You are on page 1of 9

TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 1

A Phase Grating Approach to Modeling Surface


Diffusion in FDTD Room Acoustics Simulations
Konrad Kowalczyk, Maarten van Walstijn, and Damian Murphy

Abstract—In this paper, a method for modeling diffusive the frequency domain [2], [3]. Such predictions can also be
boundaries in finite difference time domain (FDTD) room acous- made using discrete-time simulations and in addition these
tics simulations with the use of impedance filters is presented. tools can be used for broadband room auralization, which
The proposed technique is based on the concept of phase grating
diffusers, and realized by designing boundary impedance filters has applications in architectural design as well as in audio.
from normal-incidence reflection filters with added delay. These In an auralization context, good simulation predictivity is in
added delays, that correspond to the diffuser well depths, are some cases of lesser importance, the main objective shifting
varied across the boundary surface, and implemented using to enabling better control over the overall room diffusivity.
Thiran allpass filters. The proposed method for simulating sound Amongst simulation approaches, the main distinction is that
scattering is suitable for modeling high frequency diffusion
caused by small variations in surface roughness and, more gener- between geometrical methods and wave-based methods. In the
ally, diffusers characterized by narrow wells with infinitely thin former, scattering of sound rays is modeled through a ran-
separators. This concept is also applicable to other wave-based domized distribution of the ray reflection angle, for example
modeling techniques. The approach is validated by comparing using Lambert’s law [4]. Diffraction effects are not directly
numerical results for Schroeder diffusers to measured data. In addressed in the geometrical approach, but for some methods
addition, it is proposed that irregular surfaces are modeled by
shaping them with Brownian noise, giving good control over the they can be approximated using additional techniques [1]. The
sound scattering properties of the simulated boundary through main drawback of the geometrical approach is that diffusive
two parameters, namely the spectral density exponent and the properties cannot always be unambiguously assigned to a
maximum well depth. particular wall, i.e., different values of scattering coefficient
Index Terms—Architectural acoustics, FDTD methods, acous- may be required for the same wall, even under the same room
tic scattering, acoustic refraction, fractals, digital filters, acoustic condition [5]. One implication is that in practice, the boundary
signal processing diffusion parameters are often “fine-tuned” by acoustic engi-
neers operating geometrical room acoustics software in order
I. I NTRODUCTION to achieve the desired parameter values for the whole acous-
tic space such as the reverberation time or clarity. Another
OUND SCATTERING is a fundamental phenomenon
S encountered in room acoustics that plays a crucial role
in defining the acoustic properties of spaces. It encompasses
implication is that geometrical models tend to overpredict the
reverberation time, especially when the absorption is unevenly
distributed (e.g., concert halls) or in highly disproportional
edge diffraction effects as well as diffusive reflection of sound rooms [6]. In summary, geometrical room acoustic models
waves from not perfectly smooth/uniform surfaces. During the have fundamental limitations in their capacity to verify or
past few decades, several methods for modeling acoustic wave predict the performance of realistic diffusers in a real acoustic
propagation and boundary reflection have been proposed and space.
developed (see [1] for an overview). The process of designing In wave-based methods, all natural wave-related phenomena
spaces with desirable acoustics benefits significantly from such as wave interference and edge diffraction are inherently
such simulation tools, as they allow the acoustic performance modeled. The downside is that the computational costs are
to be predicted before construction begins. To date, mainly extremely high, especially so for 3D simulations. However a
the boundary element method (BEM) has been applied to lot of insight on diffusion can be gained using 2D simulations,
predicting the performance of diffusers, usually working in that are much less costly. In the long term one may expect that
the burden of the computational cost will be lessened by the
Manuscript received May 6, 2009; revised February 27, 2010; accepted
May 7, 2010. The associate editor coordinating the review of this manuscript growth in commonly available processing power, e.g., multi-
and approving it for publication was Dr. Jingdong Chen. core programmable architectures such as graphical processing
K. Kowalczyk was with the Sonic Arts Research Centre, Queen’s Univer- units.
sity, Belfast, UK. He is now with Multimedia Communication and Signal
Processing (LMS), University of Erlangen-Nuremberg, 91058 Erlangen, Ger- Among wave-based methods, the finite difference time
many e-mail: (kowalczyk@lnt.de) domain (FDTD) method has recently emerged as a possible
M. van Walstijn is with the Sonic Arts Research Centre (SARC) and the numerical tool for investigation of surface diffusion (see, e.g.,
School of Electronics, Electrical Engineering and Computer Science, Queen’s
University, Belfast BT7 1NN, UK e-mail: (m.vanwalstijn@qub.ac.uk) [7], [8]. In such studies, an irregular boundary geometry is
D. Murphy is with the Audio Lab, Department of Electronics, York Uni- usually represented directly in the grid structure, for example a
versity, Heslington, York, YO10 5DD, UK e-mail: (dtm3@ohm.york.ac.uk) staircase implementation [7], [8]. This works fine for variations
This research has been supported by the ESF and in part by the UK EPSRC
grant number EP/F013078/1. in the boundary shape that are large in comparison to the
Digital Object Identifier grid spacing, but adapting to smaller variations (i.e., subcell
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 2

geometries) can be problematic, since local grid refinement II. FDTD SIMULATION
leads to a significant increase in computational load [9]. A. Modeling the medium
Diffusion has also been addressed in digital waveguide
mesh (DWM) modeling, an approach that is mathematically Sound wave propagation in air is governed by the wave
equivalent to the explicit FDTD method [10]. For example, equation. FDTD schemes for numerical simulation of the wave
basic quadratic residue diffusers were modeled with integer equation are derived by approximating time and space deriva-
length delay lines in [11], [12], [13]. Unfortunately these tives with finite difference (FD) operators. This technique is
models used 1D reflection coefficients to model the boundary, usually characterized by a regular spatial grid, the size of
an approach that has recently been shown to be unphysical and which depends on the sampling frequency. The nonstaggered,
may lead to significant boundary reflection errors [14]. Most standard leapfrog rectilinear finite difference formulation of
other studies in the DWM field tend to aim toward controllable the wave equation is obtained by applying centered difference
diffusion rather than predictivity. For example, in [15] circulant operators to approximate the derivatives in the continuous-
matrices were applied, which result in random redirection of time wave equation [19]. FDTD schemes generally suffer from
reflected waves at the boundary. A similar approach was taken numerical dispersion errors, i.e., the numerical phase velocity
in [12], where in order to avoid the rotation error caused by the differs from the theoretical phase velocity, and for 2D and 3D
non-uniform distribution of connecting nodes at a boundary, it rooms this phenomenon is generally dependent on frequency
was proposed to apply the rotations at the mesh nodes adjacent and propagation direction [20]. In order to minimize the dis-
to the boundary, resulting in a so-called diffusing layer. While persion error, the Courant number λ = cT /X, where c is the
high levels of diffusion can be induced with this approach, wave speed, T is the sampling period, and X is the spatial
√ step,
a downside is the unphysicality of placing a local, point- is usually
√ chosen at the stability bound, i.e., λ = 1/ 2 and
based scattering mechanism in front of the actual boundary. λ = 1/ 3 for modeling sound wave propagation in 2D and
Furthermore, the latter two diffusion modeling techniques do 3D rooms, respectively. For top values of the Courant number,
not allow for modeling real diffusers, and therefore they are the standard leapfrog rectilinear scheme is mathematically
not suitable for the prediction of acoustics in rooms with equivalent to the rectangular digital waveguide mesh (DWM)
diffusers mounted on walls. [10]. The latter, when implemented in a FDTD form, is often
Most of these numerical studies treat diffusion separately referred to in the DWM literature as the Kirchhoff variable
from absorption, which for real rooms is frequency dependent. digital waveguide mesh (K-DWM) and its implementation
While a totally separate control over these two phenomena is identical to the implementation of the standard leapfrog
is generally not possible, predictive simulations eventually scheme [21].
still need to model both phenomena together. The current
paper addresses this issue, taking an FDTD approach in B. Modeling specular reflections from absorptive surfaces
which a smooth, flat boundary is made diffusive by varying
The material and construction of the boundary define locally
the impedance across the boundary surface, thus imposing a
the absorptive properties of the wall, and such information can
spatial variation of the ‘local well depth’. Hence our approach
be expressed by a complex impedance. A digital reflection
is based on the well-known concept of phase-grating diffusers
filter should therefore be designed to match as closely as
(see [16], [17] for an introduction to this concept). In the nu-
possible the required complex boundary reflectance at normal-
merical formulation, which is based on the digital impedance
incidence for given absorption data in each frequency band.
filter (DIF) formulation developed in previous work [14], each
Commonly available reflectance data is often in the form of
node in the boundary is represented by an impedance filter that
absorption coefficients (α) in each octave or one-third octave
is designed from a normal-incidence reflection filter consisting
frequency band, which when converted to a√set of normal-
of an absorption filter (that if left by itself would model
incidence reflection coefficients by |R0 | = 1 − α can be
specular reflections) in cascade with a fractional delay line.
used for the design of an appropriate reflection filter. However,
The proposed method overcomes the fundamental limitations
this approach does not include any phase information, and
of previous FDTD and related models discussed above; in
therefore a more complete and physically correct approach
particular, it allows the simulation of diffusers with precisely
is to use the data that defines both amplitude and phase
specified well depths, without creating inconsistencies with
changes in the reflected sound wave. Such complex impedance
the physics of real-life boundaries. It also allows a simple but
data for porous materials is for instance available in the
powerful parametrization of the diffusiveness of boundaries,
form of analytic formulae in [22], and an appropriate digital
based on the concept of fractional Brownian diffusers [18].
reflection filter R0 (z) can be designed that matches closely
The paper is structured as follows. A brief review of the
the analytically calculated continuous-time reflectance [14].
FDTD technique as well as the DIF boundary model for the
Given the variety of available methods for digital filter design
standard leapfrog finite difference scheme is given in Section
techniques, the resulting digital normal-incidence reflection
II. Then in Section III, our method for simulating reflection
filter R0 (z) can be either FIR or IIR and of arbitrary order.
phase-grating diffusers with the use of fractional delay lines
A corresponding digital filter that represents the local specific
is proposed, and ways of affecting and controlling diffusion
acoustic impedance is then calculated with
are discussed. Results obtained with numerical experiments
are analyzed and compared to other methods in Section IV, 1 + R0 (z)
ξw (z) = . (1)
followed by concluding remarks in Section V. 1 − R0 (z)
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 3

This digital impedance filter ξw (z) is always an IIR filter, the (a)
order N of which is equal to the highest order of the normal-
incidence reflection filter R0 (z).
To derive the numerical boundary model, the boundary
condition is then combined with the multidimensional (i.e., ξ(m-5) ξ(m-4) ξ(m-3) ξ(m-2) ξ(m-1) ξ(m) ξ(m+1) ξ(m+2) ξ(m+3) ξ(m+4) ξ(m+5) ξ(m+6)
(b)
2D or 3D) wave equation [23], and an explicit formulation
that is stable for any physically feasible impedance filter is
obtained by carefully constructing recursive update equations.
The set of formulae for such a boundary formulation of a
2D standard leapfrog scheme has been proposed in [14]; the (c)
example update formula for the node on a right boundary is
given as (see [14] for more details)
[
Fig. 1. (a) Irregular surface geometry, (b) surface impedance variation
pn+1
l,m = λ2 (2pnl−1,m + pnl,m+1 + pnl,m−1 ) + 2(1 − 2λ2 )pnl,m (SIV) model - a way of modeling these irregularities by spatial variation of
λ2 n ] ( λa0 )
impedance filters, and (c) staircase model. The impedance filters are designed
λa0
+( − 1)pn−1
l,m + g / 1 + , (2) from normal-incidence filters with added well-depths, thus implementing a
b0 b0 b0 phase grating diffuser.

where pnl,m denotes the pressure at node l, m at time t = nT .


The filter input x at time step n is computed from III. A NEW METHOD FOR SIMULATING DIFFUSIVE
SURFACES
a0 n+1 gn
xn = (pl,m − pn−1
l,m ) − , (3) A. A surface impedance variation approach
λb0 b0
The reflective properties of the boundary can also be made
and the explicit filter difference equation is updated with irregular by spatial variation of the wall impedance. This can
be done without changing the local absorption properties.
1( )
That is, only the reflection phase properties are spatially
yn = b0 xn + g n , (4)
a0 varied, while the reflection amplitude characteristics remain
unaltered. Both this surface impedance variation (SIV) method
where the intermediate value g is
and the staircase method described above are based on the
N (
∑ ) principal idea behind phase grating diffusers [16], [17], [24].
gn = bi xn−i − ai y n−i , (5) A schematic representation of the two models is depicted in
i=1 Fig. 1.
The starting point for the design of each of the impedance
where a0 , a1 , ..., aN are the impedance filter nominator coeffi- filters is to consider first a locally-defined normal-incidence
cients and b0 , b1 , ..., bN are the denominator coefficients. The digital reflection filter of the form
digital impedance filter (DIF) boundary formulation requires
updating the intermediate value g from the past filter input and R0 (z) = Ra (z) Rd (z), (6)
output values only, and calculating the boundary node pl,m , where Ra (z) is a digital reflection filter designed to match
and the filter input and output at time step n according to the absorption data and Rd is a fractional delay filter that
(5), (2), (3), and (4), respectively. Note that updating these adds a local ‘well-depth’. From (6) it follows directly that
equations in the proposed order avoids the necessity to update a flat surface that reflects waves in a specular fashion is
the ghost point pl+1,m (i.e., the point located outside of the simulated by setting Rd (z) = 1 for all impedance filters. Non-
modeled space) [14]. smooth surfaces thus result when varying the well depths (i.e.,
the delay lengths) along the surface. We note that this may
increase the overall surface absorption, even though the local
C. Modeling diffuse reflections absorption properties are not affected by any changes in the
The boundary formulation explained above can be employed Rd (z) filters (i.e., the magnitude response of Rd (z) filter is
to simulate diffusive boundaries by implementing an irregular flat, as explained in Section III-C).
boundary shape, i.e., by spatial variation of the boundary Once the digital reflection filter is designed, it is readily
geometry. Under the assumption that the surface is locally converted to a digital impedance filter using (1), and incorpo-
reacting, absorption can be realized as usual, so this approach rated in the boundary model as described in Section II. For
incorporates both diffusion and absorption, and in a manner any passive boundary [i.e., |R0 (z)| ≤ 1], (1) yields a positive
that is more consistent with the physics of the problem than real transfer function of a digital impedance filter. As shown in
geometrical approaches. However, one limitation is that the [14], this also guarantees the overall stability of the simulation.
geometry is confined by the spatial grid resolution (X), leading
to a staircase model (see Fig. 1(c)). One consequence is that, B. Relationship between the well depth and delay length
in modeling diffusers, the well depths are restricted to integer For a given wave velocity c, the time it takes to propagate
multiples of X. along a well-depth distance d is τ = d/c. Hence the required
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 4

delay length [the number of delays modeled with Rd (z)] is diffusion data, to account for small surface roughness or to
2d mask numerical artifacts. Modeling commercially available
D= . (7) diffusers is not suitable for this purpose since such optimized
cT
surfaces usually yield top values of the diffusion coefficient
As mentioned in Section II-A, the wave velocity actually
(see Section IV-A for a definition) for the frequencies of
differs from c for most frequencies and directions. For the
interest. Therefore we propose a technique to control the
standard leapfrog scheme, this amounts to no error in diagonal
diffusion properties of the numerical boundary that is based
directions and the largest error occurring at Nyquist for axial
on fractional Brownian diffusers [18].
directions, where the ratio
√ between the numerical and real This method works as follows. A Gaussian white noise
phase velocity becomes 1/2 [25]. Given that it is impossible
signal is first generated with the length equal to the number
to adjust the delay length of a well to all possible directions
of the boundary nodes along a diffusive surface. Such a noise
and frequencies, (7) is the best possible overall design choice.
signal is then spectrally shaped with the use of a digital filter
Note that this assumes a normal angle of incidence, which
designed to match values of the diffusion coefficient in each
is the most natural choice since diffuser wells are placed
frequency band. The maximum depth in shape, defined by
in the direction perpendicular to the boundary surface. The
the maximum delay value, puts a constraint on the lowest
resulting non-exactness of the delay for certain directions
frequencies that are affected by the diffusion. As will be shown
and frequencies may however lead to small high-frequency
in the results presented in Section IV, the lowest frequency
deviations in diffusion characteristics for some diffuser types.
at which diffusion occurs can be estimated in a similar way
to defining the theoretical design frequency of the Schroeder
C. Fractional delay implementation diffuser [16]. The maximum depth dmax should thus be half
Ideally, the filter Rd (z) should model an exact number of of the longest wavelength
delays, and this number will generally be fractional (i.e., a λ0
non-integer value). The ideal fractional delay element (i.e., dmax = , (9)
2
an ideal band-limited interpolator) is nonrealizable, since the
corresponding impulse response would be noncausal and in- which also enables computing the lowest frequencies affected
finitely long [26]. Suitable approximations can be made using by diffusion from
c
so-called fractional delay filters with a maximally flat phase f0 = , (10)
2dmax
delay response, such as Lagrange interpolation FIR filters and
Thiran allpass filters [27], [26]. The main difference between where f0 is the lowest diffusion frequency. However, in
these two filter design methods is that the Lagrange FIR practice some diffusion actually starts below this theoretical
filter introduces a low-pass filtering effect, while the Thiran value to peak at the design frequency.
design is strictly allpass. Hence the latter were employed in The amount of diffusion is obtained by spectral shaping
this study, mainly in order to maintain full control of (local) with the use of a simple linear roll off filter, the gain of which
absorption properties through Ra (z). A simple rule of thumb is given as
1
for the stability and accuracy of such allpass filters is given G(f ) = β/2 , (11)
as [M − 0.5 ≤ D < M + 0.5], where D denotes a fractional f
delay value and M is the filter order [26]. where β denotes the spectral density exponent which in order
For efficiency reasons, it is often desirable to keep the to assure the 1D fractal shape takes the values within the range
fractional delay filter Rd (z) of lower order, such that the final of 1 and 3. Low values of the spectral density exponent yield
impedance filter also remains small in order; in this study a more spiky surface shape resulting in a highly diffusive
we therefore employed second-order allpass filters. Fractional boundary. On the other hand, for high values of β, the
delays larger than M + 0.5 can be obtained by combining the surface shape is much smoother and the diffusion is less
allpass filter with a delay line [28], hence the filter transfer pronounced at frequencies above the design frequency, i.e.,
function is the frequency for which a diffuser should yield the strongest
Rd (z) = A(z) z −N , (8) diffusion [17]. Note that for creating surface diffusion at very
low frequencies, the changes in the boundary impedance have
where A(z) denotes the second-order allpass filter and z −N
to be sufficiently slow so that long waves are effectively
denotes a delay line of N = D − 2 delays. For very small
affected, analogously to the well width of Schroeder diffusers
fractional delays [0.1 ≤ D ≤ 1.1], a first-order allpass filter
(see [17] for more details). Therefore, to obtain low frequency
has to be applied. For D ≤ 0.1, pole-zero cancelation may
diffusion in simulations characterized by a high value of the
happen in practical implementations due to inevitable round-
sample rate (and thus small grid spacing), the generated noise
off errors, and should preferably be avoided, so the minimum
sequence should be a few times shorter so that the calculated
fractional delay is set to D = 0.1 [28].
filter delay values are repeated at neighboring boundary nodes.
Furthermore, the reader is reminded that high values of the
D. Diffusion parameter control diffusion coefficient at low frequencies rarely originate from
In many practical cases, it is desirable to control the strength the surface scattering alone, but rather from the diffraction
of diffusion and the frequencies for which diffusion occurs. For effects caused by finite-length boundaries. Consequently, this
example, controlled diffusion can be used to match measured technique should not be over-used to compensate for such
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 5

Compuational domain
large-size changes. That is, in such cases the boundary shape
should instead be implemented directly in the grid structure
(i.e., staircase implementation).

IV. N UMERICAL EXPERIMENTS


A. Test setup
The diffusion coefficient characterizes the sound reflection
from a surface in terms of the uniformity of the scattered polar
distribution; the criteria for measuring the diffusion coefficient
are defined in the AES 4id 2001 standard [29]. The setup
used in the numerical experiments in this study complies 4.4m
7.7m
with these AES guidelines and is illustrated in Fig. 2. For
investigation of Brownian diffusers in Section IV-C, the size
Fig. 2. Schematic depiction of the simulation setup according to the AES
of the modeled room (4000 × 4000 grid nodes corresponding standard [29]; black squares indicate omnidirectional source positions and
to 44 m x44 m) and the simulation time (4410 samples at the black circles indicate receiver positions on a semicircle. The computational
sample rate of 44.1 kHz) were set in such a way that only the domain is indicated with outer rectangle.
reflections from the investigated boundary sample could reach
0 1 4 2 2 4 1 0 1 4 2 2 4 1
the receiver positions and at the same time sufficiently long 0
so that the whole wavelet could reach all the receivers. The 1
2D standard leapfrog scheme was used for the implementation, 2
and room and diffuser boundaries were implemented using the 3
DIF boundary model [14]. Since the focus of this paper is on 4
modeling and measuring diffusion, the absorption reflection 7 sequence 7 sequence
filter used in (6) is set in case of all modeled diffusers to
Ra (z) = 1 for simplicity. In most simulations, the diffuser Fig. 3. The well depths of the quadratic residue diffuser based on 7-ary
sample was 101 grid points wide (which corresponds to 1.1 m) sequence.
and 3 grid points deep (which corresponds to 0.033 m). Note
that the proposed diffusion method allows for modeling much
deeper wells without having to increase the external depth of were obtained as h = h1 − h2 . The virtual microphones
the diffuser’s sample. The numerical measurement consisted and virtual loudspeakers used in the experiments are treated
of 19 simulations, in which the source position was changed as omnidirectional since the input and output signals were
with a maximum angular separation of 10o on a semicircle injected into one grid node and picked-up from another grid
with a radius of 7.7 m (700 grid points), as shown in Fig. 2. node. Hence there was no need to deconvolve the loudspeaker-
Similarly, a set of receivers with an angular receiver resolution microphone response. The impulse response windowed with
of 5o was located on a semicircle with a radius of 4.4 m (400 the use of the right half of the Hanning window was Fourier
grid nodes) with the sample diffuser positioned in the center of transformed and sound pressure levels Li (in decibels) were
the semi-circle. Such distances are in a good agreement with calculated in one-third octave bands. Following [29], the
the AES standard [29], in which when true far field conditions directional diffusion coefficient was computed in each one-
are hard to ensure, the minimum criterion is that at least 80% third octave band from the set of sound pressure levels Li
of the receiver positions should be outside of the specular zone. from n receivers for a fixed source position with the formula
The same test setup was also applied in numerical measure- ( n )2 n (
∑ ∑ )2
ments presented in Section IV-B. However, since the overall 10 Li /10
− 10Li /10
diffuser width and depth were set as 3.6 m and 0.2 m, i=1 i=1
dθ = n ( )2 . (12)
respectively, the distances from the receivers and sources ∑
to the diffuser sample were increased 3 times to comply (n − 1) Li /10
10
with the AES guidelines. Furthermore, the sample rate was i=1
changed to 38.86 kHz so that the distance of 16 grid points Finally, the random-incidence diffusion coefficient (as defined
corresponded exactly to the well depth of 0.2 m, for the in [29]) was calculated as an arithmetic mean of the directional
staircase implementation. diffusion coefficients for all source positions.
Each numerical test consisted of two simulations, where
a Gaussian-shape initial condition was used for excitation
in the source position. In the first simulation, the receivers’ B. Validation of the approach
impulse responses h1 were measured with the diffuser sample The proposed surface impedance variation (SIV) model is
in the middle of the room; h1 inevitably also included the validated by comparing it to the boundary element method
“direct-path signal”. In the second simulation, the sample was (BEM), which have been shown in many previous studies
removed and impulse responses h2 were measured in the same (see, e.g., [2], [17]) to yield an excellent agreement with
positions. The isolated signals reflected from the diffuser only real measurements. In addition, a staircase model of the
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 6

1
Flat surface
BEM
Directional diffusion coefficient

0.8 Staircase model


SIV model
(a)
0.6

0.4

Frequency [kHz]
0.2

0
125 250 500 1000 2000 4000
Frequency [Hz]
(b)
Fig. 4. Directional diffusion coefficient at 0o angle of incidence for a
quadratic residue diffuser of length 7 for: the boundary element method (gray
solid line), the staircase model (black dotted line), and the SIV model (black
dashed line).
Frequency [kHz]

1
Flat surface
Fig. 6. The pressure magnitude of the reflected signal from a quadratic
BEM residue diffuser of length 7 for all receiver positions located on a semicircle
Directional diffusion coefficient

0.8 Staircase model for: (a) the SIV model and (b) the staircase model.
SIV model

0.6

(a)
0.4

0.2

0
(b)
125 250 500 1000 2000 4000
Frequency [Hz]

Fig. 5. Directional diffusion coefficient at 57o angle of incidence for a


quadratic residue diffuser of length 7 for: the boundary element method (gray (c)
solid line), the staircase model (black dotted line), and the SIV model (black
dashed line).

Fig. 7. Boundary surface shapes: (a) white Gaussian noise, (b) Brownian
noise with β = 1.73, and (c) Brownian noise with β = 3.
same diffuser was implemented to investigate any potential
differences between the two FDTD implementations. The
modeled quadratic residue diffuser (QRD) is based on a
fuser wells in the SIV model, which is a possible explanation
periodic sequence of length 7 (0, 1, 4, 2, 2, 4, 1) that defines
for a smaller diffusion coefficient value.
the diffuser well depths, as illustrated in Fig. 3. The diffuser
sample was 3.6 m wide and 0.2 m deep and consisted of 6 full
period repetitions. In FDTD simulations, each well was 7 grid C. Brownian diffusers
points wide, and the maximum well depth was implemented In this section, a method for controllable diffusion using
using 16 grid points and 45.25 sample delay for the staircase fractional Brownian diffusers implemented using the SIV
and SIV models, respectively. The BEM data (up to 5 kHz) model is investigated in the frequency domain. The numerical
for the diffuser of the same size can be found in Appendix C simulations of the fractional Brownian diffusers were con-
of [17]. ducted for three sequences, namely a white-noise and two
The directional diffusion coefficient values for two angles Brownian-noise sequences, the latter resulting from spectral
of incidence (i.e., 0o and 57o ) are compared in Figs. 4 and 5, shaping of the white noise with spectral density exponent
respectively. These plots indicate a good agreement between β = 1.73 and β = 3, respectively. These three noise
the two FDTD models and the BEM, in particular at low sequences of length 101 are illustrated in Fig. 7. In successive
frequencies. At higher frequencies, small deviations from the experiments, three maximum delay values, corresponding to
BEM results are noticeable; while the SIV model generally the maximum depth of the diffuser, were set to Dmax = (5, 10,
underestimates the diffusion coefficient, the staircase model and 20) samples for all diffuser types.
tends to show overestimation. As illustrated in Fig. 6, the The diffusion effect of increasing the maximum well depth
direct comparison of pressure magnitude plots for the SIV and for a white-noise-shaped boundary is clearly visible in the
staircase models at 0o angle of incidence again indicates a very pressure magnitude plots for a set of receivers and a normal
good match between these two implementations. However, angle of incidence (i.e., source position at θ = 0o ) depicted
more energy flows (white vertical lines) develop between dif- in Fig. 8(a-c). From the combination of (7) and (10), three
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 7

1
White noise (max 5 samples)

Random−incidence diffusion coefficient


White noise (max 10 samples)
0.8 White noise (max 20 samples)
Flat surface

(a)
0.6

0.4

Frequency [kHz] 0.2

0
250 500 1000 2000 4000 8000
Frequency [Hz]

(b)
Fig. 9. Random-incidence diffusion coefficient for a white-noise-shaped
boundary with the maximum roughness given by Dmax = 5, 10, 20 samples.

1
White noise

Random−incidence diffusion coefficient


Frequency [kHz]
Brownian noise ( β=1.73)
0.8 Brownian noise ( β=3)
Flat surface

0.6

(c)
0.4

0.2

Frequency [kHz] 0
250 500 1000 2000 4000 8000
Frequency [Hz]

Fig. 10. Random-incidence diffusion coefficient for 3 boundary shapes: a


white-noise-shaped diffuser (black solid line), Brownian-noise-shaped diffuser
(d) with β = 1.73 (gray solid line), and Brownian-noise-shaped diffuser with
β = 3 (black dashed line); the maximum delay length is 10 delay samples.

Frequency [kHz] maximum delay values (i.e., Dmax = 5, 10, 20 samples) cor-
respond to the following design frequencies f0 = 8.820 kHz,
4.410 kHz, and 2.205 kHz, respectively. Fig. 8(a-c) indicates
that the specular reflection is very well spatially scattered
(e)
by a white-noise-shaped boundary for frequencies above the
respective design frequencies, and that diffusion already occurs
over an octave below f0 . Note that such a figure is much more
indicative than just a simple polar plot for one frequency as it
Frequency [kHz] enables the observation of the diffusive boundary performance
for many frequencies simultaneously. A very high diffusivity
of the white-noise-shaped diffuser is also confirmed by high
random-incidence diffusion coefficient values, as illustrated
in Fig. 9. Again, the diffusion coefficient gradually increases
(f)
starting from over an octave below f0 to peak at the design
frequency, and then remains constant.
Spectral shaping of the input noise signal to control dif-
Frequency [kHz]
fusion strength was investigated by comparing three noise-
shaped boundaries, namely the white-noise and two Brownian-
Fig. 8. The pressure magnitude of the reflected signal for all receiver noise diffusers with a constant design frequency of 4.410 kHz.
positions located on a semicircle for the following boundary surface shape: As illustrated in Fig. 8(b,d,e), contrary to the white-noise dif-
(a) white-noise shape with Dmax = 5 samples, (b) white-noise shape with
Dmax = 10 samples, (c) white-noise shape with Dmax = 20 samples, fusers for which the incident sound is scattered evenly in many
(d) Brownian-noise shape with β = 1.73 and Dmax = 10 samples, (e) directions, the Browinan shape is characterized by scattering in
Brownian-noise shape with β = 3 and Dmax = 10 samples, (f) white-noise a reduced number of reflection directions. Increasing the value
shape with Dmax = 5 samples with rounded fractional delays to integer
values. of the power spectral density, the strength of spatial scattering
decreases, which results in a lower random-incidence diffusion
coefficient value, as depicted in Fig. 10. In addition, since
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 8

−3
x 10
0o 6
−30o 30o
3

Amplitude
(a) 0
−60o 60o

−3

−6
1550 1650 1750
−90o 90o Time [samples]
−3
−60dB −40dB −20dB 0dB x 10
6

Fig. 11. Polar magnitude plot for the Brownian diffuser with Dmax = 10
samples and β = 3, at frequency f = 2.5kHz. 3

Amplitude
(b) 0
1
Random−incidence diffusion coefficient

Fractional delays
Integer delays −3
0.8 Flat surface

−6
0.6 1550 1650 1750
Time [samples]
−3
x 10
0.4 6

0.2 3

Amplitude
0 (c) 0
250 500 1000 2000 4000 8000
Frequency [Hz]
−3
Fig. 12. Directional diffusion coefficient at 0o angle of incidence for a
white-noise-shaped boundary with Dmax = 5 samples implemented using: −6
(a) fractional delays and (b) only integer delays. 1550 1650 1750
Time [samples]
−3
x 10
6

Browinan diffusers are based on random noise sequences,


3
their scattering is neither periodical nor symmetric, which is
Amplitude

exemplified in a polar plot for a Brownian diffuser with β = 3 (d) 0


at a frequency f = 2.5kHz in Fig. 11.
Finally, the effect of placing diffuser wells at sub-grid −3

locations using fractional delays in the SIV model was in-


−6
vestigated. Two boundary shapes were designed: a white-noise 1550 1650 1750
Time [samples]
shape with Dmax = 5 delays and an analogous shape obtained x 10
−3

6
by rounding the values of the original white-noise sequence so
that only integer delay values from the range 0 to 5 resulted.
3
The comparison of pressure magnitude plots in Fig. 8(a,f)
Amplitude

shows that there are small differences, which indicates that (e) 0
a staircase model may introduce noticeable diffusion errors
that are avoided in the SIV model. The models with fractional −3

and integer delay values are further compared in terms of


−6
directional diffusion coefficients in Fig 12. 1550 1650 1750
Time [samples]

Fig. 13. The signal reflected from the sample diffuser (gray solid line) plotted
D. Time spreading against the signal reflected from the flat boundary sample (black dashed line)
The FDTD method allows direct investigation of the dif- in the time domain for the following diffusers: (a) a white-noise diffuser with
Dmax = 5 samples, (b) a white-noise diffuser with Dmax = 10 samples,
fusive properties in the time domain. Examples of time- (c) a white-noise diffuser with Dmax = 20 samples, (d) a Brownian diffuser
domain analysis of sound scattering can be found in various with Dmax = 10 samples and β = 1.73, and (e) a Brownian diffuser with
previous studies (e.g., in [8], [30]), in which the significance of Dmax = 10 samples and β = 3.
including time spreading aspects as a means of characterizing
diffusers has been highlighted. Since time spreading properties
are not taken into account when calculating the diffusion Fig. 13 shows time spreading effects for the following noise
coefficient (that describes spatial scattering), such time-domain sequences: three white-noise-shaped boundaries with Dmax =
analysis forms a complementary means of analyzing the 5, 10, 20 samples, and Brownian-noise-shaped boundaries with
performance of phase grating diffusers. Dmax = 10 samples and β = 1.73 and β = 3, respectively;
TRANSACTIONS ON AUDIO, SPEECH AND LANGUAGE PROCESSING, PREPRINT VERSION 9

each time domain plot also includes a reflected sound signal [5] Y.W. Lam, “A comparison of three diffuse reflection modeling methods
from the flat boundary sample for reference. The comparison used in room acoustics computer models,” J. Acoust. Soc. Am., vol. 100,
no. 4, pp. 2181–2192, 1996.
of white-noise sequences indicates clearly that more time [6] M. Vörlander, “International round robin on room acoustical computer
spreading occurs for a larger maximum depth. Furthermore, simulations,” Proc. Int. Congress on Acoustics, pp. 689–692, June 1995,
more variation in the boundary surface shape (i.e., more spiky Trondheim, Norway.
[7] T. Yokota, S. Sakamoto, and H. Tachibana, “Visualization of sound
shape) causes increased time spreading. propagation and scattering in rooms,” Acoust. Sci & Tech., vol. 23, no.
1, pp. 40–46, 2002.
V. C ONCLUSIONS [8] J. Redondo, R. Picó, and B. Roig, “Time domain simulation of sound
diffusers using finite-difference schemes,” Acta Acustica united with
This paper proposes a new method for modeling surface Acustica, vol. 93, no. 12, pp. 611–622, 2007.
diffusion in FDTD room acoustic simulations, based on the [9] J. de Poorter and D. Botteldooren, “Acoustical finite-difference time-
domain simulations of subwavelength geometries,” J. Acoust. Soc. Am.,
idea of spatial variation of the surface impedance. This vari- vol. 104, no. 3, pp. 1171–1177, 1998.
ation is achieved by nesting fractional delay filters in the [10] S.van Duyne and J.O. Smith, “The 2-D digital waveguide mesh,” Proc.
digital impedance filters that model the local wall reflection IEEE Workshop on Applications of Signal Processing to Audio and
Acoustics, pp. 177–180, October 1993, Mohonk, NY.
properties. A key feature of the approach - that can in [11] K. Lee and J.O. Smith, “Implementation of a highly diffusing 2-d digital
principle also be applied to other wave-based methods - is that waveguide mesh with a quadratic residue diffuser,” Proc. Int. Comp.
frequency-dependent absorption can be taken into account as Music Conf. (ICMC), pp. 309–315, November 2004, Miami, Florida.
[12] S. Shelley and D.T. Murphy, “The modeling of diffuse boundaries in
it is for smooth surfaces (i.e., using a locally specified wall the 2-D digital waveguide mesh,” IEEE Trans. on Audio, Speech and
impedance), which distinguishes it from prior studies in this Language Processing, vol. 16, no. 3, pp. 651–665, March 2008.
field. [13] Simon Shelley, Diffuse Boundary Modelling in the Digital Waveguide
Mesh, Ph.D. thesis, University of York, UK, November, 2007.
The results predicted with the proposed surface impedance [14] K. Kowalczyk and M. van Walstijn, “Modeling frequency-dependent
variation (SIV) model for non-absorbing boundaries have been boundaries as digital impedance filters in FDTD and K-DWM room
shown to be in close agreement with BEM results from acoustics simulations,” J. Audio Eng. Soc., vol. 56, no. 7/8, pp. 569–
583, 2008.
literature, and the method is further validated by comparisons [15] J. Laird, P. Masri, and N. Canagarajah, “Modelling diffusion at the
with a second, more conventional FDTD model, in which the boundary of a digital waveguide mesh,” Proc. Int. Comp. Music Conf.
shape of the diffusive boundary is directly implemented in (ICMC), pp. 492–495, October 1999, Hong Kong.
[16] M.R. Schroeder, “Binaural dissimilarity and optimum ceilings for
staircase form. concert halls: More lateral sound diffusion,” J. Acoust. Soc. Am., vol.
The main advantage of the SIV model over the staircase 65, no. 4, pp. 958–963, 1979.
model is that the effective well depths can be more pre- [17] T.J. Cox and P. D’Antonio, Acoustic Absorbers and Diffusers: Theory,
Design and Application, Spon Press, 2004.
cisely defined and adjusted, making it particularly suitable [18] T.J. Cox and P. D’Antonio, “Fractal sound diffusers,” 103rd Convention
for modeling rough surfaces and narrow-well diffusers with Audio Eng. Soc., September 1997, Preprint 4578, New York, NY.
thin separators, and generally providing a finer control over [19] J.C. Strikwerda, Finite Difference Schemes and Partial Differential
Equations, Wadsworth & Brooks, Pacific Grove, CA, 1989.
the diffusive properties. Another advantage is the lower com- [20] S. Bilbao, Wave and Scattering Methods for Numerical Simulation, John
putational cost required when modeling diffusive walls with Wiley & Sons, London, 2004.
frequency-dependent absorption, as less digital impedance [21] M. Karjalainen and C. Erkut, “Digital waveguide vs. finite difference
schemes: Equivalence and mixed modeling,” EURASIP J. Applied Signal
filters are required. Finally, the SIV model is easier to im- Processing, , no. 7, pp. 978–989, June 2004.
plement, as most boundaries can be treated as straight walls [22] M.E. Delany and E.N. Bazley, “Acoustical properties of fibrous
with varying boundary filters, thus avoiding the complexities absorbent materials,” J. Applied Acoustics, vol. 3, pp. 105–116, 1970.
[23] K. Kowalczyk and M. van Walstijn, “Formulation of locally reacting
of implementing a multitude of corners. surfaces in FDTD/K-DWM modelling of acoustic spaces,” Acta Acustica
Furthermore, the SIV model has been employed to investi- united with Acustica, vol. 94, no. 6, pp. 891–906, 2008, Special issue
gate diffusive walls defined by spectral shaping of Gaussian on Virtual Acoustics.
[24] J. Angus, “Using grating modulation to achieve wideband large area
noise; the results indicate that this approach is a promising diffusers,” Appl. Acoust., vol. 60, pp. 143–165, 2000.
way of parameterizing the diffusive properties of the modeled [25] M. van Walstijn and K. Kowalczyk, “On the numerical solution of the 2D
surface, and could potentially be used to simulate scatter- wave equation with compact FDTD schemes,” Proc. Int. Conf. Digital
Audio Effects (DAFx’08), pp. 1–8, September 2008, Espoo, Finland.
ing behavior that roughly matches a given set of diffusion [26] V. Välimäki, Discrete-Time Modeling of Acoustic Tubes using Fractional
coefficient data. Indeed, finding a precise way of mapping Delay Filters, Ph.D. thesis, Helsinki University of Technology, Finland,
diffusion data to Brownian diffuser model parameters would 1995.
[27] T.I. Laakso, V. Välimäki, M. Karjalainen, and U.K. Laine, “Splitting
be an interesting area of future research. the unit delay - tools for fractional delay filter design,” IEEE Signal
Processing Magazine, vol. 13, no. 1, pp. 30–60, January 1996.
R EFERENCES [28] J.O. Smith, Physical Audio Signal Processing for Virtual Musical Instru-
ments and Digital Audio Effects, W3K Publishing, http://books.wk3.org,
[1] U.M. Svensson, “Modelling acoustic spaces for audio virtual reality,” August 2006.
Work. on Model based Proc. and Coding of Audio (MPCA), pp. 109– [29] AES 4id 2001(r2007), “AES information document for room acoustics
116, November 2002, Leuven, Belgium. and sound reinforcement systems - characterisation and measurement of
[2] T.J. Cox and Y.W. Lam, “Prediction and evaluation of the scattering surface scattering uniformity,” J. Audio Eng. Soc., vol. 49, pp. 149–165,
from quadratic residue diffusers,” J. Acoust. Soc. Am., vol. 95, no. 1, November 2001.
pp. 297–305, 1994. [30] T.J. Cox, B.-I.L. Dalenbäck, P.D. Antonio, J.J. Embrechts, J.Y. Jeon,
[3] T.J. Cox and Y.W. Lam, “Evaluation of methods for predictiong E. Mommertz, and M. Vörlander, “A tutorial on scattering and diffusion
scattering from simple rigid panels,” J. Acoust. Soc. Am., vol. 95, no. coefficients for room acoustics surfaces,” Acta Acustica united with
1, pp. 297–305, 1994. Acustica, vol. 92, no. 1, pp. 1–15, 2006.
[4] H. Kuttruff, Room Acoustics, Applied Science Publishers Ltd, London,
1973.

You might also like