First Order Digital Filters--An Audio Cookbook

Application Note AN-11 by Christopher Moore

Introduction Assembly language example (not optimized)

The simplest digital filters are first order FIR and IIR filters, Motorola 56002
singly or in cascade. While not as powerful as higher order fil- ;Low pass 1st order filter
move X:fsum,x0 ; source sample into x0
ters, they are nevertheless of considerable interest in audio for move Y:B_1,y0 ; filter coef into y0
bass and treble controls, DC offset removal, high pass filters, mpy x0,y0, a Y:B_2,y1 ; mpy source sample,
and low pass filters. Also appealing is the low computational ; get next coef
move X:sumz0,x1 ; get z-1 delayed source
burden that first order filters place on DSP resources. : sample into x1
Despite the simplicity of the first order filter, the DSP literature move
mac
x0,X:sumz0 ; update delayed source
x1,y1,a Y:A_2,y1 ; mac 2nd sum term,
is not forthcoming with a simple methodology that allows for ; get next coef
cookbook design. The method that I present here appears in an move x:intz0,x1 ; get z-1 output into x1
macr y1,x1,a ; macr 3rd and final term
article from 1978 by Ham and is based on relating a prototype move a,X:intz0 ; update z-1 out register
analog filter to the desired digital filter using the technique of move a,X:fsum ; update output
impulse invariance. See Mitra (1998), pages 425-431, for further
information on this technique. The values of b1 and b2 shown in the tables are simplified initial
values, but the normalization factor norm was used in the freqz
The system function of a first order discrete time filter is: evaluation. Normalization was set to give unity gain at either DC
b1 + b2 z −1 or Fn, as appropriate. The normalization factor can be further
H (z) = modified according to signal dynamic range requirements. To
1 − a 2 z −1
incorporate normalization, scale b1 and b2 through multiplica-
Using the technique of impulse invariance, one ends up with tion by norm.
exponential functions that let you map from the continuous time Each filter type is illustrated with a Matlab plot showing the
filter’s corner frequencies directly to the coefficients of the dis- digital filter response computed using Matlab’s freqz function.
crete time (digital) filter. Generally, the digital filter matches the prototype analog filter
The flow chart of a first order digital filter (Direct Form 1 quite closely. However, at corner frequencies above about a
structure) is shown below. Other structures, while yielding the quarter of the Nyquist frequency (5.6kHz here), the computed
same response, may be advantageous in some cases due to finite actual corner frequency begins to lag behind the target. Fur-
word length and other effects. thermore, as the signal frequency approaches Fn, the response of
b1 some of the filters deviates from the analog version by a few dB.
X(n)
Generally, the discrepancy can be ameliorated by the appropri-
Y(n) ate activation of a pole or zero of around +/-0.12.
b2
a2
z-1 z-1
Definitions of Terms
Fs (sampling rate) = 44.1kHz here, but can be any value
z-1 (unit delay of sampling) = 1/Fs, seconds
a2 is the coefficient of the pole feedback path. A plus sign in the
tables indicates that the feedback path has a positive gain. If you Fn (Nyquist rate) = Fs/2, 22.05kHz
are using the Matlab freqz function, you must enter -a2 to satisfy f: frequency, Hz
its sign convention for poles. The coefficient a1 is not shown on f1,f2: filter corner frequency, Hz
the flow chart as it relates to the gain of the output itself, which a1, a2, b1, b2: coefficients of the system function
is always one. abs(): absolute value function
To actually implement one of these filters, you will need a DSP e (base of natural logarithms) = 2.718..
environment in which to program a chip to carry out the indi- zero location: -b2/b1
cated sample storage, multiplication, and accumulation. pole location: a2
Pole
The first order pole by itself is capable of either low pass or high
pass responses, as can be seen below. The low pass curves look
potentially useful, but, by themselves, the high pass curves don’t
do much until pretty high frequencies. The high pass curves will
come in handy to lift curves where the zero response doesn’t
quite behave.
Pole (POLE_LH.M)
b1 b2 norm a1 a2
1 0 1/(1-abs(a2)) 1 varied from 0 to +/-
0.999 in steps of 0.1

Low Pass Filter

The low pass filter utilizes only the feedback portion of the filter
to realize a pole that sets f1, the corner frequency, appropriately.
In the curves below, f1 was set to 10, 20, 50, 100, 200, 500, 1k,
2k, 5k, and 10kHz. By enabling a zero at -0.12, you can add the
extra attenuation needed to more closely match the continuous
time filter at 10kHz and above. Furthermore, if additional roll
off is required in the range of 4kHz and higher, the zero can be
set even to -1.0, which gives zero output at Fn.
Low Pass (LOWPAS2.M)
b1 b2 norm a1 a2
1.0 0.12 (1-a2)/(b1+b2) 1 −πf 1 / FN
e
Zero
The first order zero by itself is also capable of either low pass or
high pass responses. The high pass curves look potentially use-
ful, but, by themselves, the low pass curves don’t do much until
pretty high frequencies. The low pass curves will come in handy
to pull down curves where the pole response doesn’t quite be-
have. It is very hard to look at these curves and deduce how to
choose pole and zero responses that together will yield the de-
sired filter.
Zero (ZERO_LH.M)
b1 b2 norm a1 a2
-1 varied from 0 to abs(b1)+abs(b2) 1 0
+/-1.0 in steps of 0.1

High Pass Filter

The high pass filter utilizes the feed forward portion of the filter
to realize a zero at DC and the feedback portion to create a pole
that pulls the rising response down to a flat response. The re-
sponse rises from zero at DC to a shelf at f1. This characteristic
is useful for DC offset removal filters. In the curve below, f1
was set to 10, 20, 50, 100, 200, 500, 1k, 2k, 5k, and 10kHz.
 High Pass (HIGHPAS2.M) treble graphs below, one corner frequency was fixed (300Hz for
b1 b2 norm a1 a2 the bass control and 1kHz for the treble control) and the other
corner frequency was varied in log spaced steps over a range of
1 -1 (1+a2)/2 1 −πf 1 / FN 10:1. The Matlab file for the treble control is given on the last
e
page.
Shelf (BASS_BC.M)
b1 b2 norm a1 a2
-1 −πf 2 / FN 1 1 −πf1 / FN
e −e

High Pass Filter with Low Frequency Shelf

The high pass filter utilizes only the feed forward portion of the
filter to realize a shelf at low frequencies. The response begins
rising off the shelf at f2. In the curves below, f2 was set to 10,
20, 50, 100, 200, 500, 1k, 2k, 5k, and 10kHz. By enabling a
pole at -0.12, you can add the extra gain needed to more closely
Shelf (TREB_BC.M)
match the continuous time filter at 10kHz and above.
b1 b2 norm a1 a2
High Pass (HIGHPASS.M)
-1 −πf 2 / FN (1-a2)/(b1+b2) 1 −πf1 / FN
b1 b2 norm a1 a2 e e
-1 −πf 2 / FN (1-a2)/abs(b1+b2) 1 -0.12
e

Design notes
Even though the expressions used to compute the coefficients
Shelf (bass, treble) are quite simple, the computation should be mechanized if more
The shelf filter utilizes both feed forward and feed back sec- than a few curves are required. I used Matlab, creating multiple
tions. If f1 < f2, output falls from higher to lower level with in- files, one for each of the graphs seen here (the file names are
creasing frequency, and response is suitable for bass boost or shown in parantheses), but MathCAD or even a spreadsheet
treble cut. If f2 < f1, output rises from lower to higher level with could also be used.
increasing frequency, and response is suitable for bass cut or
treble boost. Shelf level difference is dB(f1/f2). In the bass and
 Example Matlab file, TREB_BC.M In conclusion—
%TREB_BC.M Creates treble control with corner fre-
Although I’ve presented a variety of equalization and filter
quency at 1kHz and sweeps from -20dB to +20dB in 20 curves here, the world of first order filters offers still more pos-
steps sibilities. The zero and the pole each can be located in two pos-
%
% revision date: March 3, 2001 sible regions: between zero and +1 (DC), and between zero and
% -1 (Nyquist frequency). This gives four possible combinations,
% constants
Fs=44100; % sampling rate
and, since there are an infinite number of possible locations,
t=1/Fs; % sampling interval many more curves are possible.
Fn=Fs/2; % Nyquist frequency
minF=10; % minimum frequency of analysis Acknowledgements
maxF=Fs/2; % maximum frequency of analysis I would like to thank the reviewers of this application note for
numPts=256; % number of points of analysis
% their helpful suggestions--Professor Douglas Preis, Tufts Uni-
% create array of frequencies to use for freqz versity; Erik Anderson; Dave Tweed; and Roman Litovsky.
Farray=logspace(log10(minF),log10(maxF),numPts);
%
% create arrays of frequencies in Hz to use for
sweeping the treble control
Bibliography
Fcarray1=[ 1000 1258 1584 1995 2512 3162 3981 5011 Ham, P. A. L. “Simple Digital Filters,” Wireless World, July
6310 7943 10000 1000 1000 1000 1000 1000 1000
1000 1000 1000 1000 ]; 1978, p 83-87.
Fcarray2=[ 1000 1000 1000 1000 1000 1000 1000 1000 McClellan, James H., Shaver, Ronald W., and Yoder, Mark A.
1000 1000 1000 1258 1584 1995 2512 3162 3981
5011 6310 7943 10000 ]; DSP First: A Multimedia Approach, Prentice Hall, NJ, 1998.
%
clf; % clear graph
Mitra, Sanjit K. Digital Signal Processing--A Computer-Based
for A_index = 1: 21 % plot 21 times for 2dB steps Approach, McGraw-Hill, New York, 1998.
of boost and cut over +/- 20dB range
% Orfanidis, Sophocles J. Introduction to Signal Processing,
title ('treb_bc.m'); % do every iteration to get Prentice Hall, NJ, 1996.
correct appearance
xlabel ('frequency Hz');
ylabel ('amplitude dB');
axis([minF maxF -20 20 ]);
grid on;
%
f2=Fcarray2(A_index); % numerator (zero) freq
f1=Fcarray1(A_index); % denominator (pole) freq
%
%-------------------------------------------------
% compute the digital filter coefficients
% Mission statement of Seven Woods Audio
L1= exp(-pi*f2/Fn); % numerator
K1= exp(-pi*f1/Fn); % denominator I am an electrical engineering consultant specializing in the con-
% ception and design of products and circuits used in audio appli-
b1=-1, %numerator coefficient X(n)
b2= L1, %numerator coefficient X(n-1)
cations. My company, Seven Woods Audio, is committed to
a1= 1, %denominator coefficient Y(n) helping manufacturers quickly create digital or analog audio
a2= K1, %denominator coefficient Y(n-1) products that generate a good return on investment, work right
normalize=(1-a2)/(b1+b2); % normalize to AV=0dB at
DC the first time, sound excellent, and please the end user. Seven
b1=normalize*b1; Woods Audio works with manufacturers of professional audio,
b2=normalize*b2; consumer audio, broadcast, telecommunications, and computer
%
%define numerator and denominator of transfer func- equipment.
tion
num=[ b1 b2 ], den=[ a1 -a2 ],
% num=[ Xn Xn-1 ], den=[ Yn Yn-1 ], rev: March 3, 2001 Copyright 2001. All rights reserved.
%
% fill array H with transfer function amplitude
[H,W]=freqz(num,den,Farray,Fs);
%
% plot digital filter magnitude response, log in x,
linear in dB in y
semilogx(Farray,20*log10(H),'w');
hold on; % Hold plot from prior to next case
%
end % A_index loop
44 Oak Avenue Belmont, MA 02478-2715 USA
voice/fax 617 489 6292
[email protected]
http://www.world.std.com/~cmoore/