Optimal Crosstalk Cancellation For Binaural Audio With Two Loudspeakers

Optimal Crosstalk Cancellation for Binaural Audio with Two Loudspeakers
Edgar Y. Choueiri
Princeton University
[email protected]
Crosstalk cancellation (XTC) yields high-spatial-fidelity reproduction of binaural audio through

loudspeakers allowing a listener to perceive an accurate 3-D image of a recorded soundfield. Such
accurate 3-D sound reproduction is useful in a wide range of applications in the medical, military
and commercial audio sectors. However, XTC is known to add a severe spectral coloration to the
sound and that has been an impediment to the wide adoption of loudspeaker-based binaural audio.
The nature of this coloration in two-loudspeaker XTC systems, and the fundamental aspects of the
regularization methods that can be used to optimally control it, were studied analytically using
a free-field two-point-source model. It was shown that constant-parameter regularization, while
effective at decreasing coloration peaks, does not yield optimal XTC filters, and can lead to the
formation of roll-offs and doublet peaks in the filter’s frequency response. Frequency-dependent
regularization was shown to be significantly better for XTC optimization, and was used to derive
a prescription for designing optimal two-loudspeaker XTC filters, whereby the audio spectrum is
divided into adjacent bands, each of is which associated with one of three XTC impulse responses,
which were derived analytically. Aside from the sought fundamental insight, the analysis led to the
formulation of band-assembled XTC filters, whose optimal properties favor their practical use for
enhancing the spatial realism of two-loudspeaker playback of standard stereo recordings containing
binaural cues.
I. INTRODUCTION as both crosstalk and room reflections would be non-

existent. However, with earphones or headphones, the
A. Background and Motivation location of the playback transducers in or very near the
ears means that non-idealities, (e.g., mismatches between
the HRTF of the listener and that used to encode the
The ultimate goal of binaural audio with loudspeakers recording, movement of the perceived sound image with
(BAL), also known as transauralization[1], is to repro- movement of the listener’s head, lack of bone-conducted
duce, at the entrance of each of the listener’s ear canals, sound, transducer-induced resonances in the ear canal,
the sound pressure signals recorded on only the ipsilat- discomfort, etc.), when above a certain threshold, can
eral channel of a stereo signal. If the stereo signal [2] was lead to difficulties in perceiving a realistic 3-D image and
encoded with the head-related transfer function (HRTF) to the perception that the sound (or some of its spectral
of the listener, and includes the proper ITD (interau- components) is inside, or too close to, the listener’s head.
ral time difference) and ILD (interaural level difference)
cues, then delivering the signal on each of the channels of Binaural playback through loudspeakers is largely im-
the stereo signal to the ipsilateral ear, and only to that mune to this head internalization of sound, for even when
ear, would ideally guarantee that the ear-brain system non-idealities in binaural reproduction are present, the
receives the cues it needs to hear an accurate 3-D repro- sound originates far enough from the listener to be per-
duction of the recorded soundfield. Since, with playback ceived to come from outside the head. Furthermore, cues
from two loudspeakers, each of them is also heard by such as bone-conducted sound and the involvement of the
the contralateral ear (crosstalk), approaching the goal of listener’s own head, torso and pinnae in sound diffraction
BAL requires an effective cancellation of this unintended and reflection during playback (even if it departs from, or
crosstalk. interferes with, the diffraction-induced coloration repre-
In addition to crosstalk cancellation (XTC), effective sented in the HRTF used to encode the binaural record-
BAL requires an abatement of sound reflections in the ing) could be expected to enhance the perceived realism
listening room, which cause degradation to the integrity of sound reproduction relative to that achieved with ear-
of the binaural cues at the listener’s ears[3–5]. While phones. These potential advantages have, implicitly or
this problem can be somewhat alleviated through pre- explicitly, motivated the development of XTC-enabled
scriptions that increase the ratio of direct to reflected BAL since the earliest work on the subject[8–10].
sound, full disambiguation of front-back sound local- In scientific applications of BAL, such as its application
ization through BAL has been shown to require XTC to study spatial hearing disabilities of elderly adults[3],
levels[6] above 20 dB, which are difficult to achieve prac- the high levels of XTC (above 20 dB) needed for highly-
tically even under anechoic conditions[3]. accurate transmission of binaural cues to a listener re-
Therefore, it would seem that the goal stated in the quire anechoic, or semi-anechoic, environments, precise
first paragraph could be more naturally reached with matching of the listener’s HRTF with that used in the
binaural audio through headphones (or earphones)[7] recording, and constrained positioning of the listener’s
E.Y. CHOUEIRI: OPTIMAL CROSSTALK CANCELLATION 2
head in the area of equalization (“sweet spot”). In many fects on XTC performance, are the main subjects of this
less stringent applications[10], modest levels of XTC, paper.
even of a few dB over a limited range of frequencies,
have the potential of significantly enhancing the 3-D real-
ism of the reproduction of recordings containing binaural B. The Problem of XTC-induced Spectral
cues. This is because, by definition, localization cues in Coloration
a binaural recording represent differential interaural in-
formation that is intended to be transmitted to the ears 1. Nature of the Problem
with no crosstalk. In other words, crosstalk cancellation,
at any level, is a reduction of unintended artifice in the
One main difficulty in implementing XTC is to reduce
loudspeaker playback of recordings containing significant
the artifice of crosstalk without adding an artifice of an-
binaural cues.
other kind: spectral coloration. Sound waves traveling
This reduction of unintended artifice through XTC from two distinct sources to the ears set up an interfer-
should also apply to the loudspeaker playback of most ence pattern in the intervening air space. Depending on
stereo recordings[11], especially those made in real acous- the frequency, the distance of the ears from the loud-
tic spaces, and even to recordings made using standard speakers, the distance between the loudspeakers, and the
stereo microphone techniques without a dummy head, phase relationship between the left and right components
since these techniques[12] all rely on preserving in the of the recorded stereo signal, the wave interference at an
recording a good measure of the natural ILD and ITD ear of the listener might be destructive, complementary,
cues needed for spatial localization during playback. We or constructive. At some of the frequencies for which the
should therefore expect that effecting even a relatively interference is destructive at the ear, XTC control (i.e.,
low level of XTC to the playback of such standard stereo signal processing that would cause the waves from the
recordings, even those lacking HRTF encoding, should loudspeakers to the contralateral ears to be nulled) would
enhance image localization compared to playback with require boosting the amplitude of the emitted waves. As
full crosstalk, as well as the perception of width and we shall see in Section II C, for typical listening configu-
depth of the sound-field, since these binaural features rations these level boosts[28] in the case of a perfect XTC
are always, to some degree, corrupted by crosstalk[13]. filter (defined as one that theoretically yields an infinite
With such promises of high-spatial-fidelity reproduc- XTC level over the entire audio band, in a free-field or
tion of binaural recordings, and enhanced realism to the anechoic environment) can easily be in excess of 30 dB,
playback of a large portion of existing acoustic stereo and therefore amount to severe spectral coloration.
recordings, the question arises as to why crosstalk can- Of course, such a “perfect” XTC filter would impose
cellation has not yet penetrated widely in the professional these necessary level boosts only at the loudspeakers in
and consumer audio sectors. such a way that, at the listener’s ears, not only the
A part of the answer to this question is related to crosstalk is cancelled, but also the frequency spectrum is
the physical constraints required for the practical im- reconstructed perfectly, i.e., with no spectral coloration.
plementation of an effective XTC-enabled BAL play- As recognized in Ref. [29, 30], and as will be further
back system. These constraints include sensitivity to discussed in Section II C, the frequencies at which the
head movements and a limited sweet spot[14–18], sen- level boosts are required correspond to the frequencies
sitivity to room reflections[4, 5], and the often-required at which system inversion (the mathematical inversion
departure from the well-established stereo loudspeaker of the system’s transfer matrix, which leads to the XTC
configuration[19, 20] (where the loudspeakers span an filter) is ill-conditioned. At these frequencies, XTC con-
angle of 60 degrees with respect to the listener). Much trol becomes highly sensitive to errors[30], so that even
research effort has been expended recently on relieving a small error in the alignment of the listener’s head in
some of these constraints and has resulted in potential the real world would lead to an effective loss of XTC
solutions, of varying degrees of practicality, which in- control at, and near, these frequencies. Therefore, not
clude: widening the sweet spot through the use of multi- only would there be undesired crosstalk at the listener’s
ple loudspeakers[21–24], providing XTC at multiple lis- ears at these frequencies, but also, and consequently, the
tening locations[25], enhancing robustness to head move- levels boosts which must necessarily be imposed at these
ment through the use of sum and difference filters[26], frequencies, would be fully hearable, even in the sweet
and dynamically moving the sweet spot to follow the lo- spot, as a coloration.
cation of the listener’s head by tracking it with optical Even in an ideal world where the loudspeakers-listener
sensors[27]. alignment is perfect, this spectral coloration imposed at
Another major impediment to the wide adoption of the loudspeakers would present three probems: 1) it
XTC-enabled BAL has been the spectral coloration that would be heard by a listener outside the sweet spot, 2) it
XTC filters inherently impose on the sound emitted by would cause a relative increase (compared to unprocessed
the loudspeakers. The fundamental nature of this spec- sound playback) in the physical strain on the playback
tral coloration, its basic features, its dependencies, and transducers, and 3) it would correspond to a loss in the
optimal methods to abate it with minimal adverse ef- dynamic range[29]. Since even professional audio equip-
ment is seldom designed to have more more than a few dB trices. We then define a set of metrics that will be useful
headroom above the levels required to reproduce realistic for evaluating and comparing the spectral coloration and
SPL peaks[31], the dynamic range of the program must performance of XTC filters, and conclude with the defini-
be decreased by more than 30 dB (minus the headroom), tion and discussion of a benchmark for such comparisons:
in the case of the “perfect” XTC filter defined above, to the perfect XTC filter.
avoid clipping. This is particularly problematic, for in-
stance, in the case of wide-dynamic-range audio recorded
in 16 bits. A. Formulation and Transformation Matrices
In order to render the analysis tractable enough so that

2. Previous Work and Present Goals
fundamental insight is more easily obtained, we make
the idealizing assumptions that sound propagation oc-
Takeuchi and Nelson[29] have developed a method that curs in a free field (with no diffraction or refelection from
not only yields excellent measured XTC performance[3, the head and pinnae of the listener or any other physi-
32], but also effectively solves the problem of spectral col- cal objects), and that the loudspeakers radiate like point
oration. However, their method, called “Optimal Source sources.
Distribution” (OSD), which is discussed in Section II C.4,
requires the use of a minimum of six transducers posi- DL DR
tioned at various angles around the listener.
The problem of XTC-induced spectral coloration for
playback with only two loudspeakers remains compelling
due to the simplicity of the two-loudspeaker set-up and
H
its compatibility with existing audio equipment. In this VR
VL
paper, we study this problem in the context of XTC op-
timization, which we define as maximizing XTC perfor-
mance for a desired tolerable level of spectral coloration
or, equivalently, minimizing the spectral coloration for a
desired XTC performance. C
In particular, we use a free-field two-point-source
model and address, analytically, the fundamental aspects l1 l1
of spectral coloration control through both constant-
parameter and frequency-dependent regularization meth- l l
l2 l2
ods. While regularization methods have been used in
the audio literature for sound source reconstruction[33]
and to control ill-conditioning in HRTF inversion[3, 23,
24, 34] their fundamental properties in the context of
XTC optimization and XTC-induced spectral coloration
have not been studied in detail, especially in the case of PR
PL
frequency-dependent regularization.
Aside from the fundamental insight we seek through ∆r
this analysis, we aim to derive analytical expressions for
the time-domain impulse responses (IRs) of such optimal FIG. 1: Geometry of the two-source free-field model. (All
XTC filters. These IRs are not only useful for shedding symbols are defined in the text.)
light on the time-domain aspects of XTC optimization,
but can also lend their benefits, through digital convo-
lution engines, to (typically non-scientific) applications In the frequency domain, the air pressure at a free-field
where the needed XTC levels are below those requiring point located a distance r from a point source (monopole)
inversion of the listener’s HRTF, and are sufficient to en- radiating a sound wave of frequency ω is given by[35]
hance the spatial realism of the two-loudspeaker playback
(in regular listening rooms) of audio signals containing iωρo q e−ikr
P (r, iω) = ,
significant binaural cues. 4π r
where ρ0 is the air density, k = 2π/λ = ω/cs the
wavenumber, λ the wavelength, cs the speed of sound
II. THE FUNDAMENTAL XTC PROBLEM
(340.3 m/s), and q the source strength (in units of vol-
ume per unit time). It is convenient to define
In this section we start with the mathematical formu- iωρ0 q
lation of the model and the governing transformation ma- V = ,
4π
which is the time derivative of ρo q/(4π), the mass flow In the time domain, α is simply a transmission delay
rate of air from the center of the source. (divided by the constant l1 ) that does not affect the shape
Therefore, at the left ear of a listener in the symmetric of the signal. Its role in insuring causality is discussed in
two-source geometry shown in Fig. 1, the air pressure due Section III B. The source vector v = [VL (iω), VR (iω)]T
to the two sources, under the above-stated assumptions, is obtained from the vector of “recorded” signals d =
add up as [DL (iω), DR (iω)]T , through the transformation
e−ikl1 e−ikl2 v = Hd, (9)
PL (iω) = VL (iω) + VR (iω). (1)
l1 l2
where
Similarly, at the right ear, we have  
HLL (iω) HLR (iω)
e−ikl2 e−ikl1 H=  (10)
PR (iω) = VL (iω) + VR (iω). (2)
l2 l1 HRL (iω) HRR (iω)
Here, l1 and l2 are the path lengths between any of the is the sought 2 × 2 filter matrix. Therefore, from Eq. (7),
two sources and the ipsilateral and contralateral ear, re- we have
spectively, as shown in that figure.
In order to maintain a connection with the relevant p = αCHd, (11)
literature, we adopt the same nomenclature used in
Refs. [19, 20, 29, 30]. Namely, unless otherwise stated, we where p = [PL (iω), PR (iω)]T is the vector of pressures at
use uppercase letters for frequency variables, lowercase the ears, and C is the system’s transfer matrix
for time-domain variables, uppercase bold for matrices
ge−iωτc

1
and lowercase bold for vectors, and define C≡ , (12)
ge−iωτc 1
∆l ≡ l2 − l1 and g ≡ l1 /l2 (3)
which, like all matrices we will be dealing with, is sym-
as the path length difference and path length ratio, re- metric due to the symmetry of the geometry.
spectively. An inspection of the geometry illustrated In summary, the transformation from the signal d,
in Fig. 1 shows that 0 < g < 1, and that the path lengths through the filter H, to the source variables v, then
can be expressed as through wave propagation from the sources to pressure
s p at the ears of the listener, can be written simply as
2
2
∆r
l1 = l + − ∆r l sin(θ), (4) p = αRd. (13)
2
where we have introduced the performance matrix, R,
s 2 defined as
∆r
l2 = l2 + + ∆r l sin(θ), (5)
 
2 RLL (iω) RLR (iω)
R=  ≡ CH. (14)
where ∆r is the effective distance between the entrances RRL (iω) RRR (iω)
of the ear canals, and l is the distance between either
source and the interaural mid-point. As defined in Fig. 1,
Θ = 2θ is the loudspeaker span. Note that for l >> B. Metrics
∆r sin(θ), as in most loudspeaker-based listening set-ups,
we have g ' 1. Another important parameter is the time We now wish to define a set of metrics by which to
delay, judge the spectral coloration and performance of XTC
filters. In this context we note that the diagonal elements
∆l
τc = , (6) of R represent the ipsilateral transmission of the signal
cs to the ears, and the off-diagonal elements represent the
defined as the time it takes a sound wave to traverse the undesired contralateral transmission, i.e., the crosstalk.
path length difference ∆l. Therefore, the amplitude spectrum (to a factor α) of
Using the above definitions, Eqs. (1) and (2) can be a signal fed to only one (either left or right) of the two
re-written in matrix form as inputs of the system, as heard at the ipsilateral ear is
ge−iωτc VL (iω)

PL (iω) 1 Esik (ω) ≡ |RLL (iω)| = |RRR (iω)|,
=α , (7)
PR (iω) ge−iωτc 1 VR (iω)
where the subscripts “si” and k stand for “side image”
where and “ipsilateral ear (with respect to the input signal)”
e−iωl1 /cs respectively, since Esiip , as defined, is the frequency re-
α= . (8) sponse (at the ipsilateral ear) for the side image that
l1
would result from the input being panned to one side. to combine Si (ω) and So (ω) into a single metric, Ŝ(ω),
Similarly, at the contralateral ear to the input signal which is the envelope spectrum that describes the maxi-
(subscript X), we have the following side-image fre- mum amplitude that could be expected at the loudspeak-
quency response: ers, and is given by
EsiX (ω) ≡ |RLR (iω)| = |RRL (iω)|. Ŝ(ω) = max [Si (ω), So (ω)] .
The system’s frequency response at either ear when the It is relevant to note that Ŝ(ω) is equivalent to ||H||,
same signal is split equally between left and right inputs the 2-norm of H, and that Si and So are the two singu-
is another spectral coloration metric. It can be obtained lar values, which can be obtained through singular value
T
from the product R · [1/2, 1/2] , which leads to decomposition of the matrix as was done in Ref. [29].
Finally, an important metric that will allow us to eval-
RLL (iω) + RLR (iω) RRL (iω) + RRR (iω) uate and compare the XTC performance of various filters
Eci (ω) ≡ | |=| |. is χ(ω), the crosstalk cancellation spectrum:
2 2
Here the subscript “ci” stands for “center image” since |RLL (iω)| |RRR (iω)| Esik (ω)
χ(ω) ≡ = = .
Eci , as defined, is the frequency response (at either ear) |RRL (iω)| |RLR (iω)| EsiX (ω)
for the center image that would result from the input
being panned to the center. The above definitions give us a total of eight metrics,
Also of importance to our discussions are the frequency (Esik , EsiX , Eci , Ssik , SsiX , Sci , Ŝ, χ), all real functions
responses that would be measured at the sources (loud- of frequency, by which to evaluate and compare the spec-
speakers). These are denoted by S, and can be obtained tral coloration and XTC performance of XTC filters.
from the elements of the filter matrix H. They are given
using the same subscript convention used above (with “k”
and “X” referring to the loudspeakers that are ipsilateral C. Benchmark: Perfect Crosstalk Cancellation
and contralateral to the input signal, respectively) by
A perfect crosstalk cancellation (P-XTC) filter is de-
Ssik (ω) ≡ |HLL (iω)| = |HRR (iω)|, fined as one that, theoretically, yields infinite crosstalk
cancellation at the ears of the listener, for all frequen-
cies.
SsiX (ω) ≡ |HLR (iω)| = |HRL (iω)|, Crosstalk cancellation, as defined in Section I A, re-
quires that the pressure at each of the two ears be that
which would have resulted from the ipsilateral signal
HLL (iω) + HLR (iω) HRL (iω) + HRR (iω)
Sci (ω) ≡ | |=| |. alone, namely, in the frequency domain, PL = αDL and
2 2 PR = αDR , where all quantities are complex functions of
An intuitive interpretation of the significance of the above frequency. Therefore, in order to achieve perfect cancella-
metrics is that a signal panned from a single input to both tion of the crosstalk, Eq. (13) requires that R = I, where
inputs to the system will result in frequency responses I is the unity matrix, and thus, as per the definition of R
going from Esi to Eci at the ears, and Ssi to Sci at the in Eq. (14), the P-XTC filter is simply the inverse of the
loudspeakers. system transfer matrix expressed in Eq. (12), and can be
Two other spectral coloration metrics are the fre- expressed exactly:
quency responses of the system to in-phase and out-of- 1

1 −ge−iωτc

[P ] −1
phase inputs to the system. These two responses are H =C = ,
1 − g 2 e−2iωτc −ge−iωτc 1
obtained simply from the product of the filter matrix H
T T T (15)
with the vectors [1, 1] and [1, −1] (or [−1, 1] ), respec- where the superscript [P ] denotes perfect XTC. For this
tively, and are given by: filter, the eight metrics we defined above become:
Si (ω) ≡ |HLL (iω) + HLR (iω)| = |HRL (iω) + HRR (iω)|, [P ] [P ] [P ] 1
Esi = 1; Esi = 0; Eci = ;
So (ω) ≡ |HLL (iω) − HLR (iω)| = |HRL (iω) − HRR (iω)|, k

X
2
[P ] 1 1
where the subscripts i and o denote the in-phase and out- Ssi (ω) = 2 −2iωτ
= p ;
k 1−g e c 4 2
g − 2g cos(2ωτc ) + 1
of-phase responses, respectively. Note that, as defined, Si
−ge−iωτc

is double (i.e., 6 dB above) Sci , as the latter describes a [P ] g
Ssi (ω) = = p ;
signal of amplitude 1 panned to center (i.e., split equally X 1 − g 2 e−2iωτc 4 2
g − 2g cos(2ωτc ) + 1
between L and R inputs), while the former describes two
1 g

1
[P ]
signals of amplitude 1 fed in phase to the two inputs of Sci (ω) = 1− iωτ
= p ;
2 g+e c 2
2 g + 2g cos(ωτc ) + 1
the system.
Since a real signal can consist of various components [P ]
g g
Ŝ (ω) = max 1 − , 1 + ,
having different phase relationships, it is more useful g + eiωτc eiωτc − g
1 spectra, we find the following amplitudes and frequencies

= max p , for the associated peaks and minima, denoted by ↑ and
g2 + 2g cos(ωτc ) + 1
! ↓ superscripts, respectively:
1
p ; (16)
g 2 − 2g cos(ωτc ) + 1 [P ]↑ 1
Ssi = at ωτc = nπ, with n = 0, 1, 2, 3, 4, . . .
χ [P ]
(ω) = ∞. (17) k 1 − g2
[P ]↓ 1 π
Therefore the perfect (χ = ∞) XTC filter gives flat fre- Ssi = at ωτc = n , with n = 1, 3, 5, 7, . . .
k 1 + g2 2
quency responses at the ears (E [P ] (ω) = constant), but [P ]↑ g
not at the sources. To appreciate the extent of spec- Ssi = at ωτc = nπ, with n = 0, 1, 2, 3, 4, . . .
X 1 − g2
tral coloration at the loudspeakers, we plot the S [P ] (ω) g π
[P ]↓
frequency responses expressed above in Fig. 2 for a typ- Ssi = 2
at ωτc = n , with n = 1, 3, 5, 7, . . .
ical value g = .985. Throughout this paper, for the sake
X 1+g 2
of illustration, we complement the non-dimensional plots [P ]↑ 1
Sci = at ωτc = nπ, with n = 1, 3, 5, 7, . . .
with dimensional calculations, which are represented by 2 − 2g
the same curves read in terms of the frequency f = ω/2π [P ]↓ 1
Sci = at ωτc = nπ, with n = 0, 2, 4, 6, . . .
on the top axis, for a typical listening geometry charac- 2 + 2g
terized by g = .985 and τc = 68 µs (i.e., 3 samples at the ↑ 1
Red Book CD sampling rate of 44.1kHz), which would Ŝ [P ] = at ωτc = nπ, with n = 0, 1, 2, 3, 4, . . .
1−g
be the case, for instance, of a set-up with ∆r = 15 cm,
(18)
l = 1.6 m, and Θ = 18◦ .)
↓ 1 π
Ŝ [P ] = at ωτc = n , with n = 1, 3, 5, 7, . . .
f HHzL
p
1 + g2 2
100 1000 104 2x104 (19)
40
For a typical listening set-up, g ' 1, say, our reference
30 ` ApE
S g = .985 case shown in Fig. 2, the envelope peaks (i.e.,
↑
20
Ŝ [P ] ) correspond to a boost of
ApE
Ssi 1
dB
10 20 log10 = 36.5 dB
1 − .985
0
[P ]↑ [P ]↑
ApE
(and the peaks in the other spectra, Ssi ' Ssi '
k X
Sci
-10 [P ]↑
Sci , correspond to boosts of about 30.5 dB.) While
-20 these boosts have equal frequency widths across the spec-
Π100 Π10 Π2 Π 2Π 3Π
trum, when the spectrum is plotted logarithmically (as
Ω Τc
is appropriate for human sound perception), the low-
frequency boost is most prominent in its perceived fre-
FIG. 2: Perfect XTC filter frequency responses at the loud- quency extent. This bass boost has long been recog-
speakers: amplitude envelope (heavy curve), side image (light nized as an intrinsic problem in XTC. While the high-
solid curve), and central image (light dashed curve). The dot-
frequency peaks could, in principle, be pushed out of the
ted horizontal line marks the envelope ceiling, which for this
case (g = .985) is 36.5 dB. The non-dimensional frequency ωτc audio range by decreasing τc (which, as can be seen from
is given on the bottom axis, and the corresponding frequency Eqs. (4) to (6), is achieved by increasing l and/or decreas-
in Hz, shown on the top axis, is to illustrate a particular (typ- ing the loudspeaker span Θ, as is done in the so-called
ical) case of τc =3 samples at a sampling rate of 44.1 kHz. “Stereo Dipole” configuration described in Ref. [19, 20],
[P ] [P ]
(Since Ssi ' Ssi when g ' 1, these two spectra are shown where Θ = 10◦ ), the “low frequency boost” of the P-XTC
k X
[P ]
filter would remain problematic.
as the single curve Ssi .) As mentioned in Section I B 1, the severe spectral
coloration associated with these high-amplitude peaks
The peaks in these spectra occur at frequencies for presents three practical problems: 1) it would be heard
which the system must boost the amplitude of the sig- by a listener outside the sweet spot, 2) it would cause a
nal at the loudspeakers in order to effect XTC at the relative increase (compared to unprocessed sound play-
ears while compensating for the destructive interference back) in the physical strain on the playback transducers,
at that location. Similarly, minima in the spectra occur and 3) it would correspond to a loss in the dynamic range.
when the amplitude must be attenuated. These penalties might be a justifiable price to pay
Using the first and second derivatives (with respect if we are guaranteed the infinitely good XTC perfor-
to ωτc ) of the above expressions for the various S [P ] (ω) mance (χ = ∞) and the perfectly flat frequency re-
sponse (E [P ] (ω) = constant) that the perfect XTC fil- the loudspeaker span to be a function of the frequency.
ter promises at the ears of a listener in the sweet spot. More specifically, after noting that typically l >> ∆r, so
However, in practice, these theoretically promised bene- that the approximation ∆l ' ∆r sin(θ) holds, and there-
fits are unachievable due to the solution’s sensitivity to fore ωτc = ω∆l/cs = 2πf ∆l/cs can be approximated by
unavoidable errors. This problem can best be appreci-
ated by evaluating the condition number of the transfer 2πf ∆r sin(θ)
ωτc ' for l ∆r, (21)
matrix C. cs
It is well known that in matrix inversion problems the
we can re-write the robustness condition (stated in
sensitivity of the solution to errors in the system is given
Eq. (20)) as
by the condition number of the matrix. (For a discussion
of the condition number in the context of XTC system
ncs

errors, see Ref. [30]). The condition number κ(C) of the Θ(f ) ' 2 sin−1 , with n = 1, 3, 5, 7, . . .
4f ∆r
matrix C is given by
Since both cs and ∆r are constant, the required loud-
κ(C) = ||C|| ||C −1 || = ||C|| ||H [P ] ||. speaker span is solely a function of the frequency f . In
(It is also, equivalently, the ratio of largest to smallest practice this prescription, called Optimal Source Distri-
singular values of the matrix.) Therefore, we have bution (OSD), can be implemented by using a crossover
s network to distribute adjacent bands of the audio spec-
2(g 2 + 1) trum to pairs of transducers, whose spans are calculated
κ(C) = max 2
− 1, from the above equation so that in each band the condi-
g + 2g cos(ωτc ) + 1
s tion number does not exceed unity by much, thus insur-
ing robustness and low coloration over the entire audio
!
2(g 2 + 1)
−1 . spectrum. It is clear, however, that this solution is not
g 2 − 2g cos(ωτc ) + 1
applicable to the case of a single pair of loudspeakers,
Using the first and second derivatives of this function, which is the focus of our analysis.
as we did for the previous spectra, we find the following We refer the reader interested in the OSD method and
maxima and minima: XTC errors to Ref. [29, 30, 32], and sum up the discus-
sion in this section by stating that, for the case of only
1+g two loudspeakers, the perfect XTC filter carries in prac-
κ↑ (C) = at ωτc = nπ, with n = 0, 1, 2, 3, 4, . . .
1−g tice the penalties of over-amplification (and the associ-
π ated loss of dynamic range) at frequencies where system
κ↓ (C) = 1 at ωτc = n , with n = 1, 3, 5, 7, . . . (20)
2 inversion is ill-conditioned, transducer fatigue, and a se-
as was also reported in Ref. [30] in terms of wavelengths. vere spectral coloration that is heard by listeners inside
First, we note that the peaks and minima in the con- and outside the sweet spot.
dition number occur at the same frequencies as those
of the amplitude envelope spectrum at the loudspeak-
III. CONSTANT-PARAMETER
ers, Ŝ [P ] . Second, we note that the minima have a
REGULARIZATION
condition number of unity (the lowest possible value),
which implies that the filter resulting from the inver-
sion of C is most robust (i.e., least sensitive to errors in Regularization methods allow controlling the norm of
the transfer matrix) at the non-dimensional frequencies the approximate solution of an ill-conditioned linear sys-
ωτc = π/2, 3π/2, 5π/2, . . . . Conversely, the condition tem at the price of some loss in the accuracy of the solu-
number can reach very high values (e.g., κ↑ (C) = 132.3 tion. The control of the norm through regularization can
for our typical case of g = .985) at the non-dimensional be done subject to an optimization prescription, such as
frequencies ωτc = 0, π, 2π, 3π . . . . As g → 1 the ma- the minimization of a cost function. Ref. [36] provides a
trix inversion resulting in the P-XTC filter becomes ill- detailed discussion of regularization methods in a general
conditioned, or in other words, infinitely sensitive to er- mathematical context, and Refs. [3, 18, 23, 33, 34] are ex-
rors. The slightest misalignment, for instance, of the amples of the use of regularization to control numerical
listener’s head, would thus result in a severe loss in XTC HRTF inversion. We discuss regularization analytically
control at the ears (at and near these frequencies) which, in the context of XTC filter optimization, which we define
in turn, causes the severe spectral coloration in Ŝ [P ] (ω) as the maximization of XTC performance for a desired
to be transmitted to the ears. tolerable level of spectral coloration or, equivalently, the
We are now in a position to appreciate the pre- minimization of spectral coloration for a desired mini-
scription proposed and implemented by Takeuchi and mum XTC performance.
Nelson[29, 32], which effectively solves both the robust- In essence, a nearby solution to the matrix inversion
ness and spectral coloration problem of the P-XTC filter problem is sought:
by insuring that the system operates always under condi- h i−1
tions where κ(C) is small. This can be done by allowing H [β] = C H C + βI CH , (22)
p
where the superscript H denotes the Hermitian operator, [β] g 4 − 2(β + 1)g 2 cos(2ωτc ) + (β + 1)2
Ssi (ω) = ;
and β is the regularization parameter which essentially k −2g 2 cos(2ωτc ) + (g 2 + β)2 + 2β + 1
causes a departure from H [P ] , the exact inverse of C. In p
g (g 2 + β)2 − 2(g 2 + β) cos(2ωτc ) + 1
[β]
this section we take β to be a constant, 0 < β 1. The Ssi (ω) = ;
X −2g 2 cos(2ωτc ) + (g 2 + β)2 + 2β + 1
pseudoinverse matrix H [β] is the regularized filter, and p
the superscript [β] is used to denote constant-parameter [β] g 2 + 2g cos(ωτc ) + 1
Sci (ω) = ;
regularization. The regularization stated in Eq. (22) can 2
2[g + 2g cos(ωτc ) + β + 1]
be shown[23, 34, 37] to correspond to a minimization of p
a cost function, J(iω), [β] g 2 + 2g cos(ωτc ) + 1
Ŝ (ω) = max 2
,
g + 2g cos(ωτc ) + β + 1
J(iω) = eH (iω)e(iω) + βv H (iω)v(iω), (23) p !
where the vector e represents a performance metric that g 2 − 2g cos(ωτc ) + 1
; (27)
is a measure of the departure from the signal reproduced g 2 − 2g cos(ωτc ) + β + 1
by the perfect filter. Physically, then, the first term in the g 4 + βg 2 − 2g 2 cos(2ωτc ) + β + 1
sum constituting the cost function represents a measure χ[β] (ω) = . (28)
2gβ| cos(ωτc )|
of the performance error, and the second term represents
an “effort penalty,” which is a measure of the power ex-
Of course, as β → 0, H [β] → H [P ] , and it can be verified
erted by the loudspeakers. For β > 0, Eq. (22) leads
that the spectra of the perfect XTC filter are recovered
to an optimum, which corresponds to the least-square
from the expressions above.
minimization of the cost function J(iω).
The envelope spectrum, Ŝ [β] (ω), is plotted in Fig. 3
Therefore, an increase of the regularization parameter
for three values of β. Two features can be noted in that
β leads to a minimization of the effort penalty at the ex-
plot: 1) increasing the regularization parameter atten-
pense of a larger performance error and thus to an abate-
uates the peaks in the spectrum without affecting the
ment of the peaks in the norm of H, i.e., the coloration
minima, and 2) with increasing β the spectral maxima
peaks in the S(ω) spectra, at the price of a decrease in
split into doublet peaks (two closely-spaced peaks).
XTC performance at and near the frequencies where the
system is ill-conditioned. f HHzL
100 1000 104 2x104
40
A. Frequency Response
30 19.5 dB
Using the explicit form for C given by Eq. (12), in the Β=0 29.5 dB
last equation above, we find: 20
S HdBL
 [β] [β]  .005

HLL (iω) HLR (iω) 10
` AΒE
H [β] =  , (24)
[β] [β] 0 .05
HRL (iω) HRR (iω)
where -10
[β] [β]
HLL (iω) = HRR (iω) -20
Π100 Π10 Π2 Π 2Π 3Π
g 2 ei4ωτc − (β + 1)ei2ωτc
= 2 i4ωτc , Ω Τc
g e + g 2 − [(g 2 + β)2 + 2β + 1]
(25) FIG. 3: Effects of regularization on the envelope spectrum at
[β]
HLR (iω) =
[β]
HRL (iω) the loudspeakers, Ŝ [β] (ω), showing peak attenuation and for-
mation of doublet peaks as β is increased. (Other parameters
iωτc
ge − g(g 2 + β)ei3ωτc are the same as for Fig. 2.)
= .
g 2 ei4ωτc + − [(g 2 + β)2 + 2β + 1]
g2
(26) To get a measure of peak attenuation and the condi-
tions for the formation of doublet peaks, we take the first
The eight metric spectra we defined in Section II B be- and second derivatives of Ŝ [β] (ω) with respect to ωτc and
come: find the conditions for which the first derivative is nil and
[β] g 4 + βg 2 − 2g 2 cos(2ωτc ) + β + 1 the second is negative. These conditions are summarized
Esi (ω) = ;
k −2g 2 cos(2ωτc ) + (g 2 + β)2 + 2β + 1 as follows: If β is below a threshold β ∗ defined as
[β] 2gβ| cos(ωτc )|
Esi (ω) = ; β < β ∗ ≡ (g − 1)2 , (29)
X −2g cos(2ωτc ) + (g 2 + β)2 + 2β + 1
2
[β] 1 β the peaks are singlets and occur at the same non-
Eci (ω) = − ;
2 2 [g 2 + 2 cos(ωτc ) + β + 1] dimensional frequencies as for the envelope spectrum
↑
peaks of the P-XTC filter (Ŝ [P ] ), and have the following Since regularization is essentially a deliberate intro-
amplitude: duction of error into system inversion, we should expect
both the XTC spectrum and the frequency responses at
↑ 1−g the ears to suffer (i.e., depart from their ideal P-XTC
Ŝ [β] =
(g − 1)2 + β filter levels of ∞ and 0 dB, respectively) with increasing
at ωτc = nπ, with n = 0, 1, 2, 3, 4, . . . β. The effects of constant-parameter regularization on
responses at the ears are illustrated in Fig. 4.
If the condition
β∗ ≤ β 1 (30) f HHzL
100 1000 104 2x104
is satisfied, the maxima are doublet peaks located at the 40
following non-dimensional frequencies:
30
2
−1 g −β+1 .005
ωτc = nπ ± cos with n = 0, 1, 2, 3, 4, . . . 20
2g Β = .05
(31)
ΧAΒE
dB
10
and have an amplitude
0 .005
↑↑ 1
Ŝ [β] = √ , (32) .05
2 β
-10 AΒE
Esi
ÈÈ
which does not depend on g. (The superscripts ↑ and
-20
↑↑ denote singlet and doublet peaks, respectively.) The Π100 Π10 Π2 Π 2Π 3Π
attenuation of peaks in the Ŝ [β] spectrum due to regu- Ω Τc
larization can be obtained by dividing the amplitude of
the peaks in the P-XTC (i.e., β = 0) spectrum by that of FIG. 4: Effects of regularization on the the crosstalk cancel-
peaks in the regularized spectrum. For the case of singlet lation spectrum, χ[β] (ω) (top two curves), and the ipsilateral
[β]
peaks, the attenuation is frequency response at the ear for a side image, Esi (ω) (bot-
k
↑
! tom two curves). The black horizontal bars on the top axis
Ŝ [P ]

β mark the frequency ranges for which an XTC level of 20 dB
20 log10 = 20 log 10 dB,
Ŝ [β]↑ (g − 1)2 + 1 or higher is reached with β = .05, and the grey bars represent
the same for the case of β = .005. (Other parameters are the
same as for Fig. 2.)
and for doublet peaks, it is given by
↑
! √
Ŝ [P ] 2 β
20 log10 = 20 log10 dB. The black curves in that plot represent the crosstalk
Ŝ [β] ↑↑
1 −g
cancellation spectra and show that XTC control is lost
within frequency bands centered around the frequen-
For the typical case of g = .985 illustrated in Fig. 3, we
cies where the system is ill-conditioned (ωτc = nπ with
have β ∗ = 2.225 × 10−4 , and for β = .005 and 0.05 we
n = 0, 1, 2, 3, 4, . . . ) and whose frequency extent widens
get doublet peaks that are attenuated (with respect to
with increasing regularization. For example, increasing
the peaks in the P-XTC spectrum) by 19.5 and 29.5 dB,
β to .05 limits XTC of 20 dB or higher to the frequency
respectively, as marked on that plot.
ranges marked by black horizontal bars on the top axis
Therefore, increasing the regularization parameter
of that figure, with the first range extending only from
above this (typically low) threshold causes the maxima in
1.1 to 6.3 kHz and the second and third ranges located
the envelope spectrum to split into doublet peaks shifted
above 8.4 kHz. In many practical applications, such
by a frequency ∆(ωτc ) = cos−1 [(g 2 − β + 1)/2g] to ei-
high (20 dB) XTC levels may not be needed or achiev-
ther side of the peaks in the response of the perfect XTC
able (e.g., because of room reflections and/or HRTF mis-
filter. (For our illustrative case of g = .985, we have
match) and the higher values of β needed to tame the
β ∗ = 2.225 × 10−4 and ∆(ωτc ) ' 0.225 for β = .05). Due
spectral coloration peaks below a required level at the
to the logarithmic nature of frequency perception for hu-
loudspeakers may be tolerated.
mans, these doublet peaks are perceived as narrow-band
artifacts at high frequencies (i.e., for n = 1, 2, 3, . . . ), but [β]
The Esi (ω) responses at the ears, shown as the bot-
k
the first doublet peak centered at n = 0 is perceived as
tom curves in Fig. 4, depart only by a few dB from the
a wide-band low-frequency rolloff of typically many dB,
corresponding P-XTC (i.e., β = 0) filter response (which
as can be clearly seen in Fig. 3. Therefore, constant-β
is a flat curve at 0 dB). More precisely and generally, the
regularization transforms the bass boost of the perfect [β]
XTC filter into a bass roll-off. maxima and minima of the Esi (ω) spectrum are given
k
by: We can then re-write Eqs. (33) and (34) as
[β]↑ g2 + 1 π [β]
HLL (z) = HRR (z)
[β]
Esi = at ωτc = n , with n = 1, 3, 5, . . .
k g2 + β + 1 2
(β + 1)

[β]↓
4 2
g + (β − 2)g + β + 1 = z− ×
Esi = 4 g2
k g + 2(β − 1)g 2 + (β + 1)2
1

1

at ωτc = nπ, with n = 0, 1, 2, 3, 4, . . . ,(37)
1 − a1 z −1 z − a2
[β] [β]
For the typical (g = .985) example shown in the figure, we HLR (z) = HRL (z)
[β]↑ [β]↓
−1/2
− (g 2 + β)z 1/2

have, for β = .05, Esi = −.2 dB and Esi = −6.1 dB, z
k k = ×
showing that even relatively aggressive regularization re- g

sults in a spectral coloration at the ears that is quite 1 1
.(38)
modest compared to the spectral coloration the perfect 1 − a1 z −1 z − a2
XTC filter imposes at the loudspeakers.
In sum, we conclude that, while constant-parameter Since 0 < g < 1, and β ≥ 0, we see from Eqs. (35)
regularization is effective at reducing the amplitude of and (36) that 0 ≤ a1 < 1 and a2 > 1, and therefore
peaks (including the “low-frequency boost”) in the enve- |a1 z −1 | < 1 and a2 > |z|. This allows us to express the
lope spectrum at the loudspeakers, it typically results in terms 1/(1 − a1 z −1 ) and 1/(z − a2 ) in the last two equa-
undesirable narrow-band artifacts at higher frequencies tions as two convergent power series (whose convergence
and a rolloff of the lower frequencies at the loudspeakers. insures that we have a stable filter), and thus write the
This non-optimal behavior can be avoided if the regu- last two equations as
larization parameter is allowed to be a function of the
frequency, as we shall see in Section IV.
Before we do so, it is insightful to consider the effects [β] [β]
HLL (z) = HRR (z)
of constant-parameter regularization on the time-domain
(β + 1)
response of XTC filters. = z− ×
g2
∞ ∞
! !
X X
m −m
a1 z −a2−m−1 z m (39)
B. Impulse Response m=0 m=0
[β] [β]
HLR (z) = HRL (z)
We start by making the substitution z = ei2ωτc in −1/2
− (g 2 + β)z 1/2

z
Eqs. (25) and (26) to get = ×
g
∞
! ∞ !
[β] [β]
HLL (z) = HRR (z) X
m −m
X
−m−1 m
a1 z −a2 z (. 40)
z 2 g 2 − z(β + 1) m=0 n=0
= , (33)
z2 g2 + g2 − z [(g 2 + β)2 + 2β + 1]
[β] [β] The filter is now in a form that can be readily trans-
HLR (z) = HRL (z)
formed into a time-domain filter, h[β] , represented by
z gz −1/2 − g(g 2 + β)z 1/2

= 2 2 . (34)
z g + g 2 − z [(g 2 + β)2 + 2β + 1] [β] [β]
 
hLL (t) hLR (t)
h[β] =  . (41)
The two expressions above have the same quadratic de- [β] [β]
hRL (t) hRR (t)
nominator, which can be factored as
We do so by substituting back ei2ωτc for z in Eqs. (39)
z 2 g 2 + g 2 − z (g 2 + β)2 + 2β + 1 = g 2 (z − a1 )(z − a2 ),

and (40), and taking the inverse Fourier transform (IFT)
to get
where
Z ∞
[β] 1 [β]
HLL (iω)eiωt dω
p p
a− a2 − 4g 4 a+ a2 − 4g 4 hLL (t) =
a1 = , a2 = , (35) 2π −∞
2g 2 2g 2 Z ∞
[β] 1 [β]
= hRR (t) = H (iω)eiωt dω
and 2π −∞ RR

β+1
= δ(t + 2τc ) − δ(t) ∗ ψ(t), (42)
a = (g 2 + β)2 + 2β + 1. (36) g2
Z ∞
[β] 1 [β] 1.0 æ
æ
HLR (iω)eiωt dω
æ
hLR (t) = ææ
ææ
2π −∞
ææ
ææ
ææ
Z ∞ Β=0 ææ
ææ
ææ
[β] 1 [β] 0.5
ææ
ææ
H (iω)eiωt dω
ææ
= hRL (t) = ææ
æææ
2π −∞ RL
æææ
æ
δ(τc − t) (g 2 + β)δ(t + τc )

= − ∗ ψ(t), 0.0
g g
(43) ççç
çç
ççç
çç
-0.5 çç
çç
çç
ç
where the asterisk denotes the convolution operation, and çç
çç
çç
çç
çç
ψ(t) is the IFT of the product of the two series appear- ç
ç çç
ç
ç
ing in Eqs. (39) and (40), and is given by the following -1.0 ç
convolution of two trains of Dirac delta functions: -200 -100 0 100 200
1.0
∞
X ∞
X
ψ(t) = am
1 δ(t − 2mτc ) ∗ −a−m−1
2 δ(t + 2mτc ), Β = .005
æ
m=0 m=0 0.5 æ
æ
(44) ç
ç
ç æ
æ
æ
ç æ
We see that the first train evolves forward in time and ç çç
ç çç
ç æ
ææ
ææ
æææ
çç ææææ
ççççç
the second evolves in reverse time. 0.0 ç
æçæçæçæçæçæçæçæçæçæçæçæçæçæçæçæ æ
ææææ
æææ
ææ ççç
ççççç
ææææææææææ
æææææ
çççççççççççççç
ç
ææ çç
The impulse response (IR) represented by Eqs. (42) ææ
æ
æ ç
ç
çç
ç
and (43) is plotted in Fig. 5 for three values of β. æ
æ
ç
ç
The IR of the perfect XTC filter is shown in the top -0.5 ç

ç
panel of that figure and consists of two trains of decay-

ing and inter-delayed delta functions of opposite sign.
-1.0
Mathematically, it is the special case of β = 0, for which -200 -100 0 100 200
Eqs. (37) and (38) simplify to 1.0
[P ] 1
[P ] Β = .05
HLL (z) = HRR (z) = , (45)
1 − a1 z −1 0.5 æ
ç
[P ] [P ] gz −1/2 ç
æ
HLR (z) = HRL (z) = − , (46) ç æ
1 − a1 z −1 çç
ç
ç æ
æ
ææ
0.0 çæçæçæçæçæçæçæçæçæçæçæçæ
ææ
çç
ææææææ
ç ç ç ç ç ç çæçæçæçæçæçæç
æ ç
æ
from which, through the inverse Fourier transform, we re- æ
ç
ç
cover the IR of the perfect XTC filter derived in Ref. [20]: æ

ç
-0.5
∞
[P ] [P ]
X
hLL (t) = hRR (t) = an1 δ(t − 2nτc ) (47)
n=0 -1.0
[P ] [P ] -200 -100 0 100 200
hLR (t) = hRL (t) t HsamplesL
∞
X
= −gδ(t − τc ) ∗ an1 δ(t − 2nτc ), (48) FIG. 5: Impulse responses hLL (t) = hRR (t) (filled circles)
[β] [β]
n=0 [β] [β]

and hLR (t) = hRL (t) (empty circles) for three values of β.
where a1 = g 2 (obtained by setting β = 0 in Eqs. (35) (g = .985, τc = 3 samples.)
and (36)) is the pole of the filter. We see that the perfect
XTC IR starts at t = 0 with an amplitude of unity and
decays to an amplitude an1 = (l1 /l2 )2n after a time 2nτc . to the empty circles in the same plot). These two series of
Its physical significance has been discussed by Kirkeby pulses are delayed by τc with respect to each other so that
et. al.[20] who, along with Atal et. al.[9] before, recog- after the first positive pressure pulse arrives at the left
nized the recursive nature of XTC filters. Briefly, a phys- ear, then reaches the right ear with a slightly smaller am-
ical appreciation of the perfect crosstalk cancellation IR plitude, it is cancelled there by a negative pressure pulse
can be obtained by considering the hypothetical case of of equal amplitude (that was emitted a time τc earlier by
a positive pulse whose duration is much smaller than τc , the right loudspeaker), which in turn is cancelled at the
fed into only one of the two inputs of the system, say the left ear by a positive pressure pulse, and so on. The net
left input. From Eq. (9), we see that this pulse, dL (t), is result is that only the first pulse is heard and only at the
emitted from the left loudspeaker as a series of positive left ear, i.e., with no crosstalk.
pulses dL (t)∗hLL (t) (corresponding to the filled circles in The effects of regularization on the XTC IR can be
the top panel of Fig. 5) and from the right loudspeaker as gleaned from a comparison of the three panels of Fig. 5.
a series of negative pulses dL (t) ∗ hRL (t) (corresponding When β is finite, the IR has a “pre-echo” part, i.e., it
extends in reverse time (t < 0) as shown in Fig. 5. As spectrum, γ is bounded by the inequalities:
also can be seen in that figure, and inferred from Eq. (44),
the delta functions in the t < 0 and t > 0 parts have 1
1≤γ≤ , (52)
opposite signs. With increasing regularization, the t < 0 1−g
part increases in prominence and the IR becomes shorter ↑
in temporal extent, which correspond in the frequency where the last term is Ŝ [P ] , given by Eq. (18).
domain to a spectrum with abated peaks. The frequency-dependent regularization parameter
To insure causality, a time delay must be used to in- needed to effect the spectral flattening required
clude the t < 0 part of the IR. In practice (e.g., when by Eq. (49) is obtained by setting Ŝ [β] (ω), given by
dealing with numerical HRTF inversion), this can be Eq. (27), equal to γ and solving for β(ω), which is now a
done through a “modelling delay” that accommodates function of frequency. Since the regularized spectral en-
both the non-causal part of the IR and the transmission velope, Ŝ [β] (ω), (which is also ||H [β] ||, the 2-norm of the
delay regularized XTC filter) is the maximum of two functions,
we get two solutions for β(ω):
l1
δ −t p
cs g 2 − 2g cos(ωτc ) + 1
βI (ω) = −g 2 + 2g cos(ωτc ) + − 1,
γ
associated with the factor α in Eq. (8). (53)
The length of a filter having a pole close to the unit
circle, |z| = 1, is inversely proportional to the distance
p
g 2 + 2g cos(ωτc ) + 1
between the pole and the unit circle[38]. As β is in- βII (ω) = −g 2 − 2g cos(ωτc ) + − 1.
γ
creased the poles pull away from the unit circle as per (54)
Eqs. (35) and (36), and therefore the length of a finite-β The first solution, βI (ω), applies for frequency bands
IR is reduced by a factor of where the out-of-phase response of the perfect filter (i.e.,
1 − a1 the second singular value, which is the second argument
of the max function in Eq. (16)) dominates over the in-
1 − g2 phase response (i.e., the first argument of that function):
with respect to the length of the perfect XTC IR. This 1
factor (which is based on a1 since 1 − a1 < |1 − a2 | ) is So[P ] = p
accurate as long as 1 − g 2 << 1 and 1 − a1 << 1. For g2 − 2g cos(ωτc ) + 1
instance, for the IR shown in the middle panel of Fig. 5 [P ] 1
≥ Si =p . (55)
we have β = .005 and g = .985, which give a1 = .86 and 2
g + 2g cos(ωτc ) + 1
the IR is about 4.5 times shorter than the perfect XTC
IR. Similarly, regularization with βII (ω) applies for frequency
[P ] [P ]
bands where Si ≥ So . Therefore, we must distin-
guish between three branches of the optimized solution:
IV. FREQUENCY-DEPENDENT
two regularized branches corresponding to β = βI (ω)
REGULARIZATION
and β = βII (ω), and one non-regularized (perfect-filter)
branch corresponding to β = 0. We call these Branch I,
In order to avoid the frequency-domain artifacts dis- II and P, respectively, and sum up the conditions associ-
cussed in Section III A and illustrated in Fig. 3, we seek ated with each as follows:
an optimization prescription that would cause the enve-
[P ]
lope spectrum Ŝ(ω) to be flat at a desired level Γ (dB) Branch I: applies where Ŝ [P ] (ω) ≥ γ and So[P ] ≥ Si ,
over the frequency bands where the perfect filter’s enve-
and requires setting Ŝ(ω) = γ, β = βI (ω);
lope spectrum exceeds Γ (dB). Outside these bands (i.e.,
[P ] [P ]
below that level), we apply no regularization. This can Branch II: applies where Ŝ (ω) ≥ γ and Si ≥ So[P ] ,
be stated symbolically as:
and requires setting Ŝ(ω) = γ, β = βII (ω);
[P ] [P ]
Ŝ(ω) = γ if Ŝ (ω) ≥ γ, (49) Branch P: applies where Ŝ (ω) < γ,
Ŝ(ω) = Ŝ [P ]
(ω) if Ŝ [P ]
(ω) < γ, (50) and requires setting Ŝ(ω) = Ŝ [P ] (ω), β = 0.
where the P-XTC envelope spectrum, Ŝ [P ] (ω), is given Following this three-branch division, the envelope
by Eq. (16), and spectrum at the loudspeakers, Ŝ(ω), for the case of
frequency-dependent regularization is plotted as the
γ = 10Γ/20 , (51) thick black curve in Fig. 6 for Γ = 7 dB. This value
was chosen because it corresponds to the magnitude of
with Γ given in dB. We will take Γ ≥ 0 dB and, since Γ the (doublet)√ peaks in the β = .05 spectrum (i.e., Γ =
cannot exceed the magnitude of the peaks in the Ŝ [P ] (ω) 20 log10 (1/2 β)), which is also plotted (light solid curve)
as a reference for the corresponding case of constant- • Bands 2, 6, 10, 14, 18, . . . , 4n+2 belong to Branch P,
parameter regularization. (We call a spectrum ob- and are bounded by
tained with frequency-dependent regularization and one
obtained with constant-β regularization “corresponding 2nπ + φ ≤ ωτc ≤ (2n + 1)π − φ; (57)
spectra,” if the peaks in Ŝ [β] (ω), whether singlets or dou-
blets, are equal to γ.) It is clear from that figure that • Bands 3, 7, 11, 15, 19, . . . , 4n + 3 belong to Branch
II, and are bounded by
f HHzL
100 1000 104 2x104 (2n + 1)π − φ ≤ ωτc ≤ (2n + 1)π + φ; (58)
40 Band 1 2 3 4 5 6 7
30 • Bands 4, 8, 12, 16, 20, . . . , 4n+4 belong to Branch P,

Β=0
and are bounded by
20
(2n + 1)π + φ ≤ ωτc ≤ 2(2n + 1)π − φ; (59)
S HdBL
10 Β = Β HΩ, G = 7 dBL
`
where n = 0, 1, 2, 3, 4, . . . and
0
Β = .05 2 2
g γ + γ2 − 1

−1
-10 φ = cos . (60)
Branch I P II P I P II
2gγ 2
-20
Π100 Π10 Π2 Π 2Π 3Π For instance, applying this hierarchy to the case of g =
Ω Τc .985, and Γ = 7 dB (i.e., γ = 107/20 = 2.24), shown
in Fig. 6, we have the following set of eight consecutive
FIG. 6: Envelope spectrum at the loudspeakers, Ŝ(ω), for the frequency bounds for the seven consecutive bands be-
case of frequency-dependent regularization with Γ = 7 dB tween ωτc = 0 and 3π: {0, 0.45, 2.69, 3.59, 5.83, 6.74,
(thick black curve) and for the corresponding reference case 8.97, 9.42}, which correspond to dimensional frequen-
of β = .05 (grey curve). The benchmark case of the perfect cies, f (Hz) (with τs = 3 samples at 44.1 kHz) given
XTC filter is also shown (dashed grey curve). The vertical by the set: {0, 1061.5, 6288.5, 8411.5, 13638.5, 15761.5,
dotted lines show the frequency bounds of the resulting seven
20988.5, 22000}, as marked by the vertical lines in Fig. 6.
bands, which are numbered consecutively at the top of the
plot, and labeled with the corresponding branch name at the Bands 1 and 5 belong to Branch I and are regularized
bottom. (Other parameters are the same as for Fig. 2.) with β = βI (ω); Bands 3 and 7 belong to Branch II and
are regularized with β = βII (ω); and Bands 2, 4, and 6
belong to Branch P and are not regularized. In general,
the low-frequency boost and the high-frequency peaks of
successive bands, starting from the lowest-frequency one,
the perfect XTC spectrum, which would be transformed
are mapped to the following succession of branches: I, P,
into a low-frequency roll-off and narrow-band artifacts,
II, P, I, P, II, P, . . .
respectively, by constant-β regularization, are now flat
at the desired maximum coloration level, Γ. The rest of
the spectrum, i.e., the frequency bands with amplitude
B. Frequency Response
below Γ, is allowed to benefit from the infinite XTC level
of the perfect XTC filter and the robustness associated
with relatively low condition numbers. The amplitude envelope of the frequency response at
the loudspeakers, given by Eqs. (49) and (50), was al-
ready shown in Fig. 6. The other optimized metric spec-
A. Band Hierarchy tra can be derived as follows:
[O]
The three-branch prescription therefore splits the au- YI (ω) = Y [βI (ω)] (ω), for Branch-I bands; (61)
[O] [βII (ω)]
dio spectrum into a series of adjacent frequency bands, YII (ω) = Y (ω), for Branch-II bands; (62)
which we number consecutively starting with Band 1 for [O] [P ]
the lowest-frequency band. The frequency bounds for YP (ω) = Y (ω), for Branch-P bands; (63)
each band can be found by setting Ŝ [P ] (ω), given by
where Y represents any of the eight metric spectra we
Eq. (16), to γ and solving for ωτc . This results in the fol-
defined in Section II B, the superscript [O] denotes the
lowing hierarchy of bands and their associated frequency
sought optimized version of that metric spectrum, the
bounds:
subscript I, II, or P denotes one of the three branches, and
• Bands 1, 5, 9, 13, 17, . . . , 4n + 1 belong to Branch I, the superscripts [βI (ω)] and [βII (ω)] denote regularization
and are bounded by following the formulas for the regularized metric spec-
tra in Section III A, but with β taken to be frequency-
2nπ − φ ≤ ωτc ≤ 2nπ + φ; (56) dependent according to Eqs. (53) and (54).
For example, following the above hierarchical prescrip- is the XTC enhancement over that attained with the cor-
tion, and using Eqs. (28), (53), (54), and (17), the opti- responding constant-β regularization.
mized crosstalk cancellation spectrum becomes Furthermore, this XTC enhancement occurs with no
p relative penalty to the frequency response at the ears,
∓γx2 + (g 2 + 1) xγ ± g 2 ∓ x + 1 as can be seen by comparing the Esik (ω) spectrum with
[O]
χI,II (ω) = ∓ p ,
frequency-dependedent regularization (solid grey curve)
|x| γg 2 ∓ γx + γ − g 2 ∓ x + 1
to that with β = .05 (dashed grey curve) in the same
(64) figure.
[O] [p] It can be verified through Eqs. (28) and (64) that
χP (ω) = χ (ω) = ∞, (65)
constant-β regularization yields an XTC level that is
where, for compactness, we have used the definition equal to that obtained with the corresponding frequency-
x ≡ 2g cos(ωτc ) and combined both branches into one ex- dependent regularization only at the discrete frequencies
pression using the double subscripts “I,II” and the double at which to the peaks in the corresponding Ŝ [β] (ω) spec-
sign (± or ∓) with the top and bottom signs associated trum are located, i.e., at
with Branches I and II, respectively. Similarly, the opti-
mized version of the ipsilateral frequency response at the 1
ωτc = nπ, if < (g − 1)2 ;
ear for a side image, Esik (ω), becomes 4γ 2
2
g −β+1
[O] = nπ ± cos−1 ,
Esi (ω) = 2g
k I,II
1
if (g − 1)2 ≤ 2 1,
p
±xγ 2 (g 2 ∓ x + 1) + (γ 2 g + γ) g 2 ∓ x + 1 4γ
p (66)
g 2 ∓ x ± 2γx g 2 ∓ x + 1 + 1 with n = 0, 1, 2, 3, 4, . . . (68)
[O] [P ]
Esi (ω) = Esi (ω) = 1. (67)
kP k (where the inequalities are those conditioning singlet or
These spectra are plotted in Fig. 7 where it is im- doublet peaks in the corresponding Ŝ [β] (ω) spectrum,
and are derived from Eqs. (29), (30) and (32)). At
f HHzL all other frequencies, frequency-dependent regularization
100 1000 104 2x104
yields superior XTC performance to that obtained with
40 constant-β regularization. This behavior, which can also
be seen graphically in the χ(ω) curves of Fig. 7, is due to
30 Β = Β HΩ, G = 7 dBL the fact that forcing the envelope spectrum to be flat (in
Β = .05
bands belonging to Branches I and II) through frequency-
20 dependent regularization clamps the effort penalty term
in the cost function (second term in the sum in Eq. (23))
dB
10 leading to a minimization of the performance error. This

Χ in turn leads to a maximization of XTC level, which ex-
0 ceeds the corresponding constant-β XTC level at all fre-
EsiÈÈ
quencies (except at those given by Eq. (68), where both
-10
corresponding Ŝ spectra reach the same value, γ), since
-20 the corresponding constant-β envelope, Ŝ [β] (ω), is lower
Π100 Π10 Π2 Π 2Π 3Π than (or equal to) γ (as seen in Fig. 6).
Ω Τc Therefore, we conclude that if we define XTC filter
optimization as “the maximization of XTC performance
FIG. 7: Crosstalk cancellation spectrum, χ(ω) (black curves), for a desired tolerable level of spectral coloration” as
and ipsilateral frequency response at the ear for a side image, we did earlier, only frequency-dependent regularization
Esi (ω) (light curves), for the cases of frequency-dependent
k leads to an optimal XTC filter over all frequencies, while
regularization (solid curves) and β = .05 (dashed curves). constant-β regularization leads to an XTC filter that
The frequency ranges for which an XTC level of 20 dB or is optimized only at the discrete frequencies given by
higher is reached are marked on the top axis by black hori-
Eq. (68).
zontal bars for the case of β = β(ω) with Γ = 7 dB. (Other
parameters are the same as for Fig. 2.)
mediately clear from the χ(ω) curves that frequency- C. Impulse Response: The Band-Assembled
dependent regularization yields a significant enhance- Crosstalk Cancellation Hierarchy Filter
ment of XTC level over that obtained with constant-β
regularization. We can also deduce from this plot that In the frequency domain, the optimized XTC filter is
the higher the desired minimum level of XTC, the larger given by the following matrix:
the following two IRs:

 [O] [O] 
HLL (iω) HLR (iω) [O] [O]
[O]
hLL (t) = hRR (t) = (ψ0 + γψ1 ) ∗ ψa , (74)
I,II I,II
H = , (69)
[O] [O] [O] [O]
HRL (iω) HRR (iω) hLR (t) = hRL (t) = [∓ψ0 + γgδ(t + τc ) ∗ ψ1 ] ∗ ψa ,
I,II I,II
(75)
whose elements are derived following the same hierarchi-
cal prescription (i.e., Eqs. (61)-(63)) we used to get the
where
optimized metric spectra, namely by substituting β1 (ω)
and β1 (ω) from Eqs. (53) and (54), and β = 0 into each
ψa = ±ψ2 ∗ ψ3 ± (ψ1 ∓ ψ4 ) ∗ ψ5 ψ6 (c1 ) ∗ ψ6 (c2 ),
of Eqs. (25) and (26), to get the Branch I, Branch II and
Branch P versions of the filter’s matrix elements. This (76)
leads to 2 2 2 2
ψ0 = −g γ δ(t) ± gγ δ(τc − t) ± gγ δ(t + τc )
[O] [O] −g 2 γ 2 δ(t + 2τc ), (77)
HLL (iω) = HRR (iω) ∞
I,II I,II 1

X 1 −m
p
γ 2 ±x − g 2 1 + e2iωτc + γ g 2 ∓ x + 1 ψ1 = 2 (∓g)m g 2 + 1 2 ×
m
= p , m=0
g 2 ± x 2γ g 2 ∓ x + 1 − 1 + 1 m
X m
δ(2kτc − t − mτc ), (78)
(70) k
k=0
[O] [O] ∞
HLR (iω) = HRL (iω) 1 X − 21

I,II I,II
ψ2 = ± (−1)m ×
4gγ m=0 m
p
∓γ 2 ±x − g 2 1 + e2iωτc + gγeiωτc g 2 ∓ x + 1
= p , 2m
g 2 ± x 2γ g 2 ∓ x + 1 − 1 + 1 X 2m
(−1)k+m 4−m δ(t + 2kτc − 2mτc ),
k
(71) k=0
[O]
HLL (iω) =
[O]
HLL (iω) =
[P ]
HLL (iω) =
[P ]
HRR (iω), (72) (79)
P P ∞
− 21

[O] [O] [P ] [P ]
X − 1 −m
HLR (iω) = HRL (iω) = HLR (iω) = HRL (iω), (73) ψ3 = (∓g)m g 2 + 1 2 ×
P P m
m=0
m
where, again,

X m
δ(2kτc − t − mτc ), (80)
k
x ≡ 2g cos(ωτc ), k=0
ψ4 = 2gγδ(τc − t) + 2gγδ(t + τc ), (81)
and we have followed the same subscript and sign con- ∞ 3

1 X −2
ventions used to compact the XTC spectrum in Eq. (64). ψ5 = ± (−1)m ×
(4gγ)3 m=0 m
Eqs. (72) and (73) give the Branch-P elements of the ma-
trix of the optimized filter, which are also the elements 2m
X 2m

of the perfect filter’s matrix given by Eqs. (45) and (46), (−1)k+m 4−m δ(t + 2kτc − 2mτc ),
k
whose inverse Fourier transforms had given us the IRs k=0
expressed in Eqs. (47) and (48). Therefore we need to (82)
derive only the IRs associated with Branches I and II of ∞ ∞
p X
− p2

X ±c
the optimized filter. ψ6 (c) = (−1)n ×
2g n
To do so, we follow, albeit through more cumber- p=0 n=0
some algebra, the same approach we used to obtain the 2n
X 2m
constant-β IRs in Section III B; namely, we seek to factor (−1)k+m 4−m δ(t + 2kτc − 2mτc ).
the frequency-domain response of the filter into a prod- k
k=0
uct of terms, whose IFT can be readily found, or which (83)
can be expressed as convergent series of functions whose
IFT can be readily found. The complete IR is then the with the constants c1 and c2 given by
convolution of the IFTs of all the terms in the factored
frequency-domain response of the filter. The challenge p
16γ 2 (g 2 + 1) + 1 ∓ 1
is to carry out the factorization in such a way that all c1 = , (84)
the invoked power series expansions converge over the 8γ 2
p
parameter space of interest. − 16γ 2 (g 2 + 1) + 1 ∓ 1
c2 = . (85)
The derivation is carried out in Appendix A, where we 8γ 2
also discuss the convergence of the adopted series expan-
sions. The resulting filter in the time domain is given by The impulse responses are valid for values of γ and g that
[O]
satisfy the condition: flat envelope spectrum, ŜI (ω) = γ. The agreement is
p √ ! excellent only in the bands belonging to the branch for
5+ 5 1 which the IR is intended (which, in the case illustrated in
max ,1 ≤ γ ≤ , (86)
that plot, are the first and fifth bands). In other bands,
p
2 g2 + 1 1−g
not only is the IR not valid, but, as discussed in the
which is shown graphically as a region plot in Fig. 9, in appendices, its application may lead to singularities as-
Appendix A. sociated with the divergence of some of the series that
constitute it (see for instance the singularities appearing
1.0 in the Branch-P bands in Fig. 10).
Branch I
Therefore, in principle, the application of the op-
æ timal filter requires that, prior to XTC filtering, the
0.5 ç
ç recorded signal, [dLi (t), dRi (t)]T , be passed through a
çææ
æ
ç
æ æ crossover filter whose crossover frequencies are set to
ç
0.0æç
ç æç
æç æç
æç æç
æç æç
æç æç
æç æç
æç æç
æç
çç ççç
æææçæ ç
çææ
ææç ç ç
ç
ç ç
æ
æææ
ææææææ
çç
æææææææææææææææææææææ
ççççççççççççççççççççççççç
the band bounds given by the hierarchical prescription
æ ç
ææ
æ
ç
çç
ç
in Eqs. (56)-(59). The resulting bands are then as-
æ
æ
sembled into three groups (I, II and P) according to
-0.5 æ
their branch identity. The combined recorded stereo sig-
ç
nals in each group can thus be represented by a vector
[dLi (t), dRi (t)]T , where the index i stands for Branch I,
-1.0 II or P. The loudspeakers source vector, in the time do-
-100 -50 0 50 100 main, needed for optimal crosstalk cancellation is then
1.0
given by the time-domain version of Eq. (9):
Branch II   
  [O] [O] 
0.5 vL (t) h
X  LLi (t) h LRi (t) dL (t)
  i
æ
æ
ç  =  ∗ ,

vR (t) [O] [O] dRi (t)
æ çæ
æç
ç
çæ
æ
ç
ç æ æ
ç
æ ç
i h
RLi (t) hRRi (t)
æ çç æ ç ç æææ ç ç æ ç
0.0æçæç
æçæç
æçæç
æçæç
æçæç
æçæç
æç æç
æç æ ç
ææ æç
ææçæç
ææç ç æç æçç ç
ææç æ ææ ç ççæ æçæç ç
ææç æ çæç ç ç
çæç
ææ æçç ç
æ çæ çææç
æçæç
æçæç
æçæç
ææçç
æç
ç æ ç æçæ çç ç ææ ç ç æ ç æ æ ççæ æ æç ç æ æç æ
æ ç
ç
æ æ
ç
æ ç
ç
æ
æ
ç
ç æ ç æ ç æ ç
æ ç
(87)
æ ç
æ
ç where the summation is over the three branches, and
-0.5
the convolution operates in the same fashion as matrix
ç
æ
multiplication.
æ
ç Causality is insured by calculating the IRs with a “pre-
-1.0 delay,” starting back at a time t < 0, whose exact tem-
-100 -50 0 50 100 poral extent is not important as long as it allows the in-
t HsamplesL clusion of the salient part of the IR. For the IRs in Fig. 8,
this pre-delay should start at about t = −100 samples.
[O]
FIG. 8: Impulse response of the optimal XTC filter: hLL (t) =
[O] [O] [O]
hRR (t) (filled circles) and hLR (t) = hRL (t) (empty circles),
for Branch I (top panel) and Branch II (bottom panel). (Γ = V. APPLICATION AND PRACTICAL
7 dB, g = .985, and τc = 3 samples, as in Fig. 2.) CONSIDERATIONS
The impulse responses for Branch-I and Branch-II of While the first goal of the preceding analyses was
this optimal filter are shown in Fig. 8 for our typical to provide insight into the theory and fundamentals of
case of g = .985 and τc = 3 samples, and, along with XTC optimization, the resulting optimized IR can offer
the perfect filter IR shown in the top panel of Fig. 5, practical benefits in audio applications where the spa-
completely specify the optimal XTC filter. tial fidelity of a recording is degraded by the unintended
Compared to the corresponding (β = .05) finite-beta crosstalk inherent in playback through loudspeakers. As
IR in the bottom panel of Fig. 5, the optimal XTC IRs we argued in Section I A, this is the case of not only
shown in Fig. 8 are more complex in their structure. Fur- pure binaural recordings made with dummy head micro-
thermore, each IR consists of a train of deltas that are phones, but also of the vast majority of standard stereo
spaced by τc as opposed to the 2τc intervals we had for recordings, since they generally contain ILD and ITD
the perfect and finite-beta filters. cues, which would be degraded by unintended crosstalk.
These IRs are difficult to interpret physically because,
as they stand, they also include the time response asso-
ciated, in the frequency domain, with frequency bands A. The Value of Analytical XTC Filters
where the IR is not valid. This is illustrated in Ap-
pendix B, in the bottom panel of Fig. 10, where the Analytical XTC filters cannot rival the performance of
envelope spectrum obtained from the Fourier transform numerical HRTF-based XTC filters in ideal situations,
of the Branch-I optimal IR is compared to the expected i.e., when 1) sound reflections in the listening room are
negligible or non-existent (anechoic or semi-anechoic en- listening and audio mastering applications, it may not
vironments), 2) the recording was made with the indi- be desirable to have Γ exceed 3-5 dB, while for home-
vidualized inverted HRTF of the listener, 3) the XTC theatre applications, audio (spectral) fidelity may be in-
filter includes the individualized inverted HRTF of the tentionally compromised with higher values of Γ for the
listener, and 4) the listener’s head is constrained in a re- advantage of having more XTC headroom for reproduc-
stricted sweet spot. Any departure from these idealities ing surround effects with the two loudspeakers.
would cause the effective XTC level at the listener’s ears The choice of loudspeaker span is particularly impor-
to drop, and the spectral coloration that is necessarily tant. In cases where it is constrained to a set value, as
imposed at the loudspeakers to become more audible at for compatibility with the so-called “standard stereo tri-
the listener’s ears. angle”, i.e., Θ = 60◦ , the value of Θ becomes a fixed
Since in many, if not most, practical listening situa- input to the design process and is used, along with l, to
tions in non-anechoic environments all of the four ide- calculate g and τc from Eqs. (3)-(6). (The inequality in
alities listed above are compromised to a certain degree, Eq. (86), which is typically easy to satisfy, must hold for
the practically achievable XTC level of numerical HRTF- that particular combination of γ = 10Γ/20 and g. If not,
based XTC filters seldom exceeds 13 dB over a wide fre- one of the input parameters, usually Γ, must be adjusted
quency range[3]. An optimal analytical XTC filter, even accordingly before proceeding further with the design).
one based on a free-field model, such as the one derived in In cases where Θ is not constrained to a preset value,
the previous section, can become competitive especially it becomes a useful variable in the filter design process
in situations where it is calculated for, and used with, and can be used to simplify the filter, as discussed in
a loudspeaker span that is small enough to diminish the Section V C below.
relative importance of head-shadowing effects. In such With γ, g, and τc specified, one has all the parame-
applications, an optimal analytical XTC filter can offer ters needed to calculate the spectra associated with the
the following advantages over a numerical HRTF-based XTC optimal filter, as described in Sections IV A and
XTC filter: IV B, and thus evaluate the various aspects of the fil-
1. The simplicity of using a single filter for all indi- ter. (These evaluations are more conveniently done in
viduals. terms of the dimensional frequency, f , in Hz, by select-
ing the intended sampling rate.) In particular, a plot
2. Shorter filters which incur lower CPU loads on the of the XTC spectrum according to Eqs. (64) and (65)
digital processor. allows the evaluation of the XTC performance of the fil-
ter (defined as the frequency extent over which a desired
3. Low spectral coloration for listeners inside and out-
minimum XTC level is reached or exceeded) which, by
side the sweet spot (and the associated decrease in
virtue of the implicit optimization (i.e., minimization of
the physical loading of the transducers).
the cost function in Eq. (23)), is the maximum achievable
(The third advantage could, in principle, be neutralized XTC performance for that particular set of input param-
by applying the optimal regularization method, described eters. If the calculated XTC performance is judged by
in the previous section, to the design of a numerical some empirical standards to be above that achievable in
HRTF-based XTC – at the price of eroding some of the the intended listening environment (for instance, sound
advantages associated with the use of an individualized reflections in a reverberant room may limit the achiev-
HRTF.) able XTC to only a few dB over a good part of the audio
With this justification for the usefulness of spectrum), the calculation can be repeated with a lower
analytically-derived optimal XTC filters, we turn value of Γ, thus leading to even higher spectral fidelity.
our attention to some practical issues related to their Conversely, a lower than desired XTC performance can
specific design and their application to real listening be amended by raising Γ.
situations. Once the target XTC performance and coloration level
are reached, one proceeds to the time domain by calcu-
lating the Branch-P IRs from Eqs. (42)-(44), and the
B. Filter Design Strategy Branch-I, and II IRs from Eqs. (74)-(85). The loud-
speakers source vector can then be calculated according
Of course, filter design strategies depend on perfor- to Eq. (87), following the prescription given in the text
mance requirements (desired maximum tolerable col- preceding that equation, i.e., by appropriately convolv-
oration level or minimum XTC level) and the specifics ing the 3-part IR with the recorded stereo signal after
and constraints of the listening configuration (constraints having passed the latter through a crossover filter whose
on the listening distance, l, and the loudspeaker span, Θ, crossover frequencies are set to the band bounds given
and to some extent, the sound reflection characteristics in Eqs. (56)-(59). The convolution operations can be
of the listening room). carried out digitally, and in real time if desired, using
One approach to filter design is to start with the spec- a digital convolution plugin. (Such software plugins rely
ification of the maximum tolerable coloration level, that on FFT-based algorithms[39, 40] for fast convolution and
is, Γ in dB. For instance, for critical (e.g., audiophile) have become readily available in the commercial and pub-
lic domains for use as IR-based reverb processors.) It is relevant to mention in the context of loudspeaker
span that keeping Θ small offers advantages that have
been recognized since Kirkeby and co-workers presented
C. Simplified Implementation their analysis[20] of the “stereo dipole” configuration,
which has a span of only 10◦ . Objective and subjective
An XTC system consisting of the properly configured evaluations of the effects of loudspeaker span in XTC
crossover filter, the three XTC IR matrices, and the mul- systems have indicated that such a low-Θ configuration
tiple instances of convolution plugins, can be considered gives a larger sweet spot than that obtained with larger
as a single filter, having stereo inputs and outputs, which loudspeaker spans[18]. This can be attributed to the rel-
acts as a linear operator. Therefore, once assembled, it ative insensitivity of the path length difference, ∆l, to
can be “rung” once by a single delta impulse, applied head movements when the span is small. On the other
to one of its two inputs, and the recorded stereo output hand, the same study favored larger spans partly because
would then represent one of the two columns of the 2x2 increasing the span, with the distance l fixed, lowers the
IR matrix of the entire filter. Because of symmetry, the value of g and consequently decreases the magnitude of
other column of the IR matrix is obtained by simply flip- the coloration peaks and condition numbers. We should
ping the two recorded outputs. This results in a single however expect, in light of our study of regularization,
IR, representing the entire three-branch multi-band fil- that an optimal XTC filter in which regularization is used
ter, and simplifies any future application of Eq. (87) to to flatten these peaks and lower the condition numbers,
a simpler one (with no crossover filtering) in which the while maintaining good XTC performance, should tip the
summation and indices are foregone. balance in favor of lower values of Θ. This remains to be
verified experimentally.
Another argument in favor of small loudspeaker spans
D. The Role of Loudspeaker Span is particular to the use of analytical filters based on a
free-field model, such as those discussed in this paper.
Another important simplification arises in applications Since the free-field model ignores the presence of the lis-
where the loudspeaker span, Θ = 2θ, is not constrained tener’s head, it should be expected that filters based on
to a preset value, such at the 60◦ of the standard stereo it perform better when the effects of head shadowing are
triangle, and therefore can be a variable in the filter de- minimized. This can be achieved by decreasing the span
sign process. Since τc depends on the loudspeaker span, angle as can be seen, for instance, in Fig. 3.13 of Ref. [27],
the bounds of the bands can be moved by varying θ. By where the inter-aural transfer function (the ratio of the
setting θ equal to a particular value, θ∗ , the higher bound frequency responses at the two ears) of a typical human
of the second band (which belongs to Branch P) can be head, measured as a function of the azimuthal position
made to coincide with a cutoff frequency, fc , above which of a sound source, is small (about -2dB) and flat (within
XTC is psychoacoustically not needed. Such a band- 2 dB) for a small horizontal source azimuth (θ = 5◦ ), but
limited optimal XTC filter has the advantage that it re- worsens with increasing azimuths.
quires only a 2-band crossover filter, and its IR consists
of only the Branch-I and Branch-P parts, thus leading to
significant simplifications in the design and implementa- E. An Example
tion of the filter.
To find an expression for θ∗ as a function of fc , under To illustrate the above design guidelines and discus-
the typically valid approximations g ' 1 and l ∆r, sions, we give the example of a listening situation whose
we set ωτc equal to the upper bound of the second band only two design requirements are a distance l = 1.6 m
(which, from Eq. (57), is π − φ), use Eq. (21), and solve and a maximum coloration level of Γ = 7 dB. From
for θ, to get Eq. (88), with fc ' 6 kHz, and ∆r = 15 cm[42], we
get θ = 9◦ , which we take as half the loudspeaker span.
 2 
−1 2γ − 1
 cs π − cos 2γ 2 From Eqs. (3)-(6), we can then calculate g = .985 and
θ∗ ' sin−1 

. (88) τc = 3 samples at a sampling rate of 44.1 kHz. These are
 2πfc ∆r 
precisely the dimensional and non-dimensional parame-
ters chosen for the calculations that are illustrated in the
A number of studies[27, 41] have suggested that XTC plots throughout this paper. The Branch-P and Branch-I
above a frequency of about 6 kHz is not critical or per- IRs are therefore given by those shown in the top panels
haps even necessary. Therefore, we set fc to that value of Figs. 5 and 8, respectively. The Branch-II IR is not
in the above equation, solve for θ∗ , design the filter for needed as the XTC filter is limited to 6 kHz, which, by
a loudspeaker span of 2θ∗ , use a 2-band crossover filter design, was made to be the upper bound of the second
to separate the first two bands, apply the Branch-I and band (Branch-P). The spectra associated with this filter
Branch-P parts of the filter to the first and second bands, are given by the solid curves in Figs. 6 and 7, with the
respectively, and allow the part of the audio spectrum dimensional frequency read off the top axes of the plots,
above fc to bypass the filter. up to the cuttoff frequency of 6 kHz. In particular, we
note that the XTC performance (top curve in Fig. 7), ex- free-field model can be useful in practical situations
ceeds 20 dB for a wide range of frequencies that extends where individualized HRTF-based XTC filters are
from the 6 kHz cuttoff down to 850 Hz, then drops off either too cumbersome to implement or not needed
with decreasing frequency, reaching 5 dB at 290 Hz. to attain the XTC levels required for enhancing the
spatial fidelity of playback in non-anechoic environ-
ments. We described a strategy for designing such
VI. SUMMARY optimal filters that meets practical design require-
ments, and we gave an illustrative example for a
We distill the main points and findings of this work in typical listening configuration.
the following bullets.
• 3-D reproduction of binaural audio with two loud- Acknowledgments
speakers requires cancellation of the crosstalk be-
tween the loudspeakers and the contralateral ears
The author wishes to thank Ralph Glasgal for intro-
of the listener. A perfect XTC filter (i.e., one with
ducing him to the XTC problem, Angelo Farina for his
infinite crosstalk cancellation) can be easily de-
X-Volver convolution plugin, Robin Miller for his inde-
signed but causes severe spectral coloration to the
pendent analytical auditioning of the filters, and J. S.
sound emitted by the loudspeakers due to the ill-
Bach for his Mass in B Minor.
conditioned inversion of the system’s transfer func-
tion.
• The coloration produced by the perfect XTC filter APPENDIX A: DERIVATION OF THE IMPULSE
consists of peaks in the frequency spectrum that RESPONSE OF THE OPTIMAL XTC FILTER
can typically exceed 30 dB, and thus strain the
playback transducers and significantly reduce the Here we carry out the derivation of Eqs. (74) to (83)
dynamic range of the playback system. Further- following the approach outlined in Section IV C.
more, the coloration is heard throughout the listen- We start by factoring the expressions appearing
ing space and, due to extreme sensitivity to errors in Eqs. (70) and (71), which, we note, have the same
in the system, it is also heard by the listener in the denominator, into the following product of terms:
sweet spot.
[O] [O]
HLL (iω) = HRR (iω) = (Ψ0 + γΨ1 ) Ψa (A1)
• Using a two-source free-field model, we have shown I,II I,II
[O] [O]
that constant-parameter regularization, which has (iω) = ∓Ψ0 + γgeiωτc Ψ1 Ψa ,

HLR (iω) = HRL
I,II I,II
been used previously to design HRTF-based XTC
systems, can tame these peaks but can produce a (A2)
bass roll-off and high-frequency artifacts in the fil-
ter’s frequency response. Furthermore, we demon- where
strated that constant-beta regularization does not
Ψ0 = γ 2 ±x − 1 + e2iωτc g 2 ,

(A3)
lead to the optimization of XTC filters across all p
frequencies, but rather only at discrete, widely- Ψ1 = g 2 ∓ x + 1, (A4)
spaced frequencies. 1
Ψa = p . (A5)
• Full optimization can be achieved through 2 2
g ± x 2γ g ∓ x + 1 − 1 + 1
frequency-dependent regularization, and requires
that the audio spectrum be divided into a hier- The term Ψa can be factored as
archical set of adjacent frequency bands, each of
which belonging to one of three solution branches Ψa = ±Ψ2 Ψ3 ± (Ψ1 ∓ Ψ4 )Ψ5 Ψ6 (c1 )Ψ6 (c2 ),
that make up the complete optimal filter. We de-
rived analytical expressions for the three branches where
of the filter in terms of series expansions, which we
1
showed are convergent for typical listening situa- Ψ2 = , (A6)
tions. The corresponding impulse responses were 2γx
then obtained analytically, and expressed as con- 1
Ψ3 = p , (A7)
volutions of trains of Dirac deltas. 2
g ∓x+1
• Aside from seeking fundamental insight into the na- Ψ4 = 2γx, (A8)
ture and characteristics of optimal XTC filters, we 1
Ψ5 = , (A9)
addressed a number of issues related to their appli- 8γ 3 x3
cation. In particular, we argued that analytical fil- 1
ters derived under the simplifying assumptions of a Ψ6 (c) = , (A10)
1 − cx−1
and Since cosm (ωτc ) can be written as the finite sum

p
16γ 2 (g 2 + 1) + 1 ∓ 1 m
X m

c1 = , (A11) m
cos (ωτc ) = 2−m e−i(2k−m)ωτc , (A20)
8γ 2 k
p k=0
− 16γ 2 (g 2 + 1) + 1 ∓ 1
c2 = . (A12)
8γ 2 and since the IFT of e−i(2k−m)ωτc is
Z ∞
In the time domain, the filter expressed by Eqs. (A1) 1
e−i(2k−m)ωτc dω = δ(2kτc − t − mτc ),
and (A2) becomes: 2π −∞
[O]
hLL (t) = hRR
[O]
(t) = (ψ0 + γψ1 ) ∗ ψa , (A13) the IFT of Ψ1 can be expressed as
I,II I,II
[O] [O] ∞ 1

hLR (t) = hRL (t) = [∓ψ0 + gγδ(t + τc ) ∗ ψ1 ] ∗ ψa , X 12 −m
I,II I,II ψ1 (t) = 2 (∓g)m g 2 + 1 ×
m
(A14) m=0
m
X m
where δ(2kτc − t − mτc ). (A21)
k
k=0
ψa = ±ψ2 ∗ ψ3 ± (ψ1 ∓ ψ4 ) ∗ ψ5 ψ6 (c1 ) ∗ ψ6 (c2 ). (A15)
• Ψ2 : Explicitly, Eq. (A6) is
The ψi terms are functions of time, and are the IFTs of
the Ψi terms, which are functions of frequency. sec(ωτc )
Ψ2 = .
We now seek the IFT of each of the Ψi terms given 4gγ
above.
• Ψ0 : The IFT of the expression in Eq. (A3) can be The problem is that the IFT of sec(ωτc ) cannot be ex-
readily found by substituting back 2g cos(ωτc ) for x and pressed in terms of real delta functions. However, the
carrying out the IFT integration: function sec(ωτc ) can be expressed as
1
Z ∞ 1
ψ0 (t) = γ 2 ±2g cos(ωτc ) − g 2 − e2iωτc dω
sec(ωτc ) = q (A22)
2π −∞ 1 − sin2 (ωτc )
= −g 2 γ 2 δ(t) ± gγ 2 δ(τc − t) ± gγ 2 δ(t + τc ) π π
if n2π − < ωτc < n2π +
−g 2 γ 2 δ(t + 2τc ). (A16) 2 2
with n = 0, 1, 2, 3, 4, . . .
• Ψ1 : Making the substitution b = g 2 + 1 in Eq. (A4),
we get Furthermore, we note that since
√ 1
Ψ1 = b ∓ x, (A17) 1≤γ≤ and 0 < g < 1, (A23)
1−g
which can be expressed as the series expansion
the arguments of the inverse cosine function in Eq. (60)
∞ 1
obeys the condition:
1
X
Ψ1 = 2 (∓x)m b 2 −m (A18)
m g2 γ 2 + γ 2 − 1
m=0
≥0 (A24)
2gγ 2
where we have used the binomial coefficient
which leads us to write
k k!
= if 0 ≤ m ≤ k π
m m!(k − m)! 0≤φ≤ . (A25)
= 0 if m < 0 or k < m. 2
In light of this expression and Eq. (56), we conclude that
Since 0 < g < 1, we have x = 2g cos(ωτc )| < g 2 + 1 = b, the conditions for the validity of Eq. (A22) are always
and the series in Eq. (A18) always converges. How- satisfied in Branch-I bands.
ever, as g → 1, b → 2, and when ωτc → n2π with Similarly, we find that sec(ωτc ) can be expressed as
n = 0, 1, 2, 3, 4, . . . , x → b and the series converges slowly. q
Replacing x and b by their explicit values, we get −1/ 1 − sin2 (ωτc ) for conditions that are always satis-
fied for Branch-II bands. Therefore, we can write
∞ 1

1
X
Ψ1 = 2 (∓2)m g m (g 2 + 1) 2 −m cosm (ωτc ). 1
m sec(ωτc ) = ± q , (A26)
m=0
(A19) 1 − sin2 (ωτc )
for which we wish to use the expansion where, following the same arguments as in the case of Ψ2 ,
the function sec3 (ωτc ) can be expanded in a convergent
∞
− 12

1 X series of the form of that in Eq. (A28), but with the
√ = (−1)m um . (A27) fraction −1/2 inside the binomial coefficient replaced by
1 − u m=0 m
−3/2. Therefore, by analogy with the result expressed
However, this series converges only for |u| < 1. For our in Eq. (A31), we have
particular case, u = sin2 (ωτc ) and the series diverges ∞
− 23

1 X
at ωτc = nπ/2, with n = 1, 3, 5, 7, . . . From the band ψ5 = ± (−1)m ×
division conditions in Eqs. (56) and (56) we see that these (4gγ)3 m=0 m
values of ωτc are always outside Branch-I and Branch-II 2m
X 2m
bands; therefore the convergence of the series is assured (−1)k+m 4−m δ(t + 2kτc − 2mτc ).
k
and this allows us to express Eq. (A26) as k=0
∞
(A35)
− 12
X
sec(ωτc ) = ± (−1)m sin2m (ωτc ). (A28) • Ψ6 : Eq. (A10) can be written as
m
m=0
1
Ψ6 = (A36)
Since sin 2m
(ωτc ) can be written as the finite sum 1 − y(c)
2m where
X 2m c c
sin 2m
(ωτc ) = (−1)k+m 4−m e2i(k−m)ωτc , y≡ = , (A37)
k x 2g cos (ωτc )
k=0
(A29) and c represents either c1 or c2 , given by Eqs. (A11) and
and since the IFT of e2i(k−m)ωτc is (A12), respectively. We wish to expand the function in
1
Z ∞ Eq. (A36) into the power series
e2i(k−m)ωτc dω = δ(t + 2kτc − 2mτc ), (A30) ∞
2π −∞ X
σ(c) ≡ y p (c) (A38)
the IFT of Ψ2 can be expressed as p=0
∞ but this series converges only for

1 X − 12

ψ2 = ± (−1)m × |y(c)| < 1. (A39)
4gγ m=0 m
2m We now show that this convergence condition leads to a
X 2m
(−1)k+m 4−m δ(t + 2kτc − 2mτc ). restriction on the allowable range of γ and g, but that
k
k=0 this restriction does not limit the applicability of the IR
(A31) to real listening configurations.
√ The inequalities in Eq. (A25), and the band divi-
• Ψ3 : The function 1/ b ∓ x, with b = g 2 + 1, has a sion conditions in Eq. (56) and (58), imply that x =
series expansion in the form of Eq. (A18), but with the 2g cos(ωτc ) is always positive in Branch-I bands and neg-
fraction 1/2 inside the binomial coefficient replaced by ative in Branch-II bands. Furthermore, we see from
−1/2. Therefore, by analogy with the result expressed Eqs. (A11) and (A12) that, under the conditions in
in Eq. (A21), we have Eq. (A23), c1 ≥ 0 and c2 ≤ 0. Therefore, we have
∞ y(c1 ) = c1 /x ≥ 0 in Branch-I bands, (A40)
− 12

X 1 −m
ψ3 (t) = (∓g)m g 2 + 1 2 × y(c1 ) = c1 /x ≤ 0 in Branch-II bands, (A41)
m
m=0
m and
X m
δ(2kτc − t − mτc ), (A32) y(c2 ) = c2 /x ≤ 0 in Branch-I bands, (A42)
k
k=0
y(c2 ) = c2 /x ≥ 0 in Branch-II bands. (A43)
which has the same convergence behavior as that of ψ1 . If we define η + (c) and η − (c) to be the non-dimensional
• Ψ4 : The IFT of Ψ4 = 2γx = 4γg cos(ωτc ) is straight- frequencies, ωτc , at which y(c) = +1 and y(c) = −1,
forward: respectively, we can, in light of the expressions above,
restate the convergence condition in Eq. (A39) as:
ψ4 = 2γgδ(τc − t) + 2γgδ(t + τc ). (A33)
σ(c1 ) converges in Branch-I bands if
• Ψ5 : Explicitly, Eq. (A9) is φ ≤ η + (c1 ) (A44)
3
sec (ωτc ) σ(c1 ) converges in Branch-II bands if
Ψ5 = , (A34)
(4gγ)3 η − (c1 ) ≤ π − φ (A45)
1.0
and
σ(c2 ) converges in Branch-I bands if

φ1 ≤ η − (c2 )) (A46) 0.8
σ(c2 ) converges in Branch-II bands if

η + (c2 ) ≤ π − φ. (A47)
0.6
Therefore, for σ(c) to converge in Branch-I and Branch-II
g
bands, all four inequalities must be satisfied. To express
these convergence conditions explicitly (i.e., in terms of 0.4
conditions on γ and g), we first set y(c) = +1 and y(c) =
−1, and solve for η + (c) and η − (c), respectively, to find
0.2
+ −1 f (g, γ) − 1
η (c1 ) = cos (A48)
16gγ 2
0.0
1 2 3 4 5 6 7 8 9 10

f (g, γ) + 1
η − (c1 ) = cos−1 − (A49) Γ
16gγ 2
FIG. 9: Region plot showing the allowed values for g and γ
and
(white). The black-shaded region is where the series conver-
gence condition in Eq. (A54) is not satisfied, and the grey-

+ −1 f (g, γ) − 1
η (c2 ) = cos − (A50) shaded region is where the general condition in Eq. (52) is
16gγ 2
violated.

− −1 f (g, γ) + 1
η (c1 ) = cos (A51) range of allowable γ and g, and is not relevant to real
16gγ 2 listening geometries, where g ' 1.
Aside from the series convergence condition above, γ
where, for compactness, we have used the function f (g, γ)
must satisfy the general condition given by Eq. (52)
defined as
(whose region of violation is shaded in grey in Fig. 9).
p
f (g, γ) ≡ 16 (g 2 + 1) γ 2 + 1. Therefore, we combine both conditions in the following
expression:
Using these four explicit expressions, along with the defi- p √ !
nition of φ given by Eq. (60), we find that the inequalities 5+ 5 1
max p ,1 ≤ γ ≤ , (A55)
in Eqs. (A44) and (A47) lead to the same explicit con- 2
2 g +1 1−g
vergence condition:
where the first argument of the max function comes from
f (g, γ) + 7 setting the left-hand side of the convergence condition in
≤ 1; (A52)
8(g 2 + 1)γ 2 Eq. (A54) to 1, and solving for γ.
Now that we have found the convergence condition for
and the inequalities in Eqs. (A45) and (A46) lead to the series in Eq. (A38), we can express Ψ6 as that series
and proceed to find its IFT. Replacing y and x in that
f (g, γ) + 9 series by their explicit values, we write
≤ 1. (A53)
8(g 2 + 1)γ 2 ∞ p
X c
Since both of these inequalities need to be satisfied, and Ψ6 = secp (ωτc ). (A56)
p=0
2g
since the latter condition is more stringent than the for-
mer, we must satisfy the latter. We can finally state The secp (ωτc ) term can be expanded in a convergent se-
the condition for σ(c) to converge in both Branch-I and ries of the same form as the series in Eq. (A28), but with
Branch-II bands explicitly in terms of g and γ: the fraction −1/2 inside the binomial coefficient replaced
p by −p/2, and this leads to:
16 (g 2 + 1) γ 2 + 1 + 9
≤ 1. (A54) ∞ ∞
p X
8(g 2 + 1)γ 2 − p2

X ±c
Ψ6 = (−1)m sin2m (ωτc ).
2g m
This convergence condition is illustrated in the region p=0 m=0
plot of Fig. 9, where the black region denotes the values of (A57)
g and γ for which the convergence condition is violated. Finally, recalling the finite sum in Eq. (A29), and the
It is clear that this restriction only slightly limits the associated IFT in Eq. (A30), we arrive at the sought
expression for the IFT of Ψ6 (c): more than the first few (5-10) terms of the infinite se-
ries in the expressions for all the ψ functions constitut-
∞ ∞
p X
− p2

X ±c ing the IR, except for ψ1 and ψ3 , which, due to their
ψ6 (c) = (−1)m × slow convergence at and near the frequencies ωτc = n2π
2g m
p=0 m=0
( n = 0, 1, 2, 3, 4, . . . ), require taking a larger number of
2m
terms. Approximating the infinite series in the expres-

X 2m
(−1)k+m 4−m δ(t + 2kτc − 2mτc ). sions for ψ1 and ψ3 by a sum having a finite number of
k
k=0 terms causes departures from the correct amplitude spec-
(A58) tra at and near these frequencies. Due to the logarithmic
frequency scale, the n = 0 departure appears as a slight
The complete impulse response of the optimal XTC bass roll-off in the first band (seen as the first dot in the
filter is assembled according to Eqs. (A13) to (A15), and first Branch-I band in the bottom panel of Fig. 10), and
is valid under the condition stated in Eq. (A55). the n ≥ 1 departures appear as narrow-band spikes (such
as the one appearing as three vertical dots in the fifth
band in the same plot). Increasing the number of terms
APPENDIX B: NUMERICAL VERIFICATION in the series above 1000 reduces the amplitude of the bass
roll-off and pushes it into the subwoofer frequency range,
The optimal XTC IR derived in the previous appendix where XTC is not needed, and causes the n ≥ 1 spikes
was evaluated for the typical case of g = .985 and to diminish in amplitude and frequency extent so as to
Γ = 7 dB, and plotted in Fig. 8. To verify the IR’s valid- become inaudible. (The XTC spectrum is more immune
ity and assess the effect of the number of terms in the se- from the aforementioned departures, as seen in the top
ries expansions, we calculated its Fourier transform and panel, because it is a ratio of left to right spectra.)
compared the resulting spectra to those obtained from
the frequency-domain expressions of Section IV B. An
example is shown in Fig. 10 for the Branch-I part of the
XTC spectra (top panel) and envelope spectra (bottom
panel). A similar analysis of the Branch-II part of the IR is not
We found that excellent agreement (within a few tenths shown as the resulting spectra exhibit the same behavior
of a dB) over all frequencies does not require taking as that described above.
[1] D. H. Cooper and J. L. Bauk, J. Audio Eng. Soc. 37, 3 [13] While, as mentioned above, XTC levels above 20 dB are
(1989). often required for robust front-back image disambigua-
[2] Throughout this paper, the words “recording” and “sig- tion, the larger portion of the direct sound content in
nal” are used interchangeably and are meant to also rep- acoustic recordings, e.g., performed music, is of frontal
resent a live feed, or the HRTF-encoded signal for the origin and, with playback through frontal loudspeakers
artificial placement of sounds in a virtual acoustic space. at modest XTC, is largely immune to such localization
[3] M. A. Akeroyd et al., J. Acoust. Soc. Am. 121, 1056 confusion.
(2007). [14] D. B. Ward and G. Elko, IEEE Signal Process. Lett. 6,
[4] D. B. Ward, J. Acoust. Soc. Am. 110, 1195 (2001). 106 (1999).
[5] A. Sæbø, Ph.D. thesis, Norwegian University of Science [15] J. J. Lopez and A. Gonzalez, IEEE Signal Process. Lett.
and Technology, Trondheim, Norway, 2001. 6, 106 (1999).
[6] Throughout this paper, the word “level” is meant to rep- [16] T. Takeuchi, P. A. Nelson, and H. Hamada, J. Acoust.
resent, generally, a frequency-dependent amplitude. Soc. Am. 109, 958 (2001).
[7] J. Blauert, Spatial Hearing (The MIT Press, Cambridge, [17] J. Rose, P. Nelson, B. Rafaely, and T. Takeuchi, J.
MA, 2001). Acoust. Soc. Am. 112, 1992 (2002).
[8] B. B. Bauer, J. Audio Eng. Soc. 9, 148 (1961). [18] M. R. Bai and C. C. Lee, J. Acoust. Soc. Am. 120, 1976
[9] B. S. Atal, M. Hill, and M. R. Schroeder, Apparent Sound (2006).
Source Translator, US patent No. 3,236,949. Application [19] O. Kirekby, P. A. Nelson, and H. Hamada, J. Acoust.
filed Nov. 1962, granted Feb 1966. Soc. Am. 104, 1973 (1998).
[10] P. Damaske, J. Acoust. Soc. Am. 50, 1109 (1970). [20] O. Kirekby, P. A. Nelson, and H. Hamada, J. Audio Eng.
[11] An exception could be made for recordings in which the Soc. 46, 387 (1998).
specific placement of sound images was made with full ac- [21] H. S. Kim, P. M. Kim, and H. B. Kim, ETRI J. 22, 11
counting for crosstalk during playback, e.g., the case of (2000).
stereo soundfields constructed with pan-potted mono im- [22] J. Yang, W. S. Gan, and S. E. Tan, Acous. Res. Lett.
ages and monitored over loudspeakers, common in pop- Online 4, 47 (2003).
ular music recording. [23] M. R. Bai, C. C. Tung, and C. C. Lee, J. Acoust. Soc.
[12] C. Hugonnet and P. Walder, Stereophonic Sound Record- Am. 117, 2802 (2005).
ing (John Wiley & Sons, Chichester, UK, 1998). [24] M. R. Bai and C. C. Lee, J. Sound & Vibration 117,
f HHzL 2802 (2005).

100 1000 104 2x104 [25] Y. Kim, O. Deille, and P. A. Nelson, J. Sound & Vibra-
40 æ
æ
æ
tion 297, 251 (2006).
æ æ
æ æ
æ
[26] L. H. Kim, J. S. Lim, and K. M. Sung, IEICE Trans.
30 æ æ
æ
æ
æ æ
æ æ
æ
Fundamentals E85 A, 2159 (2002).
æ
æ
æ æ
æ æ æ
æ æ
[27] W. G. Gardner, 3-D Audio using Loudpseakers (Kluwer
æ æ æ æ
20 æ
æ æ
æ
æ
æ
æ
æ ææ ææ æ
æ æ æ
æ ææ ææ æ
Academic Publishers, Boston, 1998).
ææ
æ
æ
æ
æ
æ
ææ
ææ æ
æ
æ
æ æ
ææ æ
æ æ
æ
æ
æ
æ
æ ææ ææ
æ æ ææ
æ æ
æ
æ
ææ
æ
ææ
[28] Conversely, XTC control at frequencies where the inter-
æ æææ æ
æ
æ
æ
ææ
æ
æ
æ
æ æ æææ
æ
ææ
æ
æ
ference at the ear is constructive requires attenuation in-

ææææ æ æ
ææ æ ææ æ
dB
æææææææææææææ æ æ
10 ΧAOE æ
æ
æ
æ
æ
æ
æ
ææ
ææ
ææ
æ
æ
ææ
æ æ
æ ææ
stead of boosting (and implies a dynamic range gain, in-
æ
æ ææ
I æ
æ
æ
æ
æ ææ
ææ
ææ
æ
æ
0
æ
æ
æ
ææ
æ
æ
æ
æ
ææ
æ
æ
ææ
æ
ææ
æ
æ
ææ
æ
ææ
æ
æ
ææ
æ
ææ
ææ
ææ
ææ
æ
ææ
ææ
æ
ææ
ææ
æ
ææ
ææ
æ
æ
æ
æ
æ
æ æ
æ
æ
æ
æ
æ
æ
ææ
æ
ææ
æ
ææ
æ
æ
ææ
æ
æ
ææ
stead of loss). As was shown by [29, 30], and as will be
ææ
æ æ
æ æ
æææ
reviewed in Section II C, these attenuations are not prob-

æ
æ ææ æ
æ æ æ æ
æ
æ æ
ææ æ
æ æ æ
æ æ
æ æ ææ
æ
-10
ææ
æ
æ
æ
æ
æ
æ
æ
æ lematic as they correspond to frequencies where XTC
æ
Branch I P æ
æ II æ
æ
P I æ
control is most robust.
-20 æ
æ
[29] T. Takeuchi and P. A. Nelson, J. Acoust. Soc. Am. 112,
Π100 Π10 Π2 Π 2Π 3Π 2786 (2002).
Ω Τc [30] P. A. Nelson and J. F. W. Rose, J. Acoust. Soc. Am.
118, 193 (2005).
[31] B. Katz, Mastering Audio (Focal Press, Oxford, UK,
2002).
[32] T. Takeuchi and P. A. Nelson, J. Audio Eng. Soc. 5, 981
(2007).
[33] A. Schuhmacher, J. Hald, K. B. Rasmussen, and P.
C.Hansem, J. Acoust. Soc. Am. 113, 114 (2003).
[34] H. Tokuno, O. Kirkeby, P. A. Nelson, and H. Hamada,
IEICE Trans. Fundamentals EE80-A, 809 (1997).
[35] P. M. Morse and K. U. Ingard, Theoretical Acoustics
(Princeton University Press, Princeton, NJ, 1968).
[36] P. C. Hansen, Rank-Deficient and Discrete Ill-Posed
Problems (SIAM, Philadelphia, PA, 1998).
[37] P. A. Nelson and S. J. Elliott, Active Control of Sound
(Academic Press, London, UK, 1993).
[38] M. Bellanger, Digital Processing of Signals (John Wiley
& Sons, Chichester, UK, 2000).
[39] W. G. Gardner, J. Audio Eng. Soc. 43, 127 (1995).
[40] A. Torger and A. Farina, IEEE Workshop on Applica-
FIG. 10: XTC spectrum of optimal filter for Branch-I bands, tions of Signal Processing to Audio and Acoustics (PUB-
[O]
χI (ω), shown in top panel, and the associated envelope spec- LISHER, New Paltz, New York, 2001), paper No. 198.
[O]
trum, ŜI (ω), shown in bottom panel. The small dots repre- [41] M. R. Bai and C. C. Lee, EURASIP J. on Adv. in Sign.
sent the spectra calculated by taking the Fourier transform of Process. 2007, 1 (2006).
the Branch-I part of the IR derived in the previous appendix. [42] This value for the effective inter-ear separation, ∆r =
(The IR is shown graphically in the top panel of Fig. 8.) Only 15 cm, is justified by the relatively small loudspeaker
the first 20 terms of the infinite series representing the ψ func- span, following the guidelines in Ref. [29], where it is re-
tions were taken, with the exception of the series for ψ1 and ported that good correlation between the peak frequen-
ψ3 , for which the first 2500 terms were used. The hard curve cies in the data calculated using a free-field model, and
in the top panel is the Branch-I XTC spectrum calculated those measured with the KEMAR dummy head, can be
directly from Eq. (64), and the horizontal line in the bottom obtained by taking an effective ∆r ' 13 cm for low val-
[O]
panel is the Branch-I envelope spectrum ŜI (ω) = γ, with ues of θ, and ∆r ' 25 cm for large source azimuths. The
Γ = 7 dB. (Other parameters are the same as for Fig. 2.) larger value, which is much larger than the minimum
Since these spectra are valid only in Branch-I bands, all other distance between the entrances of the ear canals of the
bands are shaded in grey. (The vertical dashed lines represent dummy head, reflects the effects of diffraction around the
the frequency bounds of the successive bands, and the branch head.
numbers of the first five bands are given in the bottom half
of each panel.)

Optimal Crosstalk Cancellation For Binaural Audio With Two Loudspeakers

Uploaded by

Copyright:

Available Formats

Optimal Crosstalk Cancellation For Binaural Audio With Two Loudspeakers

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimal Crosstalk Cancellation For Binaural Audio With Two Loudspeakers

Uploaded by

Copyright:

Available Formats

Optimal Crosstalk Cancellation for Binaural Audio with Two Loudspeakers

Crosstalk cancellation (XTC) yields high-spatial-fidelity reproduction of binaural audio through

I. INTRODUCTION as both crosstalk and room reflections would be non-

In order to render the analysis tractable enough so that

1 spectra, we find the following amplitudes and frequencies

 [β] [β]  .005

by: We can then re-write Eqs. (33) and (34) as

The IR of the perfect XTC filter is shown in the top -0.5 ç

panel of that figure and consists of two trains of decay-

HLR (z) = HRL (z) = − , (46) ç æ

cover the IR of the perfect XTC filter derived in Ref. [20]: æ

n=0 [β] [β]

30 • Bands 4, 8, 12, 16, 20, . . . , 4n+4 belong to Branch P,

10 leading to a minimization of the performance error. This

the following two IRs:

and Since cosm (ωτc ) can be written as the finite sum

∞  but this series converges only for

σ(c2 ) converges in Branch-I bands if

σ(c2 ) converges in Branch-II bands if

Therefore, for σ(c) to converge in Branch-I and Branch-II

f HHzL 2802 (2005).

ference at the ear is constructive requires attenuation in-

reviewed in Section II C, these attenuations are not prob-

You might also like

∞ but this series converges only for