40 and the pole-based predictor 18 to produce the predicted value X.sub.j. It should be noted that the coefficients of predictor 40, adaptive as signified by arrow 44, are adaptive only in dependence upon difference signal E.sub.j as signified by broken line 46 and it has no feedback loop. A corresponding predictor 140 and adder 142 are provided in the receiver 12B.
It is envisaged that where predictor 18 has two coefficients, the zero-based predictor 40 might have six coefficients. The zero-based predictor coefficients may be derived in accordance with the teachings of U.S. Pat. No. 4,317,208.
Where an additional predictor 40 is provided, the coefficients of the pole-based predictor 18 may be adapted in dependence upon the output of the additional predictor 40, as well as the difference signal E.sub.j.
Such an arrangement is illustrated in FIG. 4, in which the transmitter 10C differs from FIG. 3 by the inclusion of a third adder 50 which sums the difference signal E.sub.j and the partial predicted value X.sub.jo at the output of the zero-based predictor 40 to provide an adaptation signal which controls adaptation of the coefficients of the pole-based predictor 18. The receiver 12C has a corresponding third adder 150 connected in like manner.
In both the transmitter and receiver, the equations for deriving the coefficients will be the same as for other embodiments, except that the term E.sub.j is replaced by E.sub.j, defined as the sum of the difference signal E.sub.j and the partial predicted value X.sub.jo.