See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/269102647

MATLAB program for loudspeaker equalization and crossover design

Conference Paper · September 1998

DOI: 10.13140/RG.2.2.23696.20486

CITATIONS READS

2 389

1 author:

Malcolm John Hawksford

University of Essex
231 PUBLICATIONS   964 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Down sampling-rate-conversion DSRC using spectral domain matching View project

Audio Research View project

All content following this page was uploaded by Malcolm John Hawksford on 09 August 2016.

The user has requested enhancement of the downloaded file.

 I

ENGINEERING
REPORTS

MATLAB Programfor LoudspeakerEqualizatiOnand

CrossoverDesign*

M. O. J. HAWKSFORD, AES Fellow

Centre for Audio Research and Engineering, University of Essex, UK C04 3SQ

A digital design filter program is presented written in the MATLAB environment. Impulse
response time editing is implemented together with various options for spectral domain
processing. The filter inputs time-domain impulse response data and outputs filter coefficients
for both FIR and ILR implementations. Comprehensive display options are incorporated,
including minimum-phase processing and CDS, and consideration is given to both specular
and diffuse loudspeakers. Applications include the design of digital equalization filters and
digital crossover filters for loudspeaker systems.

0 INTRODUCTION observe the effect on the overallequalization error.

The filter design can take account of a drive unit's
Digital filters enable accurate loudspeaker equaliza- individual response (which is entered as an impulse re-
tion and the implementation of near-theoretic crossover sponse) and matches this to a user-specified target fre-
alignments [1], [2]. As such it is possible to implement quency response. In its simplest form this target is con-
loudspeaker systems which yield extremely accurate stant with frequency, but can be adapted to include a
overall frequency responses by taking full account of number of equalization functions. The target function
the drive units' inherent response irregularities. This can also include a crossover filter, and a procedure is
technique is not only attractive for digital and active described in Section 2.3 (see [3] for more detail).
loudspeaker systems, but it also applies to passive loud- The program incorporates a range of data display op-
speakers that are preceded by a digital filter to fine-tune tions including time-domain data, frequency domain and
the frequency response, cumulative decay spectrum (CDS) together with the cu-
To develop crossover and equalizer filters, design mulative decay root (CDR) spectrum variant reported
tools are required to compute the digital filter coeffi- earlier [3], and the energy-time curve.(ETC).
cients that can accommodate a rangeof target responses, Also, a modified linear-phase Butterworth crossover
including those of the crossover. This engineering report filter is incorporated as well as "tilt" and psychoacoustic
describes a filter design program that is written within filters [4]. Pivotal to the design procedure is minimum-
the MATLAB 1environment (version 5.2). The program phase signal processing, which is used in both the filter
inputs data in the form of a loudspeaker's impulse re- synthesis and the data display options, including impulse
sponse and can embed specific equalization target func- and CDS. As such the program can be used to design digital
tions, including frequency response shaping and cross- filters for two-way digital and active loudspeakers [5].
over design. A flexible design is offered, where either The background to the signal processing techniques
an FIR or an IIR filter structure can be selected and the was described in a recent paper [3], and it is suggested
number of coefficients in the numerator and the denomi- that this be read as a companion to the present text. The
nator specified. An excess-phase equalizer is also corn- program listing is available on request, 2while this report
puted to enable correction of phase distortion. This allows presents an overview of the computation procedures. To
the user to change the number of coefficients and then to illustrate the capabilities of the program two loudspeaker
examples are presented. The first is a conventional small
* Presented at the 105th Convention of the Audio Engi- two-way system using moving-coil drive units, whereas
neering Society, San Francisco, CA, 1998 September 26-29; the second is a distributed-mode loudspeaker (DML) [6].
revised 1999 June 29.
i MATLAB is the trade name of a commercial matrix-based
processinglanguage. 2E-mailaddress:[email protected].

706 J.AudioEng.Soc.iVol.47,No.9,1999September
ENGINEERING REPORTS LOUDSPEAKER EQUALIZATION AND CROSSOVER DESIGN

These widelydiffering designs were chosen specifically linear-phase response.

to enable a deeper appreciation of the characteristics of a) Constant gain with frequency (no additional
a DML when compared against conventional specular equalization).
radiating devices. However, it should be observed that b) Constant slope filter (dB gain against logarithmic
a DML is a spatiotemporal dispersive radiator. As such frequency scale) of N dB per octave (both positive and
the example presented here is specific to one measure- negative slopes enabled). This modification has been
ment location. In practice there is stochastic variation found to be a useful tool in matching loudspeakers to
in the polar response, different room acoustics.
c) Linear-phase loudspeaker crossover response us-
1 OVERVIEW OF DIGITAL EQUALIZATION ing a modified Butterworth amplitude alignment. Both
FILTER DESIGN low-pass and high-pass options are available, and cross-
over frequency is specified at -6-dB gain. H°Wever,
The equalization filter design procedure is reviewed the Butterworth response is modified to guarantee that
briefly in this section, high-pass filter and low-pass filter sum to a unity-
1) Impulse response measurement. A direct measure amplitude composite response. (See Section 2.3 for
of the loudspeaker impulse response is captured using, more detail, as well as [3].)
for example, a maximum-length sequence (MLS) d) Psychoacoustic weighted frequency response,
excitation, available for subjective tailoring in the midrange fre-
2) Time-domain editing of measured impulse response quency region where the ear is most sensitive. This char-
a) Preresponse editing. Redundant samples prior acteristic has found favor in Germany [4].
to the main impulse response can be removed using on- The program allows these options to be selected individ-
screen editing. However, any prering due to antialiasing ually and mixed if required.
filters must be considered in this truncation process. 5) Calculation of equalizer amplitude-frequency re-
b) Postresponse editing. First and subsequent sponse. The equalizer amplitude response is calculated
boundary reflections from within the measurement envi- by inverting the product of the loudspeaker amplitude-
ronment can be removed together with noise that often frequency response [which has been both time-domain
contaminates the low-level "tail" of an extended impulse and frequency-domain edited, see 2) and 3)] and the
response measurement. This truncation includes a short target equalizer amplitude frequency response. This is
raised-cosine window to smooth the endpoint data. performed only on the magnitude response.
3) Frequency domain editing of the transformed, time- 6) Calculation of minimum-phase impulse response of
edited impulse response. Because the loudspeaker im- equalizer. Using the Hilbert transform, the minimum-
pulse response has been captured over a finite time win- phase impulse response of the equalizer is evaluated.
dow and at a finite sampling rate, the high-frequency and 7) Digitalfilter design. The equalizer's impulse re-
low-frequency measured responses are in error. Also, in sponse is divided into two regions within the program.
designing an equalization filter the extrema of the high- The first region defines an FIR filter, whereas an IIR
frequency and low-frequency responses may be defined filter represents the second region where coefficients are
so that the equalizer is not required to provide excessive calculated using either the Prony method 3 [7] or a least-
and inappropriate gain. A degree of modification can be mean-square (LMS) technique. The method that yields
performed again by on-screen editing of the displayed the lower error can then be selected. A general digital
amplitude-frequency response, filter is then defined with both numerator and denomina-
a) Low-frequency editing of amplitude-frequency tor coefficients.
response. Two edit points are selected on screen (using 8) Excess-phase equalization of loudspeaker. Using
the mouse), and a curve is computed using a quadratic the Hilbert transform, the truncated loudspeaker impulse
approximation. Consequently the inverse Fourier trans- response is decomposed into minimum- and excess-
form of the impulse response now extends in length phase components and then described in the time domain
beyond the original truncated response. For example, [3]. Using truncation and time reversal of the excess-
where an impulse response is truncated in 2) to 256 phase impulse response, the filter coefficients required
samples and subsequently represented .for processing by for phase correction are estimated. The phase equalizer
4096 samples, the additional samples computed to match can then be implemented either as an additional cascaded
the edited frequency response are virtually noiseless, filter or convolved with the numerator coefficients de-
b)High-frequency editing ofarnplitude-frequency rived in 7) to form a single filter.
response. A similar two-point selection editing pro-
cess is performed at high frequency, allowing the se- 2 PROGRAM FUNCTIONALITY
lected high-frequency curve to be replaced by a qua-
dratic approximation. This section highlights some of the mathematical pro-
4) Linear-phase target amplitude-frequency re- cedures used within the main program anddescribes the
sponse. The following four options have been included functionality in pseudo-MATLAB code.
in the program for modifying the overall frequency re-
3 The Prony function was authored by L. Shure and is written
sponse. They are performed only on the amplitude re- into the MATLAB signal processing suite as a function. Details
sponse. Consequently their native form results in a can be found in [7].

J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September 707

HAWKSFORD ENGINEERING
REPORTS

1 and that the first derivative is zero. Also, the curve

2.1 Minimum-Phase Processing must match in level at two adjacent samples x = m with
The minimum-phase signal-processing algorithm used a corresponding amplitude y(m) and x = m + 1 with a
in the program has been described previously [3]. The corresponding amplitude y(m + 1).
algorithm enables a straightforward method of comput- Now for a cubic curve,
ing the peak-level normalized minimum-phase impulse
response eit from a magnitude-frequency response xefa y = ax 3 + bff + cx + d .
(which is a symmetrical Fourier transform about fs/2,
where fs is the sampling frequency), that is, Then

eit = real(ifft(exp(conj(hilbert)log(xefa)))))); dy = 3ax2 + 2bx + c

eit = eit./max(abs(eit)); dx

2.2 Low-Frequency and High-Frequency Editing where for x = 1, dy/dx = O,

of Amplitude-Frequency Response
This editing routine modifies the amplitude-fre- 0 = 3a + 2b + c.
quency response at both low frequency and high fre-
quency. Two options are analyzed, a cubic fit and a In addition, matching levels at x = m and x = m + 1
quadratic fit. generates four equations,

o '_ 3 2 i o]
y(m+ 1) (m + 1)3 (m + 1)2 (m + 1) 1

y(m) m3 m2 m 1

y(1) 1 1 1 1]

Hence matrix inversion allows the coefficients

2.2,_ Cubic Fit Approximation [a b cd] to be calculated, and the cubic curve is
In each of these frequency regions two edit points are fully defined. A similar calculation procedure can be
marked and a cubic-approximation curve is generated to performed at high frequency.
replace the actual response. As shown in Fig. 1 the curve
2.2.2 Quadratic Fit Approximation
is matched in both level and first derivative by forcing
a fit to two adjacent samples, x = m and x = m + 1, A similar procedure can be followed for a quadratic
to smoothly link with the amplitude response. For low approximation, except that there is one less degree of
frequency, the first derivative is also set to zero for freedom. Thus here, as shown in Fig. 2, the derivative
sample x = 1, whereas at high frequency this is per- can be matched only at one of the extreme ends of the
formed at f/2. By way of example, the procedure is spectrum.
shown for the low-frequency cubic fit. Now for a quadratic curve,
Assume that the cubic curve has a value y(1) at x = y = axa + bx + c .

Cubic fit c ontinu ous derivative

- .... 7(m+l)
cubic curve

first derivative zero /_ .... '- -'y(m)

/
y(1)- x

x=l x=-m x=m +1

Fig. 1. Low-frequency cubic substitution (continuous derivative at x = m).

708 d.AudioEng.Soc.,Vol.47,No.9, 1999September

ENGINEERING REPORTS LOUDSPEAKER EQUALIZATION AND CROSSOVER DESIGN

Then, characteristic aj to enable a two-way digital crossover

filter to be designed and made compatible with aj.
dy The frequency range over which the computation is
dx = 2ax + b performed is defined by a frequency vector mx with an
element range {1:m2}, where
where for x = 1, dy/dx = 0,

0 = 2a + b. mx(l:m2) = ml*(l:m2);

Two options now exist. Either the derivative at the ex- Here mi is the lowest frequency in the vector whereas
treme frequency range can be equated to zero, or the m2*ml = fs/2, where fs is the sampling rate.
function can be matched to the derivative of the signal The process commences by calculating an amplitude-
by again forcing a fit to two adjacent points at x = m frequency response based on a Butterworth magnitude
and x = m + 1. This gives rise to two matrix equations, filter template.
which can be solved for [a b c] by matrix inversion, The crossover frequency xof (scaled by ml to form a
that is, either for a zero first derivative at low-frequency, corresponding element number in the vector mx) and
the asymptotic attenuation slope of the filter N dB per

design. The order ord of the Butterworth filter is then

y(rn) = 2 m 1 determined,

[ y(1)
0 ] Ii 1 01 ord = abs(N/(20*logl0(2)));
restrictions on the filter) are specified as input to the
octave (N need not be an integer as there are no "analog"

or for a matching first derivative of the actual fre- from which the 3-dB break frequencies xofl and xofh of
quency response, the low-pass and high-pass filters are calculated,

[yml[m
mil]II
xomxof
y(m)

y(1)ora / =
1
]
rn2

1
m

Matching at high or low frequency is similar, although

there are detailed changes of variables in the equations.
1

1
xofh = ml*xof*(.5 A( -- 2/ord) - 1) A .5;

The amplitude-frequency responses of the low-pass and

high-pass filters ail and ajh based on the Butterworth
magnitude frequency response then follow as

In practice the quadratic fit with a forced zero first deriv- ail = aj./(1 + (mx/xofl). A2). A(orr/2);
ative at the frequency extremes was found to give the
best overall results and has been adopted in the current ajh = aj.*(((mx/xofh), n 2)./(1 + (mx/xofh). n 2)). ^ (orr/2);
version of the program.
However, the composite zero-phase summation (ail +
2.3 Pseudo-Butterworth Linear-Phase ajh) does not generally sum to the target response aj,
Crossover Alignment which itself may not be fiat due to other frequency shap-
The program includes a routine to embed either a low- ing characteristics selected in the program. Co^se-
pass or a high-pass filter in the target equalization quently a symmetrical (about the crossover frequency

Quadraticl-i

y(m+l)
disc on t/nu ous de_hvative _-_----_
quadratic curve _ / ',
first derivative zero / ._ .... '_-'-3'(m)

y(1)- x I

x= 1 x=m x=m +1
Fig. 2. Low-frequency quadratic substitution (discontinuous derivative at x = m).

J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September 709

HAWKSFORD ENGINEERING
REPORTS

xof) compensation to ajl and ajh must be made. The impulse response eit can form a digital filter di-
For ajl, the response is modified in the frequency rectly, although in general it represents an excessive
range 1 to xof, number of coefficients. In the present program the im-
pulse response is subdivided into two regions. The first
ajl(l:xof) = ajl(l:xof) + aj(l:xof)- ajl(l:xof region of length en forms the first-stage FIR filter di-

-ajh(l:xof); rectly, whereas the remainder of the impulse response

is represented by a second-stage FIR-IIR filter and is
while for ajh, the response is modified in the frequency designed using either the Prony method or an LMS
range xof to m2, method. Both processes are supported in MATLAB. In-
tuitively this appears a logical subdivision as the FIR
ajh(xof:m2) = ajh(xof:m2) + aj(xof:m2) response carries much of the fine detail required of the
equalizer. However, the low-level tail of the impulse
- ajl(xof:m2) - ajh(xof:m2); response is primarily a result of bandwidth constraints

Now when the summation (ajl + ajh) is formed over such as the natural low-frequency rolloff of a loud-
{l:m2}, the composite amplitude response is aj, which speaker. As such it may be surmised to have a simpler
form more amenable to a low coefficient representation.
is the target response of the equalizer. However, because en can be user selected together with
2.4 Tilt Filter the second-stage filter numerator coefficients nx and de-
nominator coefficients nd required to represent the tail,
A simple frequency tilt filter is included where the
a wide range of filters can be chosen extending from
target equalization response aj can be tilted at a constant pure FIR to pure IIR. As the second-stage filter is an
slope of Nx dB per octave (as described in a logarithmic FIR-IIR form, the total number of numerator coeffi-
space). The response is normalized to unity gain at the cie^ts in the overall filter is (en + rix). However, it
geometric frequency mean fj and computed as
turns out that when the tail response is calculated, it is

aj = (mx/fj). ^ (Nx/(20*log 10(2))); not generally minimum phase. Consequently additional

numerator coefficients nx can be specified. However,
2.5 Psychoacoustic Equalizer the number of coefficients in each of the three categories
remains under user control, and through the use of time
A psychoacoustic filter was described earlier [3], [4]. and frequency displays the impact of a given design
The filter adds a mild reduction in the midband frequency selection can be observed. Also, after en is specified,
region where the ear is most sensitive. This can prove the program can be selected to interrogate the tail re-
a subjectively useful modification for some (poorer) re-
sponse and to recalculate en so that the highest value in
cordings. This filter is included in the present program the initially selected tail impulse becomes the new first
as an option and is defined by the routine
sample of the tail. The stage-one filter length en is then

np = ml*fix(3000/ml); reassigned. An autoselect procedure can be selected to

ajj = [ - 2.5*logl0((ml:ml:np)/3000) 6.4*logl0(np + mi:mi:rah)/3000)];

ajj ajj - (ajj(m2) + min(ajj))/2; facilitate this process. However, if only an FIR filter is
required, then en remains fixed and a short raised-cosine
aJj = (10*ones(size(1 :m2))). n (ajj/20); window is applied to the end of the FIR filter.

aj = aj.*ajJi The tail response is selected according to

tail = eit(en + l:m2);
where again aj is the target amplitude-frequency response.
to which both a Prony [7] and an LMS filter design
2.6 Two-Stage FIR-IIR Filter Using Prony's procedure (the latter using the MATLAB invfreq func-
Method tion 4) are applied to determine an FIR-IIR filter with
Once the amplitude-frequency responses of the target nx numerator and nd denominator coefficients. In the
aj and of the loudspeaker c are determined, the exact LMS procedure an option is included to both bandlimit
equalizer amplitude-frequency response xefa can be and power-weight the tail spectrum, where appropriate
calculated, parameterscan be user selected. The MATLABfunc-
tions have a form
xefa = aj./c;
from which the normalized minimum-phase equalizer [tnp tdp] = prony(tail(l:tmx),nx,nd);

impulse response eit follows, [tnl tdl] = invfreqz(fft(tail) ,rix - 1,nd,);

eit = real(ifft(exp(conj(hilbert(log(xefa))))));
4 Written as a function in MATLAB, coauthored by J. N.
eit = eit./max(abs(eit)); Little, J. O. Smith, Lennart Ljung, and T. Krauss.

710 J.AudioEng.Soc.,VoL47,No.9, 1999September

 ENGINEERING REPORTS LOUDSPEAKER EQUALIZATION AND CROSSOVER DESIGN

After selecting the desired tail approximation using a However, in the program the practice reported earlier
normalized LMS error calculation, the FIR and IIR filter [3] of eliminating the noncausal distortion by minimum-
sections can be spliced to form a complete filter. In the phase processing is taken. That is, a vector is formed
program up to three vectors are outputted to describe representing the magnitude Fourier transform of et,
the overall filter. The first filter Nl(z) of length en repre-
sents the first-stage FIR filter while the second recursive fa = abs(fft(abs(et)));
filter has a numerator polynomial N2(z) of length nx and
a denominator polynomial D2(z) of length nd (the first from which a minimum-phase impulse response is corn-
coefficient being unity and not included in nd). The puted which has the same amplitude spectrum as fa,
complete z-domain equalizer response EQ(z) is then cal-
culated as ett = real(ifft(exp(conj(hilbert(log(fa))))));

EQ(z) = Nl(z) + N2(z)z -(eh+ ]) The option for displaying the ETC and the minimum-
D2(z) phase ETC is provided together with an excess-phase
where a tap delay of length {eh + 1} is included in the corrected ETC [3], where the magnitude of ett is con-
volved with the excess-phase impulse response f, that is,
second-stage filter to locate the approximated tail in its
correct position within the overall impulse response.
ere = abs(conv(abs(ett),f));
2.7 CDS Display
The generation of a CDS has been described [3] in The function ete then has a similar precursive response
to the envelope of the impulse response by including
MATLAB where either a linear or a logarithmic magni-
both minimum-phase and excess-phase attributes.
tude display can be formed. The option of applying a
two-dimensional Gaussian filter mask is also included
where fine detail may require smoothing for presenta- 3 FILTER DESIGN EXAMPLES FOR
tional reasons. In the program, options exist for in- LOUDSPEAKERS
eluding minimum-phase processing based on the algo-
rithm described in Section 2.1. This is a useful tool as This section presents example filter designs for two
classes of loudspeaker, the broad-band passive system
it gives often a better description of a loudspeaker's
and the digital and active loudspeaker system [5]. Dis-
performance stripped of the excess-phase distortion
cussion as to the advantages of digital and active loud-
which generally has lower subjective significance yet
may introduce dominant features in the impulse response speaker technology is also included.
displayed in the CDS. The CDS is described by a two- 3.1 Broad-Band Equalization Filters for FUll-
dimensional matrix cd, where Range Passive Loudspeakers

cd = 20*logl0(abs(fft(hankel(e(l:cm))))); This section compares two filter examples with the

results are shown in Figs. 3-12. The (a) parts are for
The MATLAB "mesh" function then translates this ma- a small two-way conventional drive unit loudspeaker
trix into a corresponding three-dimensional display whereas the (b) parts are for a single DML. All curves
forming the CDS. were generated from within the MATLAB design pro-
gram and for comparison 150 coefficients are force se-
2.8 CDR Display leered in the first FIR stage with 50 denominator and 30
The CDR [3] is used here to display the equalizer numerator coefficients in the second-stage tail filter. The
impulse response ei. First the roots rt of ei are calculated same filter size was used for both two-way and DML
and sorted into rank order. Then they are edited and loudspeakers.
converted into a corresponding magnitude-frequency The target response for filter set (a) includes a tilt of
response. Finally minimum-phase impulse responses are 0.5 dB per octave, whereas in set (b) it is flat. For the
calculated from the set of magnitude responses and con- DML its native response persists for well over 1000
verted into corresponding ETC (see Section 2.9), from samples, including diffuse room reflections. However,
which a two-dimensional matrix is assembled to describe this was truncated to about 250 samples for processing,
the CDR. During processing a frequency response ap- which is not strictly representative of this class of loud-
proximation is displayed to check the validity of the speaker. For proper analysis of a DML, large-space
CDR after the roots are sorted and edited, measurements are required to extend 'the first reflection
arrival time together with the option for spatial averaging
2.9 ETC Display with Excess-Phase Response data derived from the long impulse response and spa-
Correction tially diffuse character of these transducers.
The ETC et is calculated directly from the magnitude
envelope of the Hilbert transform of an impulse response 3.9 Digital Equalization Filters Including
function a, where Crossover Alignment
A two-way digital and active system is shown in Fig.
et = abs(hilbert(a)); 13. Here the digital-to-analog converters (DACs) are

· J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September 711·

HAWKSFORD ENGINEERING REPORTS

Times 10 1

0.2 0.8

0.6
0
0.4

-0.2 0.2

-0.4 0

-0.2 J ......

-0,4 Times I0 .......................

-0.8 -0.6

-1 -08
50 100 150 200 250 50 100 150 200
(a) (b)

Fig. 3. Time-edited impulse response derived using MLS measurement. (a) Two-way dynamic loudspeaker. (b) DML.

09 ............................................................................................................
°9 ........,.........._.........i..........!....................................................
0.8 ........................................................................................................... 0.8 ........i.........._.........i..........i .........i..........i ...................i........._..........L.
o,1.............................................................................................................
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .o'. .}.. .. .i. .. . .i . . . .: . . . .i.:.. .. _. . . . .i.. . . ._. . . . _. . . . J. . . . .i .
0.6 ......... _--......... ri T mes 0 ...... i.......... _......... _.......... ?......... i.......... i"'

...........................................................................................................
05 .........................

0_,,..,
_..........................

_.............
'
i......................................................
:

,............
0.4

_.............
_............
..............
_............. 0_ ._
°' ..

o. . .,_. ,.o. . . . . ,. . . ,. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .o,ll

,V. :_::_lJ-:?: ',oi.........
0'2_"F'"/'"'"-'_'
"ii
.........
_:_ .............
'i
............
{.............
!.............
0.1' ......................
:................................
o.1.........i...........
:.............
'...........
_............
i............
i.............
_............. o
i_ Tree i........
.....
.....
_' i i i i i m i i i
50 _00 180 200 250 300 360 80 100 180 =00 280 308 380 400 480 800
(a) (b)

Fig. 4. Minimum-phase impulse response. (a) Two-way dynamic loudspeaker. (b) DML.

0.4

02 .........................................................................................................
08i.-,
i !
..........,
i
.........
,..........
i
_..........
,,
_
:
.........._..........
:
',
..........
i. ,
.........,...........
o6 ........ ¢..........¢.........i..........i..........;..........i ..........i..........i .........i...........
o ---_--_---:
.............
'.........: 0.4.......i..........
::......... (..........
4.......... {..........
i..........
?..........
F.........
!........
7
-0.2 ........................

-0.4 ....................................
,....................................................................

-04 .-4..........i.........i..........i..........i..........i..........i.........._.........i...........

-o.8 ................................................................................
i_........................... -o.8 .....i ..........i.........i..........:._.........i..........i..........i..........i .........i...........
_86..........................................................................................................
-1 m
80 100
m
150
_
200
_
250
m
300
m
350 400
-1[!!!t; _ i i ! i i i
200 400 600 800 1000 1200 1400 1600 1800

(a) (b)
Fig. 5. Excess-phase impulse response. (a) Two-way dynamic loudspeaker. (b) DML.

712 J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September

 ENGINEERING REPORTS LOUDSPEAKER EQUALIZATIONAND CROSSOVER DESIGN

interfaced directly to the power amplifiers, resulting in rate drive unit equalization. Although analog crossover
a welcome reduction of analog signal processing. Using filters can be synthesized, a more precise strategy is to
digital signal processing (DSP) it is straightforward to define the low-pass filter AL(z) and the high-pass filter
integrate near-ideal crossover filters together with accu- Au(z) as delay derived, as this eliminates the nonlinear

l iiii
i
-'°
.......
_1_
-20
''...........
ii......
.......

101
.............
i
....
:.........
........
[....................................................... ii=''''''"'"=_'',
i ; i;;ill

102
i i ilJiiii

103
; ;
_ '_,,
'_;'.:il i iiii
,_'i
;ii;i;;

104
: : ::::

105 101
: : :::fill

102
i i iiiiiii

103
i :::;:iii

104
i i ;;;[

105

(a) (b)
Fig. 6. Target, loudspeaker, and equalized frequency responses. (a) Two-way dynamic loudspeaker. (b) DML.

-lO -lO

-2o 20

-3o -3o

-4o -4o

-5o -50
o o

SO 50

100 60 80 _
100 60 80
40 40
20 20
1500 1500

(a) (b)
Fig. 7. CDS. (a) Two-way dynamic loudspeaker. (b) DML.

0. 0_

-10. -10.

-20, -20.

-30. -30,

-40 -40,

-50 -50.
0 0

50 50
80 80
60 100 60
100 40 40
20 20
1500 1500
(a) (b)

Fig. 8. Minimum-phase CDS. (a) Two-way dynamic loudspeaker. (b) DML.

J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September 713

I
HAWKSFORO ENGINEERING
REPORTS

phase distortion of an all-pass crossover alignment, synchronization and exact replication.

where 3) Delay compensation to align on-axis acoustic cen-
ters of drive units is straightforward to implement using
AL(z) + AH(z) = z-_ a memory-based digital delay.
4) Simple adjustment for individual drive unit sensi-
where z- x is an appropriate delay, tivities without wasting power in passive elements of
Section 2.3 described a pseudo-Butterworth design a crossover, thus enabling amplifier power to be used
that matches this criterion and is included within the efficiently. Also, use of digital gain control (with dither)
filter design Program as an integral part of the target guarantees exact tracking of each filter channel.
function generator. 5) Because individual power amplifiers are directly
Some performance and system attributes of a digital coupled to each drive unit, there is a corresponding divi-
and active loudspeaker are summarized as follows, sion of signal power. This reduces the voltage and cur-
l) DSP enables the crossover target transfer functions, rent demand placed on an individual amplifier, where
including amplitude and phase compensation of each usually an individual drive unit has a benign terminal
drive unit together with enclosure characteristics, to be impedance compared with passive systems with compli-
specified to a high degree of accuracy. Thus the polar cated crossovers.
response can be optimized and the overall on-axis fre-: 6) Momentary clipping of a single power amplifier can
quency response can approach a constant-amplitude, be softened by preprocessing and has only a localized, in
linear-phase characteristic, the sense of frequency, impact on subjective perform-
2) Crossover filters can include options for either ance. Hence the system is more overload tolerant.
sharp frequency transition bands or more gentle fre- 7) Close coupling of amplifiers and drive units (com-
quency response slopes, including all-pass analog align- pared to passive systems) results in tighter control of
ments. The new class of interleave crossover alignments the speech coil and -lower dependence on nonlinearity
can be incorporated [3]. Multiple filters offer guaranteed in the drive unit impedance.

-20. -20,

-40. -40.

-60. -80.
0 0
20 20
40 1O0 40 1O0

60 SO 60 50
80 80
lOO0 1000

(a) (b)

Fig. 9. CDR of equalization filter. (a) Two-way dynamic loudspeaker. (b) DML.

-10 -10_

-20. -20.

-30. -30..

-40.

-50. -50
0 0

50 50
80 80
60 100 60
100 40 40
20 20
1500 150 0

(a) (b)
Fig. 10. CDS, minimum-phasecorrectiononly. (a) EqualiZedtwo-way dynamicloudspeaker. (b) Equalized DML.

714 J. Audio Eng. Soc.,Vol. 47, No. 9, 1999September

ENGINEERING
REPORTS LOUDSPEAKER
EQUALIZATION
ANDCROSSOVER
DESIGN

8) Use of current-drive technology or mixed current- ponding low-pass and high-pass filter responses, in-
drive-voltage-drive systems can enhance system per- cluding drive-unit response equalization. Each filter
formance, used a total of 285 coefficients.
9) Each DAC only handles a band-limited audio sig- Finally, the new class of interleave-crossover align-
nal, thus reducing intermodulation distortion and low- ment [3] is illustrated by way of example in Fig. 17,
ering the probability overload in drive-unit frequency- which is also incorporated in the program. The purpose
response correction processes, of this alignment is to distribute and randomize in fre-
10) Signals can be routed to an active loudspeaker quency the interference patterns that result in the off-
system via either an optical or an electrical digital inter- axis polar response [8]. Both low-pass and high-pass
face with central commands such as standby mode, vol- filters for a stochastic-interleave function are shown that
ume level, equalization programs being remotely down- includes a mild equalization tilt function. Fig. 18 shows
loaded from a central (or indeed distributed) control the resulting composite frequency response when there
center. There is the option here to define a local-area is a time offset between drive units. A complicated inter-
network for digital or video distribution systems, ference spectrum is formed in the crossover region that
In Fig. 14 an example crossover target filter design is in some respects analogous to the behavior of a DML
is presented based on the pseudo-Butterworth algorithm [6]. Also, by comparing the inverted connection com-
of Section 2.3, whereas Figs. 15 and 16 show the corres- posite responses shown in the top curves of Figs. 14 and

2.5 t_' ETC of loudspeakerimpulseresponse

q,
2

1.5 [_ Minimum-phaseof loudspeakerETC

1 -- "-"-"'"-'"-_- ---- -

[3.5 Loudspeakerimpulseresponse

-0.5 ' ' ' ' ' ' '

0 20 40 80 80 100 120 140
(a)

2.5 _ [ ETC of loudspeakerimpulseresponse ]

,
1.5 [_ I Minimum-phase
of loudspeakerETC

0.5 [ Loudspeakerimpulseresponse }

-0.5 ' ' ' ' ' ' '

0 20 40 80 80 100 120 140
(b)
Fig. 11. Minimum-phaseETC. (a) Two-way dynamic loudspeaker. (b) DML.

J. AudioEng.Soc.,Vol.47,No.9,1999September 715
HAWKSFORD ENGINEERING REPORTS

3
t

2.5
ETC of loudspeaker impulse response

2 _ '/_Wlx'x''_'_''_

1.5 convolved with excess phase impulse response

t Minimum-phaseof loudspeakerETC

0.5 Loudspeakerimpulseresponse.

-0.5

'1

O 20 40
I

60

(a)
80
I

I00
I -- I

120
I

140

2.5 ETC of loudspeaker impulse response Ii

[

Minimum-phase of loudspeaker ETC

1.5 convolved with excess phase impulse response

0.5 Loudspeakerimpulse response

0 _

-0.5

-1 I I I I I I l

0 20 40 60 80 1O0 120 140

(b)
Fig. 12. Excess-phase corrected ETC. (a) Two-way dynamic loudspeaker. (b) DML.

-- Driveunit compensation FIR phase IjrV

and crossover network
+ digital gain control
equalization ' sta

Digital
input

-- Driveunit compensation FIR phase I/V

end
+ crossover
digital network
gain control equalization stag

Fig. 13. Two-way loudspeaker system using digital crossover filters with amplitude and phase correction.

716 J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September

 ENGINEERING REPORTS LOUDSPEAKER EQUALIZATIONAND CROSSOVER DESIGN

17 (with and without interleave function, respectively),

the effect of a stochastic interleave function is dramati-
FT(linear
phase):
yIIow+highl,
bIlowl-Jhighl,
gIlowl.
r lhighl cally illustrated. In studying these results it should be
50 _ . recalled [3] that the interference patterns are bounded
· ]
/ bytheenvelopesIA[+ A.IandIA,- A.I.

i ..........

A MATLAB program to facilitate the design of loud-

speaker equalization and crossover FIR-IIR filters has
been described. Two equalization examples were pre-
sented without crossover filters and one example with a
I. 4 CONCLUSION
-s.............. crossoverfiltertogetherwith examplesof computedout-
put data. In particular, the minimum-phase and excess-
phase data of a DML reveal interesting detail about this
class of drive unit. The main differences observed are
-los ; in the impulse response, which has an extended duration
10' l0s l02 10' l0_ and "noiselike" character. This makes equalization less
Fig. 14. Low-pass and high-pass pseudo-Butterworth target straightforward. Because of the spatially diffuse form
responses, of a DML's frequency response, it is better to perform

FT(linearphase):y Ilow+highl,b Ilowl-Ihighl,

g Ilowl,rIhighl
10, L i I 60 ,, ,, ,,
/I Drive unit response _--_o',-_. _'[

...........
-'°F
.......
JJ-'i .................
i............
\i .....
I........... :o _
-20 ...........
J

-40

'BO ................. -80 ................................................. : ..............

-7o , i i 400 '/

101 102 103 104 105 101 102 lO 2 104 105

Fig. 15. Low-pass filter response including drive-unit equal- Fig. 17. Example target response of stochastic interleave align-
ization, merit. (Compare with noninterleaved response of Fig. 14.)

Composite frequency response (linear-phase targets) with polar offset

50

Drive unit response

o __.,,',_................ _,__,,
-5o................. :.,................. .,, ...........
_.................
-100
......... 15
............................
_ -10

15o ---- _-/- ......................................

/ ii
2QO
I_ ', response ', -30-
/
-250 / i i i -35 , , i
101 102 103 104 105 101 102 10a 104 105

Fig. 16. High-pass filter response including drive-unit equal- Fig. 18. Family of four composite responses time offsets be-
ization, tweed drive units.

J. Audio Eng. Soc., Vol. 47, No. 9, 1999 September 717

HAWKSFORD ENGINEERING
REPORTS

equalization to a spatially averaged frequency response, ever, a much greater intersample variation in the pre-
In practice it is recommended that the spatial average and postresponse of the excess-phase impulse response
be performed using measurements taken over a sampled would be anticipated as the measured window encircles
hemisphere. Failing this a truncated impulse can be used the transducer. This is implied by the transducer being
together with a more limited number of coefficients in spatially diffuse. It is in these areas where fundamental
the equalization filter. It is important to appreciate that performance differences between specular radiators and
spatial averaging is an' intrinsic requirement of DML DMLs are observed, from which a number of attributes
measurement, especially when considering equalization, can be deduced, for example, those relating to imaging
as these transducers offer an almost omnidirectional po- when multiple arrays are employed.
larresponse together withdiffuseproperties. (One could The embedded crossover design function was also
argue that this new class of polar response is omnidif- demonstrated and a brief discussion given on the merits
fuse.) However, the contrary is true for most conven- of digital and active loudspeaker systems. This was used
tional loudspeakers, where these generally show much to illustrate the stochastic interleave-crossover align-
wider variations in frequency response over a measure- ment which, it is suggested, disperses the interference
ment sphere. Consequently any spatial averaging win- patterns in the off-axis frequency response, thus low-
dows have to be applied over a more restricted measure- ering polar-related coloration in the crossover region.
ment area with corresponding limitations on listening
postion and associated room interaction problems. Expe- 5 REFERENCES
rience suggests that the on-axis response forms the best
data in well-designed, conventional loudspeakers. [1] R: Greenfield and M. O. J. Hawksford, "Efficient
The method of filter design has proved efficient in Filter Design for Loudspeaker Equalization," J. Audio
terms of computational time, and using anIIRimplementa- Eng. Soc., vol. 39, pp. 739-751 (1991 Oct.).
tion to describe the impulse response tail would appear [2] M. O. J. Hawksford and R. G. Greenfield, "A
intuitively to offer advantage. However, in practice, Comparative Study of FIR and IIR Digital Equalization
comparing a pure FIR design with a pure HR design showed Techniques for Loudspeaker Systems," Proc. Inst.
little overall advantage where for an N-coefficient filter Acoust., vol. 12, pt. 8, pp. 77-86 (1990).
a good fit to the minimum-phase impulse response is [3] M. O. J. Hawksford, "Digital Signal Processing
achieved over about N samples in each case. It appears Tools for Loudspeaker Evaluation and Discrete-Time
that the total number of coefficients, irrespective of their Crossover Design," J. Audio Eng. Soc., vol. 45, pp.
distribution between the two filter sections, is the more 37-62 (1997 Jan./Feb.), Correction, ibid., vol. 45, p.
critical factor. As reported in earlier work [1], [2], the 497) (1997 June).
results demonstrate that by using a digital equalization [4] B. Stark, "Ein Fazit," in Das Lautsprecher Jahr-
filter, extremely accurate system performance is possi- buch 86/87, M. Gaedke, Ed. (Hifisound, Mfinster, Get-
hie, and that a wide range of target frequency responses many, 1986).
can be easily accommodated. [5] M. O. J. Hawksford, "Digital and Active Loud-
The two loudspeaker examples selected attempt to speaker Systems for High-Quality Monitoring," in Proc.
highlight some performance differences between a spec- Active 95 (1995 Int. Symp. on Active Control of Sound
ular radiator and a DML. In particular, the DML and Vibration, Newport Beach, CA, 1995 July 6-8),
minimum-phase impulse response showed good form, pp. 1247-1258.
although the excess-phase impulse response revealed an [6] "NXT," white paper, Huntingdon, UK.
extended noiselike structure in addition to a well- [7] T. W. Parks and C. S. Burrus, Digital FilterDe-
focused central response, confirming its temporally dif- sign (Wiley, New York, 1987), p. 226.
fuse nature. It is therefore suggested that DMLs probably [8] A. Rimell and M. O. J. Hawksford, "Reduction
exhibit low spatial variation in both the initial period of of Loudspeaker Polar Response Aberrations through the
the minimum-phase impulse response and the central Application of Psychoacoustic Error Concealment,"IEE
period of the excess-phase impulse response (where Proc. Vision Image Signal Process., vol. 145, pp.
these responses also show low noise structure). How- 11-18 (1998 Feb.).

THE AUTHOR

Malcolm Hawksford is director of the Centre for investigated delta modulation and delta-sigma modula-
Audio Research and Engineering, a professor in the De- tion (now commonly known as "bitstream" coding) for
partment of Electronic Systems Engineering at the Uni- color television and the development of a time-
versity of Essex, and Postgraduate Scheme director, compression/time-multiplex system for combining lumi-
where his interests encompass audio engineering, elec- nance and chrominance signals, a forerunner of the
tronic circuit design, and signal processing. Professor MAC/DMAC video system.
Hawksford studied electrical engineering at the Univer- While at Essex University, he has undertaken research
sity of Aston in Birmingham where he gained a First principally in the fields of analog amplifiers, digital sig-
Class Honours B.Sc. and Ph.D. His Ph.D. program, hal processing, and loudspeaker systems. Since 1982
which was sponsored by a BBC Research Scholarship, research on digital crossover networks and equalization

718 J. AudioEng.Soc.,Vol.47,No.9, 1999September

 ENGINEERING
REPORTS LOUDSPEAKER
EQUALIZATION
ANDCROSSOVER
DESIGN

for loudspeakers has culminated in an advanced digital Section of the AES.

and active loudspeaker system being designed within Professor Hawksford has published in the Journal
the university. This was one of the first systems of its of the Audio Engineering Society and at the Society's
type and in 1986 a prototype was demonstrated at the conventions on topics that include error correction in
Canon Research Centre in Tokyo. Research topics have amplifiers, oversampling techniques, jitter and MLS
also encompassed oversampling and noise shaping tech- techniques, and loudspeaker crossover systems. His
niques applied to analog-to-digital and digital-to-analog supplementary activities include writing contributions
conversion, the linearization of PWM encoders, and 3- for Hi-Fi News and Record Review (a magazine at
dimensional spatial audio and telepresence including which he is a technical adviser) and Stereophile maga-
multichannel sound reproduction, zinc as well as designing high-end analog and digital
Since 1971 and throughout his career at Essex Univer- audio equipment. He is a chartered engineer and is a
sity, Professor Hawksford has lectured at university fellow of the AES, the Institution of Electrical Engi-
level in electronics and in particular the undergraduate ricers, and the Institute of Acoustics. Professor
scheme in audio engineering. He has supervised numer- Hawksford is a member of the technical committee of
ous Ph.D. and M.Sc. students, many of whom are now Acoustic Renaissance for Audio (ARA), a group that
employed within the audio industry. Malcolm is particu- has been instrumental in promoting multichannel,
larly proud of the fact that with his encouragement, two high-definition audio signals on high-capacity DVD
former research students have established the Singapore optical disks.

J. AudioEng.Soc.,Vol. 47, No.9, 1999September 719

View publication stats