Anti-alias and anti-image filtering:

The benefits of 96kHz sampling rate formats for those who

cannot hear above 20kHz.
Julian Dunn
Nanophon Limited
Cambridge, UK
Email: [email protected]
Web: http://www.nanophon.com
ABSTRACT
Reports that 96kHz sampled digital audio systems have greater
transparency than those sampling at 44.1kHz apparently conflict
with knowledge of the capability of human hearing. The bandlimiting filters required are examined for a role in producing these
differences. Possible mechanisms are presented for these filters
to produce audible artefacts and filter designs avoiding these
artefacts are illustrated.
1.0 INTRODUCTION
Anti-alias filters are used to limit the bandwidth of a signal being input to a sampled
system, such as a digital audio system, so that it is less than one half the system
sample rate. This is required to avoid ambiguity resulting from the sampling process
- which represents all signals as being within a frequency bandwidth defined by a
folding frequency of one half of the sample rate. Signals sampled at a frequency
outside this range are aliased so that they are indistinguishable from frequencies
within it.
For digital audio analogue to digital converter (ADC) sub-systems with an output
sample rate of 44.1kHz this filter would, ideally, be transparent to signals in the
region below the folding frequency of 22.05kHz and totally reject signals above that
frequency. In practice the passband response will not be transparent and the
stopband may have inadequate rejection. This may be a particular problem just
above 22.05kHz due to the finite attenuation gradient over the transition (band)
between the passband and rejection band.
In the reverse process, anti-imaging (or reconstruction) filters are used at the output
of a sampled system to eliminate spectral images of the desired signal. These are
generated at integer multiples of the sample rate. The ideal filter specification is the
same as that for the anti-aliasing filter. Any passband errors will have the same
effect for both types of filter. The consequences for the rejection band deviating
from the ideal are not the same, however, and depend on the response of the
following equipment to the images that may be generated.
Both types of filter can be implemented in the analogue or digital domains. Digital
sample rate converters also have bandlimiting filters to perform the functions of antialias and anti-image filtering. These are implemented only in the digital domain. The
issues and compromises made are similar whether the processing is digital or
analogue.

Julian Dunn 1998

First presented at 104th ASE Convention, Amsterdam, May
1998 AES preprint number 4734

In most current audio ADC and DAC sub-systems this filtering is performed by a
gentle, non-critical, analogue low-pass filter of low order in conjunction with an
oversampled converter and a high order digital brickwall filter. This allows the sharp
cut-off filtering to be performed in the digital domain using a finite impulse response
(FIR) filter of one or more stages. In a multi-stage filter only one of the stages
performs the brickwall function and relates to the issues raised in this paper.
This paper is concerned with the compromises often made in the specification of this
filter. (Many of the issues raised also apply to analogue and digital infinite impulse
response (IIR) filters.) In particular, it shall examine the effects of passband
frequency response error, or ripple, and of having the rejection band start at a
frequency above half the sample rate.
Could these effects - producing very small errors in the sub-20kHz audio band explain any of the reported differences between 44.1kHz and 96kHz sampled
systems? [1, 2, 3, 4]
If these aspects are relevant then how should the design of anti-alias and anti-image
filters for 96kHz sampled systems differ from those for 44.1kHz sample rate
systems? Could some of the benefits reported for 96kHz systems be achieved for
lower sample rate systems with different filters?

2.1

PASSBAND DEVIATIONS

2.2

Ripple

The passband response should ideally be flat (or an exact complement of other
deviations so that the total system response is flat).
The digital FIR filters used in most ADC and DAC sub-systems approximate to this
ideal in a way that minimises the maximum distortion of the real filter response from
this ideal. They are 'equiripple' filters, meaning that their responses are characterised
by the presence of ripples which are all the same size.
The REMEZ [5] filter design program works in this way. Figures 1 to 4 illustrate the
response of some example 2x oversampling filters designed using REMEZ. Notice
from graph B of each of these that the deviation of the passband (like the stopband)
has a constant ripple. This passband response can be considered as the desired flat - response with an additional error response. This error can be approximated by
a cosine wave. This is shown in graph D. For figure 1 this cosine has cycles that
are 3.3kHz apart, and an amplitude of 0.0006 (relative to the average gain of 1).
This is not to be confused with a cosine wave in the time domain; it is a transfer
function that has a cosinusoidal shape in the frequency domain. What does it
represent in the time domain?
Given a frequency response, the time response can be found by Fourier
transformation. If we approximate the error (of the figure 1 filter) to a sine wave then
the transform is simply an impulse pair of total amplitude -67dB (-70dB each)
displaced by 0.3ms (14.5 samples) relative to the main signal. (The signals are
delayed by at least this amount in the filter so causality is maintained.) This
illustrates that the periodic passband ripple indicates pre and post-echoes in the time
domain.
Page 2
Julian Dunn 1998

This result - that for equiripple filters the passband time dispersion results in discrete
time displacements - is contrary to many expectations. The spread of the impulse
response is normally quoted as being indicative of the amount of smearing. A
comparison of graphs C and E for each figure show that the pre and post echoes
occur at times corresponding to the beginning and end of the impulse response where the amplitude is smallest.
The filter of figure 1 filter is not unique. Table 1 is calculated from the characteristics
of various integrated ADCs, DACs, and digital filters. The amplitude and the pre and
post echo time displacement are calculated from the passband ripple amplitude and
periodicity estimated from data sheets and published graphs. (The filters of figures 1
and 2 are designed to have similar responses to devices 1 and 2 in this list)

1
2
3
4
5
6

Type
ADC
DAC filter
DAC
ADC
ADC
(last stage)
DAC

Ripple deviation
0.005 dB
0.000015 dB
0.035 dB
0.001 dB
0.01 dB
0.003 dB
0.0005 dB

Ripple Amplitude
-68 dB
-118 dB
-51 dB
-82 dB

Echo displacement
0.3ms
0.8ms
0.3ms
no data

-72 dB
-88 dB

0.7ms
no data

Table 1
An inspection of the columns for echo amplitude and ripple rate shows a wide
variation with amplitudes of -51dB to -118dB and timing variations of between 0.3ms
and 0.8ms. (Calculated for a sample rate of 48kHz)
The production of pre-echoes from filter ripple was reported by Lagadec and
Stockham [6],. They found the pre-echo due to a filter ripple of 0.2dB with a span
of 23Hz corresponded with echoes of -32dB at 40ms - which was found to be quite
perceptible even with untrained listeners. A question is raised of how much echo,
and in particular how much of the un-masked pre-echo, can be permitted without
producing a degradation in the highest quality reproduction system? The
psychoacoustics results, including those reported in [7] for example, do cast some
light on this. However, none of the results seem to be directly applicable. It would
seem that some kind of threshold test on the audibility of pre and post-echo pairs is
required.
Some of the work on the evaluation of 96kHz sampling systems reports an improved
resolution of the timing of impulsive test signals [3], and of an improved perception
of musical transient attacks and fast staccato passages [2].
Timing is also related to localisation cues but research into the improvement in
phantom source perception differences between 48kHz and 96kHz [4] has not
shown an improvement in localisation accuracy. The experiments may be worth
repeating with less ideal filters.

2.3

Filters with ripple-free passbands

There are methods of FIR filter design that produce filters without passband ripple,
and therefore without the discrete echoes, such as the high-fidelity filter [8] or the
constraint based FIR filter design program, METEOR [9]. These approaches are not
normally used for digital audio applications - presumably because of the requirement
for greater computation. As the cost of processing falls this situation may change.

2.4

Clipping

It may seem a trivial matter to eliminate the possibility of a signal at the ripple peak
causing a clip by reducing the passband gain by the amount of the ripple. This is
adequate for static sine-wave signals but does not allow for the worst case of input.
For an FIR filter the very maximum output value, or overshoot, will occur with an
input signal that causes the magnitude of all the coefficients to be added. This input
signal would consist of positive and negative full-scale values that all have the same
sign (or all have the opposite sign) as the coefficient with which they are aligned.
This pattern is very unlikely to exist in a real signal but represents an upper bound to
the amount of overshoot that could occur. This value is shown for each of the filters
in figure 1 to 4 and can be seen to vary between 2.9 and 5.7dB.
Clipping of DAC interpolation filters can often be observed when presenting them
with a square wave peaking at close to full scale. (This may look clean but it can be
shown to be clipping by reducing the level slightly and a clipped ringing overshoot will
be observed on the filter output.) This clipping represents a sharp, or high order,
non-linearity that should be avoided. The step response shown in graph F of each
figure illustrates the extent of step overshoot for these example designs. It can be
seen that there is not much variation between them so this may not be a useful
indicator.
One approach to addressing the overshoot problem is to reduce all the coefficient
values by the same proportion. This results in a gain reduction and will normally
reduce the dynamic range of by the same amount.
Another method of eliminating the overshoot is to design the filter so that there is
less ringing. For example, design it with a wider transition region between passband
and rejection band (this is obviously easier with a higher sample rate). Compare, for
example the maximum overshoot values for figures 3 and 4.

3.1

INADEQUATE REJECTION

The purpose of the anti alias and anti image filters is to reject signals at frequencies
above the folding frequency (0.5fs). It has also become accepted wisdom that for
high quality applications these filters should have a flat response to 20kHz and that
immediately above that frequency the response should plummet into the rejection
band. For system operating at 44.1kHz this requires a transition region from 20kHz
to 22.05kHz (0.454fs to 0.5fs) in which the filter should develop the full rejection of
the filter if it is to avoid the generation of aliases or images. In practice - as an
examination of table 2 reveals - many of the (otherwise) highest performance
integrated circuits do not achieve this.

1
2
3
4
5
6

Type
ADC
DAC filter
DAC
ADC
ADC
DAC

Stopband attenuation
100dB
110dB
60dB
110dB
117dB
90dB

Stopband edge
0.604fs
0.5465fs
0.55fs
0.5465fs
0.4979fs
0.55fs

Attenuation at 0.5fs
3dB
6dB
10dB
no data
117dB
6dB

Table 2

3.2

Filter design compromises

The minimum length of an FIR low pass filter is related to three parameters. It
increases with reduced transition region width, reduced maximum pass-band error
(or ripple), and with increased minimum stop-band rejection. (Empirical formulae
relating these parameters to filter length are given in [10])
A decimating FIR filter with N non-zero coefficients requires N multiplications to be
performed per output sample. For an interpolating stage many of the input samples
will be zero filled: In this case the filter multiplication rate is N per input sample.
(There are also methods of saving on this computation but these do not change the
importance of the three parameters of the previous paragraph.)
The length of the filter also infers a storage requirement with its associated costs.
The processing and storage requirements normally limit the size of the filter, but in
some cases the filter group delay is important. In applications such as artist return
feeds, the delay is more important than other aspects of the filter response. In that
case the filter size may be restricted (as in the low group delay mode of the CS5397
[11]), or an asymmetrical FIR (with a non-flat group delay) may be used.

3.3

Half-band filters

One of the methods of saving complexity in the low pass filter is to use half-band
filters. [12] These filters are a special case of a low pass FIR filter with a frequency
response that has the following symmetric property (where ffs is the filter sampling
frequency, equivalent to 2fs for a 2x oversampling filter):
H(f) = 1 - H(0.5ffs - f)
The difference in the frequency response from 1 below 0.25ffs and from 0 above
0.25ffs is symmetric about that frequency.
Given this property in a filter of odd length then every even coefficient - apart from
the central coefficient - is exactly 0. Graph C of figure 2 illustrates this. For a given
length of FIR this saves almost 50% in the number of multiply-accumulates that are
required.
This type of filter is used for interpolation or decimation by factors of two (then the
symmetry is around one half the lower sample rate).
There is, however, an important disadvantage to using this type of filter to form the
final decimation stage or first interpolation stage in a multi-stage oversampling filter.
The symmetry property requires that at a frequency of 0.25ffs (0.5fs) the attenuation

is 0.5. This means that it cannot provide adequate rejection to avoid images or
aliases close to that frequency.
It is also a restriction that the passband ripple and minimum rejection cannot be
specified separately. This may not be seen as much of a problem given the ability to
have a filter of twice the length for the same processing power.
4.0

IMAGING

For the listener who cannot hear signals above 20kHz it may be argued that antiimage filters for DACs operating with an input sample rate of 44.1kHz are not
required. All images will be above 22.05kHz and inaudible. (This case may be
argued even more strongly for 96kHz systems, when images are more than an
octave higher.)
Such arguments ignore the potential for non-linear behaviour in the electronic and
electromechanical stages following the DAC in the signal path. This non-linearity will
cause high frequency images above the audio band to intermodulate with signals
within the audio band. This would produce audio band intermodulation distortion
artefacts that could fall in the band.
For example, consider the half-band interpolation filter shown in figure 2.
Take, for an extreme example, a full scale input signal of 1kHz below the half sample
frequency. The image of this tone (to be suppressed by this filter) will be at 1kHz
above the half sample frequency, and at full scale. The filter attenuation at that
frequency is 25dB.
The output of the DAC following this filter will therefore have two tones 2kHz apart.
A second-order non-linearity in an amplifier or loudspeaker would produce
intermodulation products at the frequency sum and frequency difference of these
tones. The amplitude of the products would be equivalent to the product of the two
amplitudes multiplied by the coefficient of non-linearity.
If the second order non-linearity in the following stages (amplifiers and loudspeakers)
is worse than 1% at that frequency and level then the resulting distortion product at
2kHz will be approximately -70dB below the signal. Do not forget that - for our
listener at least - the signal itself is inaudible and will not have any masking effect.
Few people would choose to listen to this signal. It merely illustrates how the image
rejection performance of an interpolation filter could have a significant impact on the
audio band below 20kHz.
5.0

ALIASING

This is caused by sampling and is the reflection of the spectrum of a signal about the
0.5fs folding frequency. The effect of inadequate aliasing in a digital audio ADC subsystem is to produce frequency shifted signal in the audio band.
Most digital audio ADC sub-systems have poor alias rejection at close to the folding
frequency. Figure 1, graph A, is an illustration of the effect. The dotted line shows
the folded stopband rejection. This corresponds to the attenuation applied to aliases
that would be folded to the respective frequency. In this example the direct effect of

the poor alias rejection in the transition region would be inaudible for those who
cannot hear above 18kHz.
Signal just above the folding frequency will be hardly attenuated at all. The
intermodulation mechanisms mentioned in the previous section could also apply to
signals in this region. The frequency shifting in that region would change the
frequency of any lower frequency difference-frequency distortion. Such a change
may be perceived as unnatural and hence more noticeable. As before, this effect
depends on the linearity of the electronics (and mechanics) of the following
equipment.

6.1

FILTERS FOR 96KHZ SAMPLED SYSTEMS

The filters of figures 1 and 2 are modelled on products that can operate at 44.1kHz
with an equiripple passband to 20kHz. At that sample rate the equiripple band upper
edge is at 90% of the half sample rate.
The filters of figures 3 and 4 are specified for the equiripple region of the passband
to extend to 32% and 52% of the folding frequency respectively. These designs
illustrate the use of the extra bandwidth (provided by doubling the sample rate) for
providing a wider, or more relaxed, transition region. (Please note the distinction
between the passband and the equiripple region of the passband. For a low pass
filter with a very low ripple and a gradual transition region the traditional definition for
filter bandwidth (the -3dB frequency) will be significantly above the top of the
equiripple region, as illustrated in graph B of figures 3 and 4.)

6.2

31 tap decimation/interpolation filter

The 192kHz filter of figure 3 requires approximately the same multiply-accumulate

rate as the 63 tap filter of figure 1 operating at 96kHz (for a 48kHz output fs).
Graph A shows that this filter reaches the full attenuation of the stopband at half the
sample rate. This compares with figure 1 graph B where the attenuation of aliases
at the equivalent point is only 3dB.
The upper edge of the equiripple region is at 0.08 of the 192kHz sample rate,
15.4kHz. Above that frequency the roll-off is very gentle so that the attenuation at
20kHz is only about 0.3dB and the -3dB frequency is at 26.5kHz. This should be
compared with the filter of figure 1 which has 20kHz within the passband equiripple
region and the -3 dB frequency is at 25kHz.
The periodicity of the ripple, illustrated in graph D, at 10kHz per cycle transforms to
a timing of 0.1ms and the ripple amplitude of 0.003dB corresponds with a level of
-73dB. This compares with the 0.3ms at -67dB for figure 1.
The headroom that could be required for the worst possible input stimulus - the
maximum overshoot - is just under 3dB. This compares with 5.5dB and 5.7dB for
the first two filters.

6.3

63 tap decimation/interpolation filter

Figure 4 shows a filter with the same number of taps as that of figure 1. As the
sample rate is doubled this means that the computational rate is increased by the

same amount and that the filter length is halved in time. Like the filter of figure 3 the
stopband lower edge is aligned with half the sample frequency.
The passband upper edge of this filter is at 0.13 of the 192kHz sample rate, or
25kHz, and the graph B shows the transition region rolls off to the -3dB frequency at
34.7kHz.
This filter is long enough to reduce the passband ripple to 0.000 006 dB which
transforms to a pre-echo amplitude of -126dB. The periodicity of this ripple
transforms to echo timing of 0.18ms.
The maximum overshoot is worse than that of the previous filter, at 4.8dB, and so
there is minimal dynamic range advantage for this over the filters of figures 1 and 2.
7.0

CONCLUSION

Linear and non-linear distortion mechanisms within digital audio FIR low pass filters
used for decimation and interpolation have been described. Four filters have been
designed using the REMEZ algorithm. Two of these are based on commercially
available devices in common use and these are used to illustrate the issues.
The typical cosinusoidal passband ripple characteristic has been analysed to
estimate the time-dispersion characteristics of the filter to signals within the audio
band. This analysis is used to show significant differences in the pre and post-echo
performance between the example filters.
Susceptibility of the filters to overshoot has been illustrated and the effect of
compromises in stopband performance close to the folding frequency are discussed.
Filters designed for the higher sampling frequency of 96kHz are used to show how all
these distortion mechanisms can be reduced by filters of similar computational
requirements but with a more relaxed transition region. This results in either reduced
ripple (and hence pre-echo) amplitude or in a lower time displacement for the echo.
In both cases the effect is likely to be a reduction in the audibility of the echo. A
direct effect of the higher sampling rate is that for an identical filter design the time
displacements will scale inversely with sample rate. Hence an improvement can be
made just from raising the sample rate - even for those who cannot hear above
20kHz.
More work is required to evaluate the limits on the perception of the echo effects
described here. This should cover both the audibility of echoes and their effect on
the localisation of sound sources. One direct benefit of the increased use of 96kHz
for recording is that there will be an increasing amount of source material suitable for
this work.
The effects described here indicate that it may be difficult to distinguish any
beneficial effects of an increase in sampling frequency from the different filter
behaviour. This should be considered when making comparisons between different
rates.

8.0

REFERENCES

[1]
S. Yoshikawa, S. Noge, M. Ohsu, S. Toyama, H. Yanagawa and T.
Yamamoto Sound Quality Evaluation of 96 kHz Sampling Digital Audio Preprint
4112 of the 99th AES Convention, New York, October 1995.
[2]
Takeo Yamamoto Sound Quality of 96 kHz Sampling Digital Audio
Presentation to workshop High Bandwidth High Quality Digital Audio at the 101st
AES Convention, Los Angeles, November 1996
[3]
S. Yoshikawa, S.Noge, T. Yamamoto and K. Saito Does High Sampling
Frequency Improve Perceptual Time-Axis Resolution of Digital Audio Signal?
Preprint 4562 of the 103rd AES Convention, New York, September 1997.
[4]
Bernd Thei and Malcolm Hawksford Phantom Source Perception in 24 Bit
@ 96 kHz Digital Audio Preprint 4561 of the 103rd AES Convention, New York,
September 1997.
[5]
J. H. McClellan, T. W. Parks and L. Rabiner A Computer Program for
Designing Opyimum FIR Linear Phase Digital Filters IEEE Transactions on Audio
and Electroacoustics, Vol. AU-21, No. 6, pp. 506-526, December 1973.
[6]
R. Lagadec and T. G. Stockham, Dispersive Models for A-to-D and D-to-A
Conversion Systems Preprint 2097 of the 75th AES Convention, Paris, March 1984.
[7]
Thomas Sporer and Holger Schrder, 'Measuring Tone Masking Noise',
Preprint 3349 of the 93rd AES Convention, San Francisco, October 1992
[8]
R. H. Wilkinson, 'High-fidelity finite-impulse-response filters with optimal
stopbands' IEE Proceedings-G, Vol. 138, No. 2, pp. 264-272, April 1991.
[9]
K. Steiglitz, T.W. Parks and J.F. Kaiser, METEOR: A Constraint-Based FIR
filter design program IEEE Trans. Signal Processing, Vol. 40, No. 8, pp.1901-1909,
August 1992.
[10]
R.E. Crochiere and L.R. Rabiner Interpolation and Decimation of Digital
SIgnals - A Tutorial Review Proceedings of IEEE, vol. 69, no. 3, pp.417-448, March
1981.
[11]
Crystal Semiconductor data sheet CS5396 CS5397 120dB, 96kHz Audio A/D
Converter September 1997.
[12]
Fred Mintzer, 'On Half-Band, Third-Band, and Nth-Band FIR Filters and Their
Design' IEEE Transactions on Acoustics, Speech and Signal Processing, Vol.
ASSP-30, No. 5, pp. 734-738, October 1982.
9.0

LIST OF FIGURES

Figure 1
Figure 2
Figure 3
Figure 4

(2 pages)
(2 pages)
(2 pages)
(2 pages)

ADC decimation filter for 48kHz output sampling rate

DAC interpolation filter for 48kHz input sampling rate
A 31 tap filter for 96kHz sampling rate
A 63 tap filter for 96kHz sampling rate

63 tap decimation filter

Input fs = 96kHz, Output fs = 48kHz
REMEZ design parameters:
BAND 1
BAND 2
LOWER BAND EDGE
0.000000000
0.292000000
UPPER BAND EDGE
0.229000000
0.500000000
DESIRED VALUE
1.000000000
WEIGHTING
1.000000000

0.000000000
10.000000000

Graph A - Frequency response

20
20
(d
B)

60
100
140 0

12

18

24
30
Frequency (kHz)

36

42

48

Direct response
Alias rejection against output frequency
Graph B - Passband magnified

(d
B)

16

24
Frequency (kHz)

32

40

48

As scale
Scaled x 100
Graph C - Filter coefficients
0.5

V
al
u

10

20

30
Tap
Figure 1
Julian Dunn 1998

40

50

60

Graph D - Passband approximation

1.001

G
ai
n

0.999 0

10

15
Frequency (kHz)

20

25

30

Filter response
Cosine approximation
Graph E - Transform of approximation
0
(d
B)
100
1

0
Time (ms)

Graph F - Filter step response

1.5
1
0.5
0
0.5
1
1.5

40

30

20

10

0
10
Time (samples)

Step_overshoot = 1.092 dB
Maximum_overshoot = 5.497 dB

20

30

Fsample = 96 kHz

40

153 tap half-band interpolation filter

Input fs = 48kHz, Output fs = 96kHz
REMEZ design parameters:
BAND 1
BAND 2
LOWER BAND EDGE
0.000000000
0.273250000
UPPER BAND EDGE
0.226750000
0.500000000
DESIRED VALUE
1.000000000
WEIGHTING
1.000000000

0.000000000
1.000000000

Graph A - Frequency response

20
20
(d
B)

60
100
140 0

12

18

24
30
Frequency (kHz)

36

42

48

Direct response
Alias rejection against output frequency
Graph B - Passband magnified

0
(d
B)
2
0

16

24
Frequency (kHz)

32

40

48

As scale Scaled
x 100,000
Graph C - Filter coefficients
0.5
V
al
u

0
0

20

40

60

80
Tap

Figure 2
Julian Dunn 1998

100

120

140

Graph D - Passband approximation

1.000005

G
ai
n

0.999995 0

10

15
Frequency (kHz)

20

25

30

Filter response
Cosine approximation
Graph E - Transform of approximation
0
(d
B)
100
1

0
Time (ms)

Graph F - Filter step response

1.5
1
0.5
0
0.5
1
1.5

80

60

40

20

0
Time (samples)

Step_overshoot = 1.108 dB
Maximum_overshoot = 5.683 dB

20

40

60

Fsample = 96 kHz

80

31 tap decimation/interpolation filter

Higher fs = 192kHz, Lower fs = 96kHz
REMEZ design parameters:
LOWER
UPPER
DESIRED VALUE
WEIGHTING

BAND 1
0.000000000
0.080000000
1.000000000
1.000000000

BAND EDGE
BAND EDGE

BAND 2
0.250000000
0.500000000
0.000000000
40.000000000

Graph A - Frequency response

20
20
(d
B)

60
100
140 0

12

24

36

48
60
Frequency (kHz)

72

84

96

Direct response
Alias rejection against output frequency
Graph B - Passband magnified

(d
B)

16

24
Frequency (kHz)

32

40

48

As scale
Scaled x 100
Graph C - Filter coefficients
0.5
V
al
u

0
0

10

15
Tap

Figure 3
Julian Dunn 1998

20

25

30

Graph D - Passband approximation

1.001

G
ai
n

0.999 0

10

15
Frequency (kHz)

20

25

30

Filter response
Cosine approximation
Graph E - Transform of approximation
0

(d
B)
100
1

0
Time (ms)

Graph F - Filter step response

1.5
1
0.5
0
0.5
1
1.5

15

10

0
Time (samples)

Step_overshoot = 1.027 dB
Maximum_overshoot = 2.974 dB

10

Fsample = 192 kHz

15

63 tap decimation/interpolation filter

Higher fs = 192kHz, Lower fs = 96kHz
REMEZ design parameters:
BAND 1
LOWER
BAND EDGE
0.000000000
UPPER
BAND EDGE
0.130200000
DESIRED VALUE
1.000000000
WEIGHTING
1.000000000

BAND 2
0.250000000
0.500000000
0.000000000
1.000000000

Graph A - Frequency response

20
20
(d
B)

60
100
140 0

12

24

36

48
60
Frequency (kHz)

72

84

96

Direct response
Alias rejection against output frequency
Graph B - Passband magnified

0
(d
B)
2
0

16

24
Frequency (kHz)

32

40

48

As scale Scaled
x 100,000
Graph C - Filter coefficients
0.5
V
al
u

0
0

10

20

30
Tap

Figure 4
Julian Dunn 1998

40

50

60

Graph D - Passband approximation

1.000002

G
ai
n

0.999998 0

10

15
Frequency (kHz)

20

25

30

Filter response
Cosine approximation
Graph E - Transform of approximation
0
(d
B)
100
1

0
Time (ms)

Graph F - Filter step response

1.5
1
0.5
0
0.5
1
1.5

40

30

20

10

0
Time (samples)

Step_overshoot = 1.131 dB
Maximum_overshoot = 4.761 dB

10

20

30

Fsample = 192 kHz

40