Apparatus and Methods for Analogue-to-Digital Conversion

The present invention relates to apparatus and methods for analogue-to-digital (A-to-D) and digital-to-analogue (D-to-A) conversion.

The quantising accuracy of A-to-D and D-to-A converters is, in general, limited by the accuracy to which electronic components can be manufactured or the speed at which electronic switches can operate. The term converter is used in this specification when a reference is made to either type of converter.

A speed times accuracy product can be found which applies to each method of conversion and provides a performance indicator. This indicator is of most importance in monolithic circuit versions of converters. For example, one method such as the `flash` technique of analogue-to-digital conversion can produce quantised output in a very short time and is probably the fastest conversion method. The quantising accuracy is at the moment, however, limited to at most 8 binary digits (bits) because a potential divider is required to generate reference voltages for the (2.sup.8 -1) comparators required, and with present integrated circuit technology the potential divider cannot be manufactured easily with an accuracy or stability greater than that needed for 8 bit conversion. Techniques such as laser trimming or `zener zapping` can be used to trim the potential divider to greater accuracy, and this could extend the quantising accuracy to 12 bits, though at increased production cost.

The limiting factor with integrated circuit `flash` converters is the number of comparators which can be accommodated in a highspeed integrated circuit within an economic die size, and this limits the quantising accuracy to about 10 bits at present. The maximum sampling frequency of a `flash` 10 bit converter is in the region of 20 MHz, whereas the sampling frequency of a 1 bit converter can be as much as 500 MHz. Converters with a quantising accuracy of 16 bits are at present limited to sampling frequencies in the region of 50 KHz.

This higher accuracy can be obtained by using inherently slow-speed conversion techniques (e.g. the dual-slope ramp-converter) which do not depend on high accuracy resistor networks.

The practical constraints of integrated circuit technology therefore enable low quantitising distortion (up to 16 bits) to be achieved but at a low sampling frequency. Alternatively, a very high sampling frequency can be obtained if high quantising distortion (1 bit) is acceptable.

Oversampling is a well known technique for trading speed for accuracy, in which sampling is carried out at a frequency which is higher than the Nyquist frequency.

The r.m.s. sinewave signal to r.m.s quantising noise ratio in a bandwidth f.sub.B produced by an n bit A-to-D converter at a sampling rate of 2f.sub.B is equal to (6n+1.8) dB. If the sampling frequency is greater than the Nyquist (2f.sub.B) frequency, then the noise within the baseband up to f.sub.B is reduced. Every time the sampling frequency is doubled, the quantising noise power in the baseband up to f.sub.B is reduced by 3 dB. It can be seen that a fourfold increase in sampling frequency is necessary to reduce the baseband noise by 6 dB: equivalent to that given by an analogue-to-digital converter of 1 bit greater accuracy, sampled at the Nyquist rate.

Another problem which occurs in A-to-D and D-to-A converters is that of providing equally spaced quantising levels. Due to component tolerances such converters usually have some degree of non-uniformity in this respect.

According to a first aspect of the present invention there is provided apparatus for digital-to-analogue or analogue-to-digital conversion comprising

holding means for holding values of a signal to be converted one at a time, successive values being substituted at a first frequency,

combining means for combining the output signal of the holding means, with an interpolation signal which is periodic at the first frequency or a multiple thereof,

a digital-to-analogue or analogue-to-digital converter, connected to the addition means, and arranged to operate at a second frequency which is p times the first frequency where p is greater than one, the converter having a quantising interval d.sub.1 which is no more than ##EQU1## times the magnitude of the change effected by the combining means on the output signal of the holding means in one period of the interpolation signal, and

averaging means for averaging the output of the converter.

The combining means may comprise addition means for adding the interpolation signal to the output signal of the holding means, the interpolation signal having a maximum excursion which is at least ##EQU2## times the interval d.sub.1. Since subtraction is the addition of a negative number, the word "addition" as used in the claims of this specification includes subtraction. Alternatively the combining means may comprise means for multiplying or dividing the output signal of the holding means by a multiplier or divisor which varies periodically at the first frequency or a multiple thereof to provide the required change in magnitude.

An important advantage of the present invention as applied to binary A-to-D converters is that where the converter used has an accuracy when sampling at the second frequency of, for example, n bits, the accuracy of the output obtained from the apparatus according to the invention is (m+n) bits where p equals 2.sup.m. In D-to-A conversion if the converter has an accuracy of (n-m) bits when operating at the second frequency there are more steps in its output signal per unit time than would be available by using an n bit converter at a necessarily slower rate. A "smoother" output analogue signal can be obtained by averaging a converter output if the converter output has more output steps per unit time.

For optimum increase in signal to quantising noise ratio the said maximum excursion should be less than the quantising interval of the converter. However the above-mentioned problem of non-uniformity in quantising levels can be lessened by employing a maximum excursion which spans several quantising intervals. Such improvement is at the expense of the speed times accuracy product since as the signal at the first frequency has a greater excursion, more time is required to cover this excursion and convert the signal so provided to a digital or an analogue output.

In some circumstances p need not equal a power of two since where quantisation in the converter is not binary other integer (or even non-integer) values of p may be found useful. Further, at the expense of the speed times accuracy product, the signal which is added by the addition means may be periodic at a multiple of the first frequency.

When the apparatus is for binary A-to-D conversion, the holding means may comprise a sample-and-hold circuit operating at the first frequency. The combining means may add a sawtooth waveform having a repetition frequency equal to the first frequency and the combining means then includes a further sample-and-hold circuit operating at the second frequency for generating samples of the signals formed by addition for application to the converter. Alternatively the combining means may add a stepped waveform, having a repetition frequency equal to the first frequency and steps of duration equal to the period of the second frequency. Each step may have a magnitude equal to the said maximum excursion of the periodic signal divided by the number of step transitions. The averaging means may comprise an accumulator circuit arranged to be reset at the first frequency.

In D-to-A conversion the said quantising interval is the difference between the magnitudes of adjacent levels in the analogue output before significant smoothing takes place.

Where the apparatus is a D-to-A converter, the holding means and the addition means together may comprise an up/down counter which is arranged both to be reset at the first frequency to values of the input for conversion and to be incremented or decremented at the second frequency in each period of the first frequency.

The D-to-A converter may be capable of converting binary signals having q bits where q is less than the number of bits in the output of the up/down counter, only the q most significant bits being passed from the counter to the converter. The averaging means may be a low pass filter having a cut-off frequency at the top of the desired bandwidth for output signals.

According to a second aspect of the present invention there is provided a method of digital-to-analogue or analogue-to-digital conversion comprising the steps of

holding values of a signal to be converted, one at a time and substituting successive values of the signal at a first frequency,

combining an interpolation signal which is periodic at the first frequency or a multiple thereof with each signal value held, to generate a resultant signal,

carrying out a digital-to-analogue or analogue-to-digital conversion on each resultant signal at a second frequency which is p times the first frequency, the conversion being carried out using a quantising interval d.sub.1 which is no more than ##EQU3## times the magnitude of the change effected in one period of the interpolation signal by combining the interpolation signal with one of the signal values held, and

averaging the output from the converter.

Certain embodiments of the invention will now be described by way of example with reference to the accompanying drawings in which:

FIG. 1 is a block diagram of A-to-D conversion apparatus according to the invention,

FIG. 2 shows an interpolation waveform for use in A-to-D conversion,

FIG. 3 is a diagram illustrating the way in which averaging takes place in the circuit of FIG. 1,

FIG. 4 is a block diagram of a D-to-A conversion apparatus according to the invention,

FIG. 5 is a diagram illustrating the effect of incrementing the up/down converter of FIG. 4, and

FIG. 6 is a block diagram of a circuit which may be used as an alternative to part of FIG. 1.

FIG. 7 is a block diagram of a processor which may be used as an alternative to part of FIG. 1,

FIG. 8 is a block diagram of an alternative processor to that of FIG. 7, and

FIGS. 9 to 12 illustrate interpolation waveforms.

FIGS. 13 and 15 are block diagrams of modified arrangements, and

FIG. 14 shows interpolation waveforms using the arrangements of FIGS. 13 and 15.

In the A to D converter apparatus of FIG. 1 an input waveform to be digitized is applied to a terminal 10 and thence to a sample-and-hold circuit 11 sampling at a frequency 2f.sub.B where f.sub.B is the input bandwidth of the incoming signal. By means of an addition circuit 12 a signal from an interpolation waveform generator 13 is added to the output of the sample-and-hold circuit 11.

One form of interpolation waveform is shown in FIG. 2. It has a period of 1/2f.sub.B and decreases in steps d.sub.2 at the ends of periods 1/f.sub.S. The frequency f.sub.S is the frequency of conversion in an A-to-D converter 15, and flyback to the initial magnitude of the interpolation waveform occurs at the end of each period 1(2f.sub.B) The frequency f.sub.S =2.sup.m.2f.sub.B and therefore there are 2.sup.m -1 step transitions in the interpolation waveform excluding flyback.

For binary conversion, the maximum excursion of the interpolation waveform should, at least, equal a magnitude d.sub.1.(2.sup.m -1)/2.sup.m where d.sub.1 is the quantising interval of the converter 15. In the example shown the maximum excursion is equal to d.sub.1, and d.sub.2 =d.sub.1 /3 since there are three step transitions in each cycle of the interpolation waveform (again excluding flyback).

The output of the A-to-D converter 15 contains 2.sup.m samples for each sample at the output of the circuit 11 and this increased number of samples is averaged by means of a resettable accumulator 16 which is reset at intervals of 1/2f.sub.B. As a result the output of the accumulator 16 contains m+n bits. The accumulator adds together the values presented at its two inputs to produce a running sum. Each new input is added to the previous sum provided by way of the feedback path shown.

In order to give an (m+n) bit pcm parallel output while the accumulator 16 is operating, it is connected to a parallel output channel 17 by way of an output latch circuit 18 sampled at 2f.sub.B. The latch circuit holds the result produced by the accumulator 16 over a period 1/2f.sub.B when the accumulator is reset and a new value is derived.

Another interpolation waveform which may be used is a sawtooth waveform having a maximum excursion of d.sub.1 and an average d.c. value of zero. A further sample-and-hold circuit (not shown), sampling at 2.sup.m.2f.sub.B, is then connected between the addition circuit 12 and the converter 15. In view of the use of the further sample-and-hold circuit the maximum excursion d.sub.1 cannot be reduced to ##EQU4## as for the stepped interpolation waveform. Either of the two previously mentioned waveforms may be used when modified to have a d.c. value equal to .+-.d.sub.1 /2, and either may increase progressively with time, rather than decreasing.

A triangle waveform is another suitable interpolation waveform.

An input low-pass anti-aliasing filter (not shown) of bandwidth f.sub.B is connected at the input terminal 10 since a filter of this type is a requirement of all sampled systems.

An example of the operation of the apparatus of FIG. 1 is now given. In FIG. 3 one output level from the sample-and-hold circuit 11 over the period 1/2f.sub.B is shown by a chain dotted line 20. The vertical axis is divided according to decimal code corresponding to magnitude of input level to the converter 15. Only a portion of the axis is shown over the interval 99 to 103. In a conventional A-to-D converter the output code corresponding to a given magnitude is given for any input amplitude within .+-.1/2 a quantising interval of that magnitude; for example any magnitude in the interval 100.5 to 101.5, that is within the shaded area of FIG. 3, receives the output code 101. Thus the level 20 corresponds to an output level 101 from the converter 15 but in fact the level could be said to be 100.67 (say). When the interpolation waveform is added the resultant is as shown by the continuous line 21 and the samples obtained at the rate f.sub.S are 101, 101, 101, 100 and when these samples are averaged by the accumulator 16 the result 100.75 is provided and an increase in accuracy of 2 bits is obtained. In this example since there are (4-1) steps in the quantising signal 2.sup.m .times.4 and m.times.2.

As has already been mentioned at present 16 bit converters operating at 50 KHz and 10 bit converters operating at 20 MHz are, for example, obtainable. If 2f.sub.B is 50 KHz then a 16 bit output can be obtained without using the invention or oversampling. using the first of these converters. By oversampling at 200 KHz (not using the invention) a 1 bit increase in accuracy to 17 bits could be obtained if this were possible. On the other hand by using the invention with a 10 bit converter which samples at 20 MHz, m is between 8 and 9 and therefore the accuracy is n+m=18 or 19 bits.

An alternative way of combining the output of the sample-and-hold circuit 11 with the output of the generator 13 is to replace the addition circuit 12 of FIG. 1 with the circuit of FIG. 6 which employs a multiplier circuit 40, a divider circuit 41 and an addition circuit 42. The output V.sub.w of the sample-and-hold circuit 11 and the output V.sub.i of the interpolation waveform generator 13 are applied to terminals 43 and 44, respectively. The divider circuit 41 provides an output V.sub.i /V.sub.w to which one is added by the circuit 42 which has unity input at a terminal 45. On multiplication by V.sub.w in the multiplier 40, the output (V.sub.i /V.sub.w +1) of the circuit 42 is converted to V.sub.i +V.sub.w and appears at a terminal 46 for application to the A-D converter 15.

The divider circuit 41 may be formed, according to a known technique, by a multiplier using a negative feedback loop.

The D-to-A converter of FIG. 4 is now described. An up/down counter 25 receives an n bit pcm signal at a sample rate 2f.sub.B by way of a parallel input channel 26. The up/down counter is set at intervals 1/2f.sub.B to the current value of the n bit pcm and it is counted up by one bit at a frequency of 2.sup.m.2f.sub.B so that there are 2.sup.m times more samples in a given time at the output of the counter 25 than at its input.

These samples are passed to an OR function circuit 27 at the rate 2.sup.m.2f.sub.B but only the n-m most significant bits are passed to the OR function circuit. The circuit 27 ensures that the reduced number of bits does not result in all zeros being passed on to a D-to-A converter 28 when as a result of counting up, the counter 25 resets with all zeros. For this purpose the "carry" output of the counter 25 is connected to an input of the OR function circuit 27. The circuit 27 comprises (n-m) OR gates each receiving one input from the counter 25 and another input from the carry output from the counter 25.

The D-to-A converter 28 operates at the frequency 2.sup.m.2f.sub.B and therefore has an output level rate 2.sup.m times greater than would be obtained by applying the input at the channel 26 to a D-to-A converter operating at 2f.sub.B. The output from the converter 28 is applied to a transversal filter 30 which carries out signal averaging which is equivalent to that of the accumulator of FIG. 1. The transversal filter may have 2.sup.m -1 sections, each imparting a delay of ##EQU5## in series with the circuit input signal, and the outputs of each section taken to a common summing circuit whose output is sampled at 2f.sub.B. The tap coefficient for each summing-circuit input is unity. The output of such a filter, when used in the circuit of FIG. 4, is the same as that produced by an ideal n-bit D-to-A converter. A smoothing filter 31 is connected at the output of the filter 30 to complete the process of removing quantisation noise contained in the band above the frequency f.sub.B. Alternatively, a single low-pass filter with a high degree of attenuation above the frequency f.sub.B can be used to smooth the output of the D-to-A converter 28.

An example of the operation of FIG. 4 will now be given. Suppose three samples of a 5-bit input signal are applied to channel 26 and the three samples have values 10000, 10110 and 11011 as shown in FIG. 5 by the dashed line 33. The states taken up by the counter 25 are then as shown by the solid line 34 so that after truncation to provide an (n-m) input for the converter 28 the following samples are provided in the interval in which the three original samples occurred:

100, 100, 100, 100, 101, 101, 110,

110, 110, 111, 111 and 111.

These signals are applied to the converter 28 and the resulting output signal has more changes per unit time than the original signal. A relatively smooth analogue outout is therefore obtained from the filter 31. The outputs of the converter 28 are summed in groups of four (2.sup.m) by the transversal filter 30 to give 10000, 10110 and 11011 for the samples mentioned above, and these are the values which a 5-bit converter would have provided. In this example, the binary addition performed by the transversal filter 30 produces five-bit sums, thus providing two additional bits of resolution. The accuracy of the resulting five-bit sums is equal to the accuracy of the digital-to-analog converter 28, which in this example has three bits of resolution, but must have five-bit accuracy if the five bit sums produced by transversal filter 30 and the analog output produced by smoothing filter 31 are to have corresponding accuracy.

To contrast the invention with the prior art, if a 16 bit D-to-A converter is available, then without using the invention one of 2.sup.16 output levels could be obtained at a rate of 50 KHz for smoothing. However, if a 10 bit converter which operates at 20 MHz is used according to the invention, n-m=10 and m is between 8 and 9. Thus 2.sup.10 possible output levels at a rate of 20 MHz are provided and after smoothing between 2.sup.18 and 2.sup.19 possible output levels are obtained at a rate of 50 KHz.

The up/down counter 25 may be connected to count down during each period of 1/2f.sub.B, but a d.c. offset then occurs in the output signal. This is unimportant for conversion of audio signals and can be removed, if required, by the use of a coupling capacitor.

Two features of the A to D Converter (ADC) output used as part of the proposed interpolative encoding system enable a logical processing stage to replace the digital accumulator 16 to obtain higher resolution PCM output words.

These features are the monotonic and bivalued nature of the output stream.

Consider an n bit ADC oversampled at M times

N=2.sup.n

M=2.sup.m

for every input sampled, the ADC produces an output stream X ##EQU6##

Because of the type of interpolation waveform used (: linear staircase of amplitude less than 1 l.s.b.) there is provided ##EQU7##

Therefore in general, the ADC output stream during each sampling period consists of two values: a set of l values, k, followed by a set of M-l values, k+1, ##EQU8##

The function of the accumulator 16 is to add this stream ##EQU9## so computing the sum is equivalent to multiplying the first value in the stream by M. This can be achieved by shifting the binary value of k by m bits to the left. The binary k constitutes the n most significant bits of the output.

The remainder M-l consists the m least significant bits. This can be computed by latching the first input of any stream and comparing it with subsequent values to obtain l. M-l can be computed by decrementing a counter presetted to M.

A circuit for the implementation of this idea is illustrated in FIG. 7. An alternative arrangement which can be implemented using a smaller number of available integrated circuits is depicted in FIG. 8. The arrangement of FIG. 7 requires less gates and would therefore be of interest when the arrangement is manufactured as a single integrated circuit, as it would require less silicon area than the arrangement of FIG. 8.

This system avoids the use of adders. An implicit multiplication is performed (4), and can be implemented at higher speeds than the accumulator 16. This is because features of the ADC output stream are utilized to enable simpler, and hence faster, logic circuitry to be employed.

The limitation of this approach is in that it assumes the ADC is being used with an interpolation waveform amplitude of less than 1 l.s.b., i.e., the threshold positions of the ADC are near-ideal. If an ADC with significant threshold errors is used, then it would be normal to use an interpolation waveform amplitude of more than 1 l.s.b. The system described here could be extended using more complex logic circuitry to cope with this situation, but at the expense of speed of operation. The optimum choice between this approach and the use of a resettable accumulator 16 is one involving engineering judgement.

The circuit illustrated in FIG. 7 implements the summation operation of equation (4) using latches, logic gates and one presetable counter. The output of the ADC is latched in the n bit latch 48 at the sampling frequency 2f.sub.B. This value is then compared with successive values of ADC which are produced at the oversampling frequency f.sub.S. This comparison is performed using n exclusive or gates and inverters 50. The output of these gates and the oversampling clock are anded together to give the counting signal driving the presetable down counter 52. The counter is presetted to M-1 by wiring all its m preset inputs to 1. The comparison circuit will then produce l-1 counts for the first l ADC outputs that are the same. Using these counts to decrement the counter a value of

is obtained at the output of the counter. This is the second term in the Equation 4. The first term is obtained by the output of the n bit latch 48. These two partial results are latched together in an n+m bit latch 54 and can be clocked at 2f.sub.B to produce the output stream. The output of the down counter 52 is ready when l-1 down counts are made; that is the ADC output has changed. Therefore in general the output is ready before the M oversampled ADC outputs have arrived and been compared. A data-ready signal can be provided to indicate the end of the accumulation process.

Referring to FIG. 8, 56 is an m+n bit presetable down counter. The most significant n bits are fed from the oversampling ADC output. The remaining m bits are tied to logic `1`.

Block 58 is an m+n bit latch which provides stable parallel output signals during 2f.sub.B periods.

Consider the case when n-2 and m-2. In this situation a four times sampling frequency will result in the gain of 2 bits.

At the start of the 2f.sub.B cycle the first sample will be preset in the counter. Due to the synchronisation of the 2f.sub.B and 2.sup.m.2f.sub.B clocks, the counter will not be decremented at this time. At time 1/2.sup.m.2f.sub.B later, the next sample will be available. This is compared with the first sample by the ex.or gate. Since there are used only up to 1 l.s.b. of interpolation, any change in the samples in any 1/2f.sub.B period will be limited to 1 l.s.b. and hence can be observed in a change in the l.s.b. of the new sample. If this value is the same as the first one, then the counter will be decremented.

During subsequent samples in the 1/2f.sub.B period the counter is decremented at f=2.sup.m.2f.sub.B until an input transition is recorded. To complete the 1/2f.sub.B period, the final value reached is latched at the output.

Consider the code sequence 1,1,1,2. over one cycle at intervals t.sub.0,t.sub.1,t.sub.2,t.sub.3, ______________________________________ COUNTER ADC VALUE OUTPUT TIME (dec.) (binary) LATCH OUTPUT ______________________________________ t.sub.0 1 0111 t.sub.1 1 0110 Previous x.sup.- t.sub.2 1 0101 value t.sub.3 2 ##STR1## t.sub.4 .sup. x.sup.+ X.sup.+ 11 0101 ______________________________________

OR consider the sequence 2,3,3,3 ______________________________________ t.sub.0 2 1011 t.sub.1 3 1011 Previous t.sub.2 3 1011 value t.sub.3 3 1011 t.sub.4 .sup. x.sup.+ x.sup.+ 11 1011 ______________________________________

The analogue transversal filter 30 should have a sample and hold at its output, ie between the transversal filter 30 and the smoothing filter 31, operating at a rate of 2f.sub.B for its action to be a direct equivalent of the resettable accumulator used in the A to D converter. Resampling is essential in order to obtain a full one bit per octave increase in resolution with wide band input signals. If the signal bandwidth is very much less than f.sub.B, then neither the transversal filter 30 nor resampling would be necessary and a very simple low-pass filter is all that would be required. Other sampled data analogue filters could be implemented which would provide the same result as the transversal filter plus sample and hold. The optimum arrangement would depend on the engineering trade-offs which proved most suitable for a given application.

One feature of this conversion technique, is that a small loss of input dynamic range occurs when a 2.sup.n-m level converter is employed in the system. Ideally, an n bit converter should provide outputs which correspond to 2.sup.n different levels; with this system, some of these levels are lost because the 2.sup.n-m level dynamic range must be used partly for the interpolation signal and partly for the main signal. If one l.s.b. of interpolation is used, then the number of levels which the main signal can excurse is (2.sup.n-m -1). This loss of range may be unimportant when, for example, n-m=16. In an integrated circuit implementation of this system, it would be advantageous to use a (2.sup.m-n +r) level converter, where r is small and equals the number of l.s.b's of interpolation used. This is preferable to using a 2.sup.m-n+1 level converter as (a) much less silicon area is needed, and (b) faster effective operation of the converter becomes possible by making use of the much reduced one l.s.b. conversion time as explained under increased performance using improved l.s.b. conversion times. It is instrumentally easier to achieve fast one l.s.b. conversion times in an r level converter where the r levels change one at a time than in a conventional binary converter (particularly for D to A conversion).

For example, a (2.sup.m-n +1) level flash A to D could be producwed by adding only one extra comparator and a 2.sup.m-n +1 level binary weighted current D to A could be produced by adding an extra one l.s.b. current source. The r level part of the converter would be controlled separately from the 2.sup.m-n level part. In the D to A application this separate control would reduce the amplitude of switching transients.

A conventional binary D to A is not optimum for obtaining minimum one l.s.b. conversion time as the worst case figure must be used and due to the m.s.b. switching transient, this occurs at the one l.s.b. code change of 01111 . . . to 1000 . . . , (a full scale conversion time would be measured with a code change of 0000 . . . to 1111 . . . ). With conventional binary D/A converters major bit transistors such as 01111 to 10000 can occur with an effective one l.s.b. change in signal level. The switching transients which correspond to these changes are at amplitudes near the m.s.b. level rather than the l.s.b. level. The extra r levels change only one l.s.b. at a time and any switching transients introduced are always smaller than those due to m.s.b. transitions. Additional control logic will be required in order to select the r level portion of the D/A converter. Many arrangements are possible, depending on the value of r, but in its simplest form, the carry output of the counter 25 can be used to select the extra l.s.b. when r-1.

Simplification of the counter circuit is also possible when a 2.sup.m-n +1 level D to A is used: instead of an n-bit counter, an n-m bit latch plus an m-bit counter can be used. The n-m most significant bits of the input word would go straight to the (n-m) binary input portion of the D/A. The m least significant bits of the n-bit input words would preset an m-bit up-counter, and the carry output of this counter would control the l.s.b. portion of the D/A. This approach can be extended to cover the r level case and offers a saving in logic circuitry and the possibility of faster circuit operation.

An interpolative scheme is described which enables increased resolution to be obtained by utilizing the potentially fast settling time of one least significant bit (LSB) change in a digital to analogue converter (DAC). Results of experimental measurements indicate close to theoretical improvement is achievable in practice. The use of the arrangement in an integrated circuit design is likely to offer significant advantages in terms of lower production and testing costs than a conventional DAC of equivalent performance.

There is reported preliminary results of work aimed at improving the performance of an interpolative DAC. Interpolation achieves increased resolution from a DAC by alternately switching its output between adjacent levels. Time averaging of the DAC output using appropriate filtering can enable additional levels to be resolved between the DAC output levels and this has been likened to a process of linear interpolation.sup.1. Previous work reported in this area has been aimed mainly at the development of integrated circuit codes for speech transmission using companded pulse-code modulation.sup.2. Maximum conversion speed and scope of this approach was limited by a need for relatively complex digital circuitry and a special purpose DAC which converted on a level-at-a-time basis. In contrast with this earlier work, this letter describes an arrangement which offers the promise of much higher conversion speeds and a wider field of application because very simple digital circuitry is used and the arrangement works with conventional DAC designs.

The definition of figure of merit of an analogue to digital converter as used in the specification is the resolution in bits multiplied by conversion frequency. For a digital to analogue converter the figure of merit is the product of its resolution in bits and the increase of its full-scale settling time (bits.times.Hertz). Linear interpolation can be thought of as trading a reduction in effective conversion speed (Hertz) for an increase in effective resolution (bits), with no alteration in the figure of merit. The theoretical rate of exchange being therefore a one bit increase for every octave of speed reduction.

Whilst this exchange rate can be an attractive proposition for many applications, there are of course some in which the consequent increase in full-scale conversion time would be unacceptable. An arrangement is proposed which should enable a very much better rate of exchange to be obtained, resulting in an increase of the figure of merit.

Experiment: FIG. 13 shows the experimental arrangement. The digital circuitry required consists simply of a presettable up/down counter 60. In the experiment a commercially available 16 bit DAC 62 was used to examine the practical limits of the technique. The test signal source was a low distortion digital sinewave generator 64 which produced data words at a rate of 50,000 per second.

Experimental measurements were first made using only the eight most significant bits of the DAC in order to minimize the influence of DAC linearity errors and switching transients. This was followed by measurements using sinewave data which generated an output waveform of peak amplitude -40 dB relative to maximum available from the DAC, and centred on the mid scale point of the DAC. In this last test, all 16 data inputs of the DAC were used. Signal to quantising noise ratios were measured using a total harmonic distortion analyser.

Results: A computer program was written which used signal-flow graph theory.sup.3 to predict the likely improvement in baseband signal to quantising noise ratio and the amplitude-frequency distribution of the quantising noise. This analysis predicted a baseband improvement of 4.26 dB with interpolation waveform 1 shown in FIG. 14(a), 5.80 dB with waveform 2 in FIG. 14(b) and 5.99 dB with waveform 3 in FIG. 14(c) after the waveforms were low-pass filtered. Each of these waveforms has an average value of one half an LSB and in theory the use of an integrate and dump filter at the DAC output would enable the ideal figure of 6 dB improvement to be obtained with any of the waveforms. Practical measurements using the eight most significant bits of the DAC and a 20 kHz low-pass filter with a stop-band attenuation of better than 96 dB, gave signal to quantising noise ratio improvements of (a) 4.2 dB, (b) 5.8 dB, (c) 5.85 dB for waveforms 1, 2 and 3 respectively, showing very close agreement with the predicted values.

In the -40 dB test, signal to quantising noise ratios significantly worse than theoretical were obtained and investigations revealed that this was due primarily to switching transients (glitches) occuring when data inputs changed. The output data rate of the counter can be up to eight times more than its input rate and therefore the repetition frequency of the DAC transients can be up to 400 kHz. The importance of using a track-hold to remove such glitches is well recognised, particularly with DACs intended for digital audio.

In higher speed applications, this requirement for a fast track-hold may be undesirable as the design of such a device can present sigificant problems.

Figure of merit improvement: The minimum time between the n bit input words (conversion time) in the interpolative arrangement of FIG. 13 is simply the full scale settling time plus the time required for the sequence of LSB changes needed to produce the interpolation waveform. With most conventional DACs, the worst case LSB settling time is comparable with the full scale settling time owing to glitches which occur at input code changes. In practice, therefore, a figure of merit is calculated using a conversion time based on the full scale settling figure. The 2.sup.m +1 level DAC arrangement of FIG. 15, together with a simplified up/down counter system is suggested as a way of increasing the figure of merit available by using interpolation. In the proposed system, the m most significiant bits of the n bit word are presented to the m bit conventional section of the DAC for the duration of the conversion period. The n-m least significant bits are loaded into an n-m stage presettable up/down counter and with the use of appropriate control signals the carry output of the counter provides the waveform to control the interpolation stage of the DAC. Glitches are now no worse a problem than with a conventional DAC as the interpolation stage is inherently glitch free. The settling time for the interpolation stage should therefore by very much less than the LSB settling time for the conventional section of the DAC. This should reduce the total conversion time required significantly, with a corresponding increase in the figure of merit. Ideally the interpolation stage should be matched to the LSB stage of the conventional part of the DAC and therefore the proposed arrangement is likely to be of most use in an integrated circuit. The reduced complexity of the counter should facilitate faster operation at the integrated circuit level and reduce the area of silicon required.

Very good agreement has been obtained between theoretical predictions of signal to quantising noise ratio produced by an interpolative decoder and experimental measurements. Code-dependent switching transients have been found to be a major limitation to the application of interpolative techniques to high speed digital to analogue conversion. In the light of this work a modified LSB switching arrangement has been proposed which would eliminate the switching transient problem at the LSB level. Owing to the much reduced LSB settling time, it would enable far better resolution to be obtained for a given full scale settling figure than previously reported interpolative arrangements. It is likely that the area of silicon required for a proposed interpolative DAC will be much less than for a conventional DAC with an equivalent resolution and full-scale settling time. In the limit, it is in principle possible to use an extremely fast low-resolution DAC to suit a wide range of lower-speed higher-resolution applications. An added advantage is that high-resolution conventional DACs ideally require time consuming level-at-a-time tests in order to fully characterise their performance. Because of this, full testing of high-resolution DACs in a production environment can be uneconomic. As the analogue section of an interpolative DAC is ideally of low resolution, full testing of this can be accomplished in a relatively short time and as the time required to test the remaining digital circuitry is insignificant it should prove economic to undertake full testing of interpolative DAC integrated circuits on a routine basis.

Most practical D to A and A to D converters are able to convert a signal change of one l.s.b. amplitude in less time than a signal change of full scale amplitude. All previous estimates of the improvements available using the method described in the patent specification have been based on only full scale conversion times as this is the conventional measure of conversion speed.

If the full scale conversion time is T.sub.L and the time available for conversion is T.sub.C (=1/2f.sub.s) then the improvement in resolution would be log 2T.sub.C /T.sub.L bits.

By making use of the fact that the one l.s.b. conversion time is always less than the full scale conversion time, a further improvement in resolution can be obtained. In order to calculate this improvement, assume T.sub.C =2.sup.m.T.sub.L, T.sub.L =2.sup.K.T.sub.S, and T.sub.C =2.sup.m+K.T.sub.S, where T.sub.S is the one l.s.b. conversion time.

The first conversion must be able to cope with up to a full scale change and therefore the time remaining for l.s.b. conversions is T.sub.C -T.sub.L. An improvement of q bits more than can be obtained with just using T.sub.L is shown by the following equations. ##EQU10##

One conclusion which can be deduced from this equation is that if m is large, then it is possible to increase the resolution by up to K bits, giving a total improvement of almost m+K bits. In a given application it may be beneficial to increase the linearity of the converter at the expense of some of the m+K bits improvement in resolution.

The following information has been obtained on the amplitude of the interpolation wave form from the generator 13 in FIG. 1. This information has been obtained as a result of simulations with a 5 bit ideal quantizer. Neglecting the effect of over load:

iw* Ideal interpolation waveform amplitude/step size

iw.sub.(3 db) Waveform range that gives .+-.3 dB SNR of the maximum.

iw.sub.(3 dB/Oct) Waveform range that gives better than 3 dB/Octave improvement.

Approximate amplitude as a function of m=oversampling factor. ##EQU11##

These results are illustrated in FIGS. 9 to 12

Now that embodiments of the invention have been described, it will be clear that the invention can be put into practice in many other ways. For a further increase in accuracy or better smoothing the present invention may be combined with the previously mentioned known apparatus and methods of oversampling.

An application of the improvement in resolution can be obtained by making use of a system which combines a fast one l.s.b. conversion with a conventional full scale conversion. The improvement available can be seen by way of the following example:

Using the prior art of oversampling, an 8 bit A to D converter with a full scale conversion time of 100 ns can be used to provide a 12 bit resolution conversion by averaging 256 successive full scale conversions, giving a total conversion time of 25,600 ns.

With the technique described above with reference to FIG. 1, 12 bit resolution could be obtained from the same A to D by using 32 successive full scale conversions, giving a total conversion time of 3,200 ns. In an A to D converter system which provides for both conversion of full scale signals and successive signals which are different in amplitude by only one l.s.b., two types of A to D converters are usually required. In this case, a one l.s.b. conversion may typically be completed in one tenth of the time for a full scale conversion. By making use of this faster one l.s.b. conversion in the present example, 12 bit resolution could be obtained by making one full scale 8 bit conversion (in 100 ns) followed by 31 successive one l.s.b. conversions (each taking 10 ns). The total conversion time is now reduced to 410 ns demonstrating a significant improvement in conversion speed over both prior art and the previous specification. If, instead of reducing the conversion time, the aim were to obtain the maximum resolution for a given conversion time (e.g., 25,600 ns), then this comparison would give the following results:

using prior art oversampling techniques, 12 bits; using the technique in the circuit described above with reference to FIG. 1, 16 bits; using the technique described in the foregoing paragraph, 19 bits.