Philips tech. Rev. 36, No. 11/12 MODULATON 337 ll. Quantization and coding of analog signals An audio signal from a microphone or a video signal from a camera will become degraded by noise or other interfering signals when it is transmitted over a particular channel, either in baseband or on a carrier. On reception it is no longer possible to separate the original signal from the interference imposed on it. This is' because the signal is an 'analog' signal: within certain limits it can assume any value. When a signal is quantized, however - that is to say given a range of discrete values - the receiver will always be able to discriminate between the signal values in spite of the added interference, provided that the interfering signals are not too large compared with the differences in the signal levels. This opens up the possibility of almost completely eliminating the effects of noise and interference. The advantage when signals have to be transmitted over great distances is particularly apparent, because as the distance increases an analog signal inevitably accumulates more and more noise and interference, so that the signal is finally lost in the noise, whereas a quantized signal can be regenerated at repeater stations. To take advantage of this technique for analog signals from a microphone or camera, the signals have to be transformed into quantized signals that carry the same information. Usually this is done in such a way as to result in a digital signal; the process is then called 'digital coding'. n particular, in 'binary coding' the result is a 'bit train', a train of pulses that can only have two values ('0' or '1 '). Digital coding is now used on a large scale and to a growing extent in telephony. Apart from the advantage of a signal quality that is virtually independent of distance, it also provides the possibility of transmitting signals in time-division multiplex (TDM), with the benefits we have already seen. The system of digital coding in combination with TDM was first used on overcrowded lines between exchanges at short and medium distances, but it is now increasingly being used for long-distance communications. Owing to the high investment costs, the public telephone network is only gradually being digitized and it will probably take twenty or thirty years or more before it goes completely digital. n military communication, where digital coding is also widely used, the advantages of secrecy and the possibility of changing the code carry particular weight. For television digital coding is still in the experimental stage. n Britain there are experimental digital-coded television links between Southampton, Portsmouth and Guildford. n this part of the article we shall 'take a closer look at a number of coding methods. Pulse-code modulation (PCM) A straightforward development of the idea outlined above is to be found in 'pulse-code modulation', a name that is not really correct since what is involved is coding and not modulation. The coding takes place in three stages in each sampling interval; seefig. 58. First of all the analog signal is sampled; the sampling rate must of course be sufficiently high. Next, the sampled value is quantized, i.e. rounded off to the nearest level in a preselected scheme of levels. f the levels are equidistant, we have linear PCM. The levels are thought of as being numbered, and the number obtained is finally coded in a pulse train. Notation of the numbers in the binary system results in binary coding. f for example we reserve 8 bits per number,' we.can operate with 256 levels. We then say that we have 'words' of 8 bits. Out of the bit train received the receiver must of course 'know' which of each 8 bits together form a word; this requires 'word synchronization'. This does not have to be provided for each word separately but only now and then, for example at regular intervals after a certain number of words. Examples of sampling and bit rates in use are given in Table J. The sequence of level values obtained after decoding in the receiver is an approximation to the sequence of sampling values. After having passed thtough a lowpass filter the signal is an approximation to the original signal. The error arising from the difference between the level values and the sampling values is known as 'quantization noise'. -t Q -7-6 :::::L5-4 -3-2 -1-0 Q n~~:~~1 Wr~w3~ Fig. 58. Coding of an analog signal in PCM. a) The analog signal is sampled. b) The samples..., At, A2, A3, A4,... are each rounded offto the nearest level ofa given scheme of levels (quantization). c) The sequence of level numbers obtained is converted into a sequence of 'words'..., Vt, 1V2, W3, W4.. which together form a pulse train. n the figure the scheme consists of 8levels, which can be coded in binary form into words of three bits. The sample A4, for example, yields level5 and the word 101 (W4). -f
338 F. W. DE VRJER Philipstech. Rev.36, No, 11/12 The fineness of the scheme of levels chosen is a compromise.acoarse scheme gives a great deal of quantization noise. A fine scheme requires a large wordlength, and hence a high bit rate and a broadband transmission channel. A good compromise between quantization noise and wordlength can be reached with a nonlinear level scheme. For speech signals an optimum, nonlinear level scheme has been internationally standardized. selves, we can either divide up the region much more finely with the same number of levels, and hence reduce the quantization noise, or divide it up just as finely with far fewer levels, which implies a reduction in the bit rate required. The advantage gained here is paid for with a 'rounding off' of large transitions between the samples ('slope overload'), since a large change in signal level can only be covered by several steps; see jig.60. Table J. Examples of sampling rates fe and bit rates fb used for coding speech and television signals; b bandwidth of the information signal, 11 number of bits per sample (fb = /fc). n televisionfc is often derived from the frequency se ofthe colour sub carrier. nformation System \ b signal Speech pcm 4 khz 8kHz 8 64 kbitjs PCM 5MHz 3/se = l3.30 MHz 8 106.4 Mbit/s TV (PAL,lse = PCM 5MHz 2/se = 8.867 MHz 8 70.9 Mbit/s 4.433 MHz) DPCM 5MHz 2/se = 8.867 MHz 6 53.2 Mbit/s PCM 1 MHz 2MHz 6 12 Mbit/s Video 7 14 Mbit/s telephone 8 16 Mbit/s fe DPCM 1 MHz 2MHz 4 8 Mbit/s 1/11 fb CClT standard Differential pulse-code modulation (DPCM) n audio signals, and more particularly in speech signals, the low frequencies generally predominate. This also applies to video signals: abrupt brightness transitions in the picture are the exception, a fairly even distribution of brightness is the rule. Thus, for both audio and video signals, consecutive samples often differ little in value. For this reason it may be advantageous to code not the sampling value itself but the difference between a sample and the preceding one. This is known as differential pulse-code modulation (see Table ). Fig.59 gives a block diagram of the equipment required. n the sampling interval k the sample Ak is compared with an approximation to the previous sample Ak-l in differential amplifier D. The difference is quantized and coded (Cod). The coded signal is then transmitted, but besides being decoded in the receiver it is also decoded in the transmitter (Decad) and added in the integrator to the value Ak-l that was stored in 1. n the next sampling period, thus passes on to D an approximation Ak to Ak for comparison with the new sample Ak+l. The receiver also consists of a decoder and an integrator. f we now limit the difference to a region much smaller than that covered by the sampling values them- A level scheme that covers the whole region but is much more finely divided up for the small differences than for the large ones ('nonlinear DPCM',jig. 61) may often be a useful compromise. n television applications, for example, there is then little quantization noise in the fairly even patches of brightness in the picture, while the greater amount of noise that occurs with the larger transitions in brightness is not very annoying. The DPCM system can be refined still further by keeping the level scheme flexible and adapting it at Fig.59. Block diagram for DPCM. D differential amplifier, Cod coder, Decod decoder, integrator, fe sampling frequency. n the sampling interval k the difference between the sample Ak and an approximation to the, preceding sample Ak-l is coded. The pulse train obtained is decoded in the transmitter, and added to Ak-l in ; this yields Ak for comparison with Ak-l in the new interval.
Philips tech. Rev. 36, No. 11/12 MODULATON 339 -f -- Fig. 60. n DPCM the maximum difference per step is often made considerably less than the possible difference between two samples. This permits a reduction in quantization noise. A considerable difference in consecutive samples can then only be bridged, however, in several steps, so that the transient is rounded off ('slope overload'). Solid line: original signal; dashed line: approximation after DPCM. -5-4 _,.3 /2 -:::::::='.--1-0 --~ --,J -2 --':::3 --4 Fig. 61. Nonlinear level scheme for PDCM. Because of the fine division for the small differences the quantization noise remains low during slowly changing signals, and slope overload is avoided by the presence of large differences (coarsely divided). --5 any given moment to the situation, f, for example, a few positive steps have been made one after the other, we can then extend the level scheme to cope with the expected slope overload. This can bring considerable improvement, but does of course require more complicated equipment. Delta modulation (DM) A method that is remarkable for the simplicity of the equipment is delta modulation [12] (DM, fig. 62). This is DPCM for which only one bit per sample is used; the bit rate is equal to the sampling rate. For the sample Ak in this case it is only the sign and not the magnitude of the difference from Ak-1 that is determined in the differential amplifier. The.coding is also simple: a pulse ('1') is transmitted when Ak is larger and no pulse ('0') when Ak is smaller than Ä k -1. n the first case the integrator output makes a single (fixed) step upwards, and in the second case one step downwards. The result is a stepped curve that approximates to the signal (fig. 63). The choice of step size is a cornpromise here between a considerable chance of slope overload (small steps) and a high quantization noise level (large steps). The DM system, which is especially suitable 'for speech, has several refined variations; a system with 'digitally controlled companding' [13], for example, is used in the experimental video-telephone network in the Netherlands [14]. Effect of bit errors J 1----' Fig. 62. Block diagram for delta modulation. The differential amplifier D determines only whether AT.: is greater or smaller than Ak-1, which is translated by the coder as a pulse or no pulse. The integrator converts this into a step up or a step down, as appropriate. The result is a stepped curve as shown in fig. 63. Owing to noise peaks during transmission, bit errors will occasionally occur. n ordinary PCM one bit error leads to one incorrect sample value. n DPCM, on the other hand, the errors are cumulative, so that the mean level departs from the correct value. For audio signals the mean level is ofno significance; at the receiving end it is always reduced to zero. The slow accumulation of errors in DPCM and DM therefore causes no trouble in audio reception. Furthermore, an error in a difference is far less troublesome than an error in the value itself; in PCM an error in the most significant bit can be serious. For audio signals, therefore, DPCM and DM are preferable to PCM. -f 10101 1 1 1 1 100 Fig. 63. Stepped curve produced modulation. from an analog signal in delta [12] J. F. Schouten, F. de Jager and J. A. Greefkes, Philips tech. Rev. 13, 237, 1951/52. J. A_. Greefkes and F. de Jager, Philips Res. Repts. 23, 233, 1968. [13] J. A. Greefkes and K. Riemens, Philips tech. Rev. 31, 335, 1970. [14] E. A. Aagaard, P. M. van den Avoort and F. W. de Vrijer, Philips tech. Rev. 36, 85, 1976. H. P. J. Boudewijns, E. C. Dijkmans, P. W. Millenaar, N. A. M. Verhoeckx and C. H. J. Vos, Philips tech. Rev.36, 233, 1976..
340 F. W. DE VRJER Philips tech. Rev. 36, No. 11/12 n television the situation is entirely different. While it is true that the detected signal is restored to 'black' after each line of a field, it remains a drawback that when DPCM is used an error that occurs during a line fore gained at the expense of greater susceptibility to bit errors. There are methods, however, of correcting the effect of bit errors or masking it, and DPCM is in fact often preferred to PCM for television. Fig. 64. Television-picture degradation due to bit errors in transmission. above: with DPeM (Pc = 3.3 x 10-4), below: with pem ip«= 10-2). The probabilities Pc of a bit error for the two cases have been chosen so as to give about the same degradation. DPeM is thus about 30 times more sensitive to bit errors than pem. persists up to the end of the line. With DPCM, therefore, received bit errors lead to stripes in the picture, with PCM only to small flecks (jig. 64). The advantage of a lower bit rate or less quantization noise is there- Predictive coding DPCM and OM can be regarded as simple cases of 'predictive coding'. The principle is to extrapolate, in accordance with a certain rule, from preceding samples
Philips tech. Rev. 36, No. 11/12 MODULATON 341 to an expected value for the sample at the time tic. The difference between the actual value and the predicted value is coded and transmitted. At the receiver end the preceding samples and the rule are known, so that the predicted value can be determined and, with the received difference value, the actual value found from it. n DPCM and in the form of DM described above, the rule is simply that 'the predicted value is equal to the preceding sample'. Often, however,.a better predicted value can be made, for example by means of linear or quadratic extrapolation from the two or three preceding samples. This is the method used in the more refined forms of DM and DPCM. For television it is better to extrapolate in a different way. n a television picture (fig. 65) high correlations tions that has at least the capacity of a line, a field or a complete picture, as the case may be. lt is also possible to transmit nothing at all if the actual value differs from the predicted value by less than a threshold value. Because of the high correlation between successive pictures in television, the bit rate can be substantially reduced in this way. From the resultant non-constant bit flow a strongly reduced but constant bit flow can be made with the aid of a buffer memory. This does however require additional address bits to establish the picture points to which the transmitted bits relate. This is the system used for transmitting weather charts by meteorological satellites. The method is of great advantage here, because there is very little difference between successive pictures. c B o ---- ----Ë-------- A Fig. 65. Predictive coding in television. The figure represents lines and points in a small part of a television picture; the solid and broken lines represent interlaced fields. Because of the strong correlation in a picture, the luminance at P at a given instant can be reasonably accurately be predicted, for example from the signal value of the previous picture point (A) and from the values obtained one line period previously (B), immediately before and after that (C and D), one field period previously (E), and one picture period previously (P itself). exist between a given point P in the picture and its environment (A, B, C, D, E,... ). Since a complete picture in a picture period is generated by two successive interlaced line fields (picture period = 2 X field P period), it is obvious that the predicted value for the sample that will give the luminance at P at a given moment should be determined not only by the preceding sample (luminance at A) but also by the samples taken a line period before (B), immediately before and after that (C,D), a field period before (E) and possibly a complete picture period before (P itself). The predictive rule might then be: 'P shall be just as much brighter than A as B is brighter than C'. The counterpart of the advantage obtained - a reduction in bit rate - is in the first place that a memory is required for these opera-, Other coding methods To conclude, two other coding methods currently of interest will be briefly mentioned. n 'transform coding' a 'block' of samples is transformed into an equal number of other values that are subsequently coded and transmitted, The transformation may be a Fourier transform, where the coefficients of the Fourier components are transmitted, but other transformations can also be used. Two advantages are aimed at. n the first place the errors in the transmission and errors due to quaritization noise will be more uniformly distributed among,. the original samples. n the second place, certain coefficients (e.g. those of the low-frequency or those of the high-frequency components) will often prove to be relatively insignificant, and neglecting them can lead to a reduction in the required bit rate. n television a block is often chosen in which the samples belong to a square in the picture, e.g. of 16 X 16 image points [15]. The other method of coding to be mentioned in conclusion offers advantages in applications such as facsimile transmission. This involves long series or runs of identical samples ('white' or 'black'). t is then preferable not to code each sample separately but to take the sample value and the length of a run. This method, known as 'run-length coding', can be improved by taking into account the correlation between successive lines. :..,~~ [15] H. Bacchi and H. Tchen, Ann. Télécomm. 30, 363, 1975. * * *
342 Philips tech. Rev. 36, No. 11/12 An underground line amplifier (repeater) for 60- MHz cable transmission. The cable contains twelve coaxial cores, each carrying 10800 telephone conversations in frequency-division multiplex (FDM). A thirteenth core with its amplifier repeater serves as a service channel. These line repeaters are spaced at intervals of Jj2 kilometres (about 1 mile).