[Damian Murphy, Antti Kelloniemi, Jack Mullen, and Simon Shelley]

© DIGITALVISION
Acoustic Modeling
Using the Digital
Waveguide Mesh
[Recent activities in articulatory vocal tract modeling, room
impulse response synthesis, and reverberation simulation]

T
he digital waveguide mesh (DWM) is a numerical context of discrete-time physics-based modeling for sound syn-
simulation technique based on the definition of a thesis, and [3], where the DWM is examined within a rigorous
regular spatial sampling grid for a particular prob- theoretical and comparative framework for more established yet
lem domain, which in this specific case is a vibrating related wave scattering numerical simulation techniques.
object capable of supporting acoustic wave propaga-
tion resulting in sound output. It is based on a simple and intu- THE ONE-DIMENSIONAL DIGITAL WAVEGUIDE
itive premise—the latter often considered important by the The one-dimensional (1-D) digital waveguide is based on a time
computer musicians who are the primary users of a sound syn- and space discretization of the d’Alembert solution to the 1-D
thesis algorithm—yet the emergent behavior is complex, natu- wave equation. This approach to sound synthesis was first used
ral, and capable of high-quality sound generation. Hence, the in the Kelly-Lochbaum model of the human vocal tract for
DWM has been applied in many areas of computer music speech synthesis [4] and has parallels with other, more generally
research since it was first introduced by Van Duyne and Smith applied wave variable scattering modeling paradigms such as the
in 1993 [1]. This article is the first to attempt to consolidate and transmission line matrix (TLM) method [5] and wave digital fil-
summarize this work. The interested reader is also directed to ters (WDFs) [6]. However it was Julius O. Smith III who first
[2], where DWM modeling is considered in the more general proposed the term digital waveguide and used these techniques

1053-5888/07/$25.00©2007IEEE IEEE SIGNAL PROCESSING MAGAZINE [55] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
initially for artificial reverberation [7] and later for sound syn- ming the upper and lower values according to (3), being exact at
thesis [8], [9]. Digital waveguides have remained the most popu- the sampling points of the system.
lar and successful physical modeling-based sound synthesis
technique to date, due to the realistic, high-quality sounds that THE SCATTERING JUNCTION
can be generated, often in real-time and so therefore also facili- The terminations introduced in the 1-D string shown in Figure 1
tating effective user interaction. This research has also been are a special case of signal scattering. An input signal will propa-
made more widely accessible through a range of commercially gate without loss until it is incident upon a change in system
available physical modeling hardware synthesizers developed by impedance, resulting in transmission and/or reflection of the
Yamaha in the early 1990s based on digital waveguide tech- incident signal. This example leads to the formal definition for a
niques [10]. The reader is referred to [9] and [11] for a thorough lossless scattering junction, now given without loss of generality
treatment and discussion of this area and a full derivation of in terms of acoustic pressure rather than string displacement.
some of the equations that are introduced in what follows. At such a junction, system continuity must be preserved in
Consider the 1-D wave equation for transverse motion with terms of pressures and volume velocities analogous to
speed c on an ideal, infinitely long, vibrating string: Kirchhoff’s Laws for parallel connection of electrical circuit ele-
ments. Assuming N connected waveguide elements with the
∂ 2 y(t, x) 2
2 ∂ y(t, x)
= c . (1) pressure in each defined as pi and volume velocities as ui , then
∂ t2 ∂ x2
for lossless scattering the following must hold:
The d’Alembert or traveling wave solution to (1) is defined as:
p1 = p2 = . . . = pi = . . . = pN = pJ, (4)
+ − u1 = u2 = . . . = ui = . . . = uN = 0
y(t, x) = y (t − x/c) + y (t + x/c), (2) (5)

where y+ and y− are arbitrary twice-differentiable functions Note that pJ is defined as the actual pressure value at the point of
denoting wave movement to the left and right, respectively. connection for these N waveguide elements, referred to as the
Assuming that y+ and y− are bandlimited to half the sampling pressure value at scattering junction J. Scattering junctions,
rate of the system allows the discrete time version of (2) to be together with the 1-D waveguide elements described above, pro-
defined for spatial sampling points mX and sampling interval vide the basic building blocks for a digital waveguide physical
nT such that X = cT : model of a vibrating system. For instance, six 1-D strings could
be coupled together via a scattering junction to simulate the
y(nT, mX) = y+ (n − m) + y− (n + m). (3) bridge of a guitar, facilitating sympathetic resonances where
excitation on one string causes low-amplitude oscillation on one
This solution can be implemented in an efficient and straight- or more of the others due to energy transmitted through the
forward manner using two parallel digital delay lines to repre- bridge. Similarly, this scattering junction could also allow cou-
sent the left-going and right-going traveling wave components. pling to a filter to simulate the effects of body resonances. Hence,
Figure 1 shows a digital waveguide implementation of an ideal scattering junctions also act as system sampling points where
string, rigidly terminated at either end of the M-sample delay physical variables may be tapped off for coupling with other
lines, corresponding to the nut and bridge of a typical instru- aspects of the model or with the outside world. Similarly, they
ment. The system is excited with an appropriate input “loaded’’ can also be used to allow energy to be input to a system. In mod-
into the upper and lower delay lines at position xin = min cT eling a wind instrument such as a clarinet, the bore can be
and a physical output is obtained at xout = mout cT by sum- implemented as a 1-D lossless waveguide coupled with the more

M Sample Delay Line

y+(n) z−1 z−1 z−1 z−1 z−1 z−1 y+(n − M)

0.5
Nut + Bridge
y(nT, mincT) y(nT, moutcT)
Termination −1 −1 Termination

y−(n) z−1 z−1 z−1 z−1 z−1 z−1 y−(n + M)

x=0 x = McT

[FIG1] The ideal lossless 1-D digital waveguide string model, which is M-samples long and rigidly terminated at either end.

IEEE SIGNAL PROCESSING MAGAZINE [56] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
complex, nonlinear breath pressure/reed input function
via an appropriate scattering junction implementation [9].
2Y1 2YN
Figure 2(a) shows the functional block diagram for a J
− −
general lossless scattering junction J with N neighbors, + +
+
with each connected unit waveguide element having an
2Y2 + +
associated admittance Yi . The impedance of a waveguide pJ, i pi, J
− 1
is given by Z i = pi /ui and hence the admittance + ΣYi 2Yi z−1
Yi = 1/Z i . The signal p+J,i represents the incoming signal −
to junction J along the waveguide from the opposite junc- + z−1

tion i. Similarly, the signal p−J,i represents the outgoing pJ

pJ, i− pi, J−
signal from junction J along the waveguide to the oppo-
site junction i. Connecting delay lines together at scatter- (a)
ing junctions in a more general sense allows spatial and
temporal sampling grids to be defined and gives rise to
2Y1 J 2YN
families of models that are more generally known as digi-
tal waveguide networks (DWNs). The Kelly-Lochbaum +
vocal tract model and the simply terminated 1-D string as
2Y2
shown in Figure 1 are both examples of specific DWNs. A 1
DWN with a more complex arrangement of multiport ΣYi
2Yi
interconnections can be used to simulate reverberation, as + −
in the first application of digital waveguides [7] and more pJ
recently explored in [12]. However a DWN consisting of z−1 z−1
(typically) unit delay waveguide elements and N-port loss-
less scattering junctions conforming to a regularly (b)
arranged and spaced grid structure gives rise to a particu-
[FIG2] Functional block diagrams for the general lossless scattering
lar family of two- or three-dimensional (2-D or 3-D) struc-
junction J with N neighbors: (a) the W-model case and (b) the K-model
tures. These are called digital waveguide meshes and are case. Note that in each example a single connecting waveguide element
more directly analogous in construction to the physical has been connected to terminal Yi .
objects they are attempting to simulate.
previous time step, p−
i, J (n − 1). Expressing this relationship in
THE DIGITAL WAVEGUIDE MESH the z-domain gives:
The DWM was first proposed by Van Duyne and Smith [1] as an
+
extension to 1-D digital waveguide sound synthesis appropriate PJ,i = z−1 ·P−
i, J . (8)
for modeling plates and membranes, potentially leading to full
3-D object modeling. Acoustic wave propagation through a Hence, from (6) junction pressure values are calculated accord-
DWM is determined according to the scattering equations and ing to input values from immediate neighbors, output values are
associated mesh topology. For a lossless junction J according to calculated using (7) and then propagated to neighbors via the
conditions (4) and (5) or directly from Figure 2(a), the sound bi-directional waveguide elements, becoming inputs at the next
pressure pJ at junction J for N connected waveguides can be iteration according to (8). From (6), (7), and (8) via an appropri-
expressed as: ate linear transformation it is possible to derive an equivalent

N formulation in terms of junction pressure values only:
2 i =1 Yi · pJ, i+
pJ =
N . (6)
N
Yi 2 Yi · pi · z−1
i =1
i =1 pJ =
N − pJ · z−2 . (9)
Y
i =1 i
Noting from (3) that the total sound pressure pJ in a waveguide
element connected to junction J can also be defined as the sum Expression (9) can also be derived directly from a finite differ-
of the traveling waves in this element, or alternatively as the ence time domain (FDTD) formulation of the 2-D case of the
sum of the input and output gives: wave equation in (1). The functional block diagram for the scat-
tering junction implementation described by (9), equivalent to
pJ = p+J,i + p−J,i . (7) Figure 2(a), is shown in Figure 2(b). Digital waveguide models
represent signal propagation via two directional wave compo-
And finally, as the waveguide elements in a DWM are equivalent nents and schemes implemented in this way, according to (6),
to bi-directional unit-delay lines, the input to scattering junc- (7) and (8), are termed W-models or W-DWMs [2], [13], [14]. A
tion J at time index n, p+J,i (n) is equal to the output from linear transformation of a W-DWM leads to this alternative
neighboring junction i into the connecting waveguide at the implementation as a Kirchhoff variable DWM (K-DWM) [2],

IEEE SIGNAL PROCESSING MAGAZINE [57] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
[13], [14], as given in (9), and depending on physical quantities Figure 4(c)–(f) include the rectilinear [23]; tetrahedral [24],
only rather than sampled traveling-wave components. In this [25]; dodecahedral [also known as cubic close packed (CCP)]
form, and under certain conditions, a K-DWM can be computa- [26]; and octahedral structures, and a similar analysis of their
tionally equivalent to a FDTD simulation. characteristics is presented in [27].
Mixed modeling scenarios where K-DWM and W-DWM Three-dimensional DWM structures are applied to a range of
approaches have been interfaced in 1-D via a KW-pipe have sound synthesis applications. The work of [28] combines a 2-D
been proposed in [14], [15], leading to the formulation of a 2-D triangular mesh model of a drum membrane coupled to a 3-D
hybrid DWM [13], [16], [17]. The 2-D hybrid mesh combines rectilinear model of a drum-shell to give a more complete model
the computational efficiency of the K-DWM approach in terms of a percussion instrument. DWM models have been applied to
of computation time and memory use, with the flexibility of simulate 3-D resonant objects [26], [29], [30], sometimes in
scattering-based boundary termination options for complex combination or parallel with other digital waveguide models; for
geometries through the use of KW-pipes. Typically, KW-hybrid instance, to provide synthesis of complex instrument resonances
DWMs demonstrate a speed up in processing time of the order [31], [32], or to simulate a 3-D acoustic space with multiple 2-D
of 34% with a 50% decrease in use of main system memory cross-sectional simulations [33]. However, most current
[16]. A K-DWM scattering junction connected to a W-DWM research activity in 3-D DWM modeling is in its application to
scattering junction via a KW-pipe is shown in Figure 3. the accurate synthesis of acoustic spaces, and this will be dis-
The W-DWM or K-DWM scattering equations can be used to cussed in the Applications section.
implement a range of topologies/structures. In 2-D the most An additional subset of K-DWMs has also been subject to much
commonly implemented topologies are the four-port rectilinear investigation and these are based on an interpolated rectilinear
and six-port triangular mesh structures shown in Figure 4(a) mesh structure in either 2-D [34] or 3-D [35]. Interpolated DWMs
and (b). A thorough comparison of their relative characteristics, demonstrate wave propagation characteristics approaching that of
together with those of the three-port hexagonal mesh, is pre- triangular/dodecahedral topologies but without the additional
sented in [18]. Two-dimensional DWM models based on the rec- overheads of a denser and more complex topological structure.
tilinear or triangular topology have been most commonly used
for synthesis of percussion instruments such as plates, mem- DWM LIMITATIONS
branes, and gongs [19]–[21], as well as for 2-D reverberation There are a number of important factors that impose limitations
modeling [22]. Three-dimensional topologies as shown in on DWM models as an optimal solution for all sound synthesis

2Y1 2YN 2YN+1 2YM

J K
− −
+ +
+ +

2Y2
− 1
pJ, i +
1
+ ΣYi ΣYj
2Yi + 2Yi 2Yj
− −
+ −
+ + z−1 pk
pJ, 1−
z−2 z−1 z−1
pJ

[FIG3] Functional block diagram for a W-DWM scattering junction J with N neighbors connected to a K-DWM scattering junction K via a
KW-pipe connecting waveguide element.

(a) (b) (c) (d) (e) (f)

[FIG4] DWM topologies: (a) four-port 2-D rectilinear; (b) six-port 2-D triangular; (c) six-port 3-D rectilinear; (d) 12-port 3-D dodecahedral
(CCP); (e) four-port 3-D tetrahedral; (f) eight-port 3-D octahedral.

IEEE SIGNAL PROCESSING MAGAZINE [58] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
applications. One of the most significant advantages of the 1-D terminated with single one-pole all-pass filters, which may be
digital waveguide that originally made it a realistic proposition interpreted as a 1-D termination connected to an ideal spring,
for applications in sound synthesis is the computational efficien- allowing modal frequencies in the DWM to be re-tuned or cor-
cy of the approach when compared with a brute-force numerical rected appropriately. For curved boundaries, where the perime-
solution to the system wave equation. This is further improved ter of the structure being modeled is not normal/parallel to the
through the ability to commute losses to specific lumped points axes of the mesh, noninteger length waveguide elements called
in the system, significantly reducing the number of calculations rimguides can be used [20] and have been demonstrated as
required per time-step iteration. Unfortunately, the elegance of appropriate for accurate low-frequency modeling of circular
this approach is lost when moving to higher dimensions. With a membranes using a 2-D triangular mesh.
DWM-based system, acoustic wave propagation is determined by A commonly applied solution is to passively terminate a
signal interaction at the scattering junctions and hence a calcu- DWM using a simple 1-D connection that implements a change
lation must take place at every junction for every time-step. in admittance such that there is no signal return from the con-
Reducing the number of scattering junctions reduces the sam- nected boundary over a finite time duration. Hence, the associ-
ple rate of the DWM and hence the effective bandwidth of the ated input value for such a connection in Figure 2, or using (6)
system. The advantage gained with the DWM approach, howev- or (9), is set to zero. This termination acts to reflect an incident
er, is in the structural immediacy of the simulation, allowing signal according to the change in admittance of the connected
objects to be defined based only on physical and geometrical def- waveguide elements. In the simplest case, for a one-port bound-
initions, and the ability to observe and interact with the system ary-node pB connected to a single N-port scattering junction p1
at physically relevant or meaningful points. with a change in waveguide admittance from Y to YB, a reflec-
A more specific DWM limitation is dispersion error, where tion coefficient −1 ≤ r ≤ 1 is determined such that:
the velocity of a propagating wave is dependent upon both its
Y − YB
frequency and direction of travel, leading to wave propagation r= (11)
Y + YB
errors and a mistuning of the expected resonant modes. The
degree of dispersion error is highly dependent upon mesh topol- and pB can therefore be calculated as a function of the sound
ogy and has been investigated in, for example, [3], [18], [24], pressure of the incident traveling wave variable from p1 :
[27]. In 2-D both the interpolated and triangular DWMs demon-
strate dispersion characteristics that are substantially reduced to pB = (1 + r) · pB,1+ . (12)
a function of frequency only. In 3-D, minimization of dispersion
can be similarly achieved through the use of interpolated or In the equivalent K-DWM case, a passive termination is equiva-
dodecahedral topologies. Appropriate pre- and postprocessing of lently implemented as a feedback loop between waveguide ele-
results from these mesh structures allows offline frequency ment terminals in Figure 2 with unit delay as derived in [13],
warping techniques to be used to correct mis-tuned resonances [38] and given by
[34], [35]. Alternatively, frequency warping can be incorporated
directly as part of a DWM scattering junction [36], [37]. pB = (1 + r)p1 · z−1 − r · pB · z−2 . (13)
However, although accurate synthesis of resonant modes is
required for the dominant low-frequency properties of a vibrat- Note that r = 1 or r = −1 gives total reflection and r = 0
ing system, dispersion error is considered less important with approximates anechoic conditions. Full derivations of boundary
increasing frequency as modal density increases, and human conditions for the general N-port boundary termination for K-,
perception of such variations becomes less critical.Oversampling W-, and KW-hybrid cases is offered in [16] and a similar bound-
a DWM can also offer improvements such that the required ary implementation for a triangular DWM using multiport
bandwidth lies within accepted limits, typically 0.25 × fupdate reflection factors is presented in [29].
[1], where fupdate for a DWM of dimension D and spatial sam-
pling distance d is generally given by: APPLICATIONS OF THE DWM
√ The digital waveguide mesh in 2-D and 3-D has been applied to a
c D
fupdate = , (10) diverse range of applications where simulation of acoustic wave
d
propagation within an enclosed system is required. What follows
where c is the speed of sound. Ultimately fupdate dictates the is a summary of recent results and research in this area, namely
quality of audio signal output from a DWM with large sample for vocal tract synthesis, object modeling, synthesis of room
rates requiring denser meshes, more computer memory, and impulse responses, and how this method can be extended to
hence taking longer to run, limiting even the most efficient abstract higher dimensions.
large-scale K-DWMs to offline generation only.
2-D VOCAL TRACT MODELING FOR SPEECH SYNTHESIS
DWM BOUNDARY TERMINATION The well-established 1-D Kelly-Lochbaum vocal tract model [4]
There exist a number of possibilities for terminating a DWM at a is based on a linear series of concatenated cylindrical acoustic
system boundary. In [21] a 10 × 10 node 2-D rectilinear DWM is tubes, each of different cross-sectional area, and each tube

IEEE SIGNAL PROCESSING MAGAZINE [59] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
section implemented as a 1-D digital waveguide element. A area function information. A minimum impedance channel
number of developments on this basic model include nasal tract, Zmin is defined as the lowest value across the range of vowels
lip radiation, and wall losses to synthesize the singing voice [39] to be simulated, corresponding directly to the largest cross-
and the use of fractional waveguides to make lengthwise sectional area Amax , and from this a maximum tract width
changes to tract shape [40], opening can be defined. An
[41]. Standard waveguide ele- impedance map is constructed
ments have also been substitut- DIGITAL WAVEGUIDES HAVE for a particular vowel tract
ed for conical equivalents using REMAINED THE MOST POPULAR shape such that each area func-
scattering methods derived AND SUCCESSFUL PHYSICAL tion value A(x) along the
from the spherical wave equa- MODELING-BASED SOUND SYNTHESIS length of the tract walls corre-
tion. This increases model TECHNIQUE TO DATE, DUE TO THE sponds to a maximum imped-
accuracy, giving higher-order REALISTIC, HIGH-QUALITY SOUNDS ance value Z x . An impedance
area function approximation, curve varying from Z x to Zmin
THAT CAN BE GENERATED.
but adds to the computational and back to Z x at the opposite
load and introduces possible wall is then defined across the
stability problems [40], [42]. More recent work has explored the tract according to a raised cosine function, with the mini-
possibility of replacing the basic 1-D digital waveguide imple- mum impedance channel equidistant between the tract walls.
mentation with a 2-D DWM model that simulates the variation Figure 5(a) shows the cross-sectional area function informa-
in cross-sectional area along the vocal tract directly through an tion A(x) taken from MRI scans [45] as it varies along the length
appropriately shaped mesh geometry [43]. Formant patterns of the vocal tract from glottis to lips. Figure 5(b) is the corre-
produced using the 2-D DWM implementation are equivalent to sponding impedance map imposed across and along the under-
those produced by a very high-resolution 1-D digital waveguide lying rectangular 2-D DWM based on a four-port rectilinear
acoustic tube-based simulation. The 2-D model also offers simu- topology. Areas of higher impedance are represented by a lighter
lation of cross-tract modes due to the additional dimension of shading, and the minimum impedance channel can be observed
freedom for acoustic oscillation and propagation and approxi- as the darker area along the center of the map. Figure 6(a)
mately linear control over formant bandwidths via the addition- shows the resulting formant pattern for this vocal tract shape,
al reflection parameter at the side walls of the vocal tract. when excited by a noise source at the glottis and measured at
Hence, the 2-D DWM vocal tract offers improvements similar to the lip end. The dotted lines have been generated from a high-
other developments based on enhanced-order acoustic tube area resolution 1-D waveguide model, using the same area functions
function approximation, together with additional model flexibil- for comparison purposes, and measured average formant values
ity such as the ability to simulate a split in the air channel used are also shown.
in the creation of sounds such as /l/. Software developed to test the real-time dynamic behavior of
The disadvantages of this proposed voice synthesis mecha- this 2-D DWM vocal tract model is available for download and
nism rest in its inability to simulate smooth, continuous use at [47], and initial results based on this system were first pre-
dynamic changes to the tract area functions to facilitate artic- sented in [44]. This application also facilitates real-time dynamic
ulated voice synthesis, and the high mesh sample rate fupdate articulation. An example is presented in Figure 6(b) demonstrat-
required to ensure accurate tracking and mapping of vocal ing a smooth linear interpolation between area function data for
tract shape resulting in an implementation that can only work the /a/ (“bard’’) and /e/ “bed’’) vowels under noise source excita-
offline. Both of these problems have been analyzed and a solu- tion to highlight the resulting change in formant patterns.
tion proposed in a new implementation of the 2-D DWM [44]. Figure 6 shows that this new 2-D dynamically varying DWM
In this new method, rather than mapping acoustic tube area demonstrates results in terms of simulated formant frequencies
function directly to the 2-D DWM geometry, a constant-width that are in good agreement with both a high-resolution 1-D
2-D rectangular DWM, 17.5 cm long with fupdate = 44.1 kHz model and real-world values. Figure 6(a) seems to indicate that
is used. The waveguide element impedance across the width the 2-D model is closer to real-world formant values than the
of this rectangular geometry is then varied according to the high-resolution 1-D case, although this accuracy actually varies

Zmax
A(x)

x
Glottis Lips
Zmin
(a) (b)

[FIG5] Forming the impedance mapped /u/ vowel DWM: (a) cross-sectional area function; (b) rectilinear mesh with raised cosine
impedance map.

IEEE SIGNAL PROCESSING MAGAZINE [60] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
with target vowel/tract-shape and the real-world values used. In component of most musical instruments, serving to both ampli-
general, the two modeling methods are in good agreement with fy and modify the characteristics of a source excitation. Given
one another. Also, Figure 6(b) demonstrates a smooth transition that the resonating aspects of most instrument bodies are rela-
between vocal tract shapes without any discontinuity for the 1-D tively small implies that a high-resolution DWM implementa-
and 2-D cases, being of particular importance in the latter exam- tion is feasible—in real-time in the case of the vocal tract model
ple. Hence, from these generally comparable results the 2-D above—with modern computing facilities. Consider the classic
dynamically varying DWM can be seen to offer an alternative to example of a 2-D ideal stretched circular membrane. The reso-
current 1-D dynamic vocal tract models, while also offering nant frequencies fmn can be defined according to the nature of
additional advantages over these 1-D implementations as dis- their nodal regions where m represents the number of nodal
cussed above and presented in [43] for static simulations of the lines positioned along the diameter of the membrane and n rep-
vocal tract. The reader is invited to test the software presented resents the number of circular nodal lines, including the bound-
in [47] and compare the audio output from both 1-D and 2-D ary. The fundamental frequency of an ideal membrane f01 can
models under LF glottal source excitation. Informal perceptual be calculated according to its physical properties (for example,
testing has demonstrated that users consider the 2-D example to as presented in [19]). Subsequent modes are fixed relative to f01 .
be more “natural-sounding’’ than the similar 1-D case. Figure 7(a) shows the nodal regions of an ideal stretched cir-
This work is the first demonstration of a dynamically varying cular membrane with diameter 0.5 m, implemented using a
DWM model, in this case operating in real-time. Most prior highly oversampled 2-D triangular DWM, with fupdate = 192
DWM work has been based on a static representation of the kHz, resulting in a spatial sampling distance of 0.00253 m and a
acoustic system under study, partly due to the computational total of 35,742 junctions. The membrane is excited near the
resources required for real-time implementation and user inter- boundary with a lowpass-filtered impulse and an output is
action, and partly due to possible discontinuities in the output obtained at a junction near the opposite boundary. To model an
from the resulting model. Hence, this work potentially opens ideal membrane with clamped edges, the reflection coefficients r
new areas of research and application areas for DWM modeling, at the boundary of the mesh are set to −1. The modes (0,2),
possibly moving to direct user-input and feedback that has cur- (1,1), (2,1), and (3,1), are shown, with associated frequencies
rently only been possible in 1-D digital waveguide synthesis. given relative to f01 . Figure 7(b) plots the spectrum of the out-
Further work in this area will concentrate on developing appro- put against the theoretical predicted frequencies for the funda-
priate tract wall boundary filters and facilitate lengthwise shape mental and first nine modes.
changes for modeling lip protrusion required for accurate syn- Note that from Figure 7(b) there is an exact correlation
thesis of the /u/ vowel. This work also demonstrates the poten- between the predicted modal frequencies and those obtained via
tial of moving toward a full 3-D DWM model using 3-D MRI simulation, and this is due to the high mesh sample rate used,
scans of the vocal tract shape incorporating complex-shape minimizing dispersion error effects for the bandwidth studied,
cross-sectional area data. and ensuring a smooth mesh fit to the circular boundary of the
membrane without using rimguides.
2-D OBJECT MODELING An exciting possibility with physical modeling synthesis is
The DWM is often used to synthesize the acoustic properties of a that, with clear defined rules governing system behavior, it
2-D or 3-D resonant body as these objects are a fundamental becomes relatively straightforward to extend these rules to

/a/ /e/
−20 5

f1 f2 f3 f4 4
−40

3
f/kHz

−60
dB

−80 Average 1-D Formants

2-D Model 1
1-D Model
−100 0
0 1 2 3 4 5 0 0.25 0.5
kHz t/s
(a) (b)

[FIG6] Formant patterns from the impedance-mapped DWM under noise excitation: (a) /u/ vowel compared with a high resolution 1-D
model and average measured values; (b) /a/ to /e/ diphthong compared with same 1-D model.

IEEE SIGNAL PROCESSING MAGAZINE [61] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
situations that could not exist, or are difficult to control, in the aspects such as coupling a 2-D membrane to a 3-D resonator
real world. Figure 8 presents one such example as an extension [28], improved nonlinear excitation [49], and using simple
of the 2-D circular membrane and shows the two lowest modes DWM resonators to model the high-frequency characteristics of
of resonance from a DWM membrane simulation of a tri-foil complex instrument bodies [31]. Also of note is the Sounding
radiation symbol with diameter and fupdate as before, this time Object Project that has explored physical modeling, including
resulting in a model consisting of 25,549 junctions. Sound 2-D and 3-D DWMs of resonating objects, with a view to match-
examples for these simple 2-D objects are available at [48]. In ing the perception of synthesized sounds to the modeled
isolation the sounds produced from such basic 2-D membranes, objects that created them [26], [29], [50].
although percussive in nature, are somewhat uninspiring and
require a more complete model for accurate and interesting SYNTHESIS OF ROOM IMPULSE RESPONSES
object synthesis and, hence, these examples should be consid- The first application of DWMs in the field of room acoustics sim-
ered as a starting point only. Further research in improved ulation was by Savioja et al. in 1994 [23]. Fundamentally, syn-
modeling of resonant objects has considered DWNs for more thesizing the characteristics of a bounded space using a DWM is
complex theoretical multidimensional systems [3], specific exactly the same as synthesizing the sound of a vibrating

−10
f
f21 f31 f12 f11 03
f11 f22 f51
−20 f02
f11 =1.59 f01
−30
Power (dB)

−40
f21 = 2.14

−50

f02 = 2.30 −60

−70

f31 = 2.92
−80
0 500 1,000 1,500 2,000
Frequency (Hz)
(a) (b)

[FIG7] 2-D triangular DWM model of a membrane with a diameter of 0.5 m: (a) animation captures from the resulting simulation
demonstrating resonance at modes (0,2), (1,1), (2,1), and (3,1); (b) actual output spectrum compared with predicted modal frequencies.

10

−10

−20
Power (dB)

−30

−40

−50

−60
−70
−80
2,600 2,800 3,000 3,200 3,400 3,600 3,800 4,000
Frequency (Hz)
(a) (b)

[FIG8] 2-D triangular DWM model of a physically impossible system—a trifoil radiation symbol membrane with a diameter of 0.5 m:
(a) animation captures from the resulting simulation demonstrating modal resonances; (b) output spectrum.

IEEE SIGNAL PROCESSING MAGAZINE [62] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
physical object. However, in the latter example, sound output is reflection is equal to the angle of incidence, or diffuse such that
generated directly from the modeled object by reading sample the incident energy is redistributed over a range of angles.
values at a scattering junction, whereas with room acoustics Previous diffuse boundary implementations for a DWM are effec-
modeling it is a room impulse response (RIR) that is synthesized tive but limited, either in terms of accuracy [20], or by sacrific-
rather than the actual sound source. The RIR is generally of lit- ing user control for an optimal solution [59]. A new technique
tle interest in terms of its direct audio quality; however, when based on [20] simulates accurate diffusion with a high degree of
convolved with an arbitrary anechoic audio input the result is to control and consistency by rotating incoming junction signals
perceive the audio source as if placed within the modeled space. via a circulant matrix at a diffusing layer of standard N-port W-
Also, the relative size of the DWMs used as 3-D acoustic spaces DWM air-nodes adjacent to the boundary [60]. The model is
are many times larger than, for example, the 2-D vocal tract pre- lossless and allows other boundary conditions, such as frequen-
sented above, and hence take considerably longer to execute, cy-dependent absorption, to be easily incorporated.
implying offline RIR synthesis only. Much of this recent application-focused research has been
It has been shown that DWMs offer accurate RIR synthesis at incorporated as part of the RoomWeaver DWM-based room
low frequencies [51] and demonstrate natural wave phenomena acoustics research tool first presented in [17] and shown in
such as interference and diffraction [52], with high-frequency Figure 9. The purpose of the system is to allow the user to intu-
accuracy being limited by fupdate and the dispersion error of the itively set up enclosed space geometry, boundary surface, and
selected topology. This contrasts with other RIR synthesis meth- source/receiver parameters required to generate an RIR by
ods based on geometric acoustic techniques [53] that are typi- means of a simple scripting language and graphical user inter-
cally valid for high frequencies only. Other research has face. High-quality reverberation and auralization for a wide
explored how 3-D spaces, or their reverberant characteristics, range of spaces/applications are possible using high-resolution
might be simulated using 2-D models, significantly reducing 2-D triangular and 3-D mesh topologies both based on a KW-
computational resources [22], [33]. However, the accurate simu- hybrid implementation. A range of RIRs synthesized according
lation of DWM boundaries is still a key research area with a view to varied initial conditions and associated audio examples are
to how the physical properties of real materials might be mod- available for download [48]. For complete synthesis of a sound
eled. This has included how 1-D boundary termination might be event it would be desirable to incorporate a 3-D dynamically
optimized for anechoic conditions [54] [55], leading to a new variable instrument model within a larger 3-D DWM of a per-
spatially averaged approach for the 0 ≤ r ≤ 1 case [56]. formance space, requiring appropriate interfacing across DWM
Optimized absorption/reflection across a wide range of angles of types according to the space and instrument models used.
incidence for −1 ≤ r ≤ 1 has been facilitated using the admit- Although nontrivial, this has been considered in the case of a
tance boundary method [57], where a DWM is terminated with drum model using a 2-D triangular DWM membrane with a 3-D
additional layers of boundary-nodes behind the actual boundary rectilinear shell [28] and in the more general case [3]. However,
location, in turn terminated using an optimal anechoic solution. such complete synthesis could only be offered via the offline RIR
For accurate simulation of real acoustic boundaries, frequen- generation/real-time convolution processing paradigm due to
cy-dependent reflection/absorption must be implemented. In the computational expense of full 3-D space modeling. Such
[58], a boundary-node is replaced with a boundary filter defined examples are presented in [48] for a 2-D DWM vocal tract
to optimally match given frequency-dependent material reflec- processed with DWM-synthesized RIRs.
tion coefficients and implemented using a first-order infinite
impulse response filter for a 2-D rectilinear K-DWM. This THE HYPER-DIMENSIONAL DIGITAL WAVEGUIDE MESH
results in a good approximation to the target response but is From (6) it is clear that dimensionality is not inherent in the
subject to the directional-dependent characteristics of the mesh scattering equations. For example, the four-port lossless scat-
topology, being less accurate for certain angles of incidence. tering junction is the main algorithmic building block of both
The other important characteristic of a real-world acoustic the 2-D rectilinear and 3-D tetrahedral mesh. The spatial
boundary is whether a reflection is specular, where the angle of arrangement of the surrounding scattering junctions—the

[FIG9] DWM room acoustics modeling in RoomWeaver incorporating the latest implementations of frequency-dependent and diffusive
boundaries. A completely defined acoustic space followed by wave propagation snapshots through a 2-D horizontal plane of the same
space viewed in wireframe mode is shown from left to right.

IEEE SIGNAL PROCESSING MAGAZINE [63] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
mesh topology—is the determining factor and it is therefore ic modal frequency, the sound pressure value at a point (x1 , x2 ,
possible to extend the scattering junction concept beyond x3 , . . . , xN ) inside such space is determined by:
three spatial dimensions to hyper-dimensional DWMs [23],  
[61] that have been shown as appropriate for simulating artifi- 
N
ni xi
pn1 n2 n3 ...nN (x1 , x2 , x3 , . . . , xn) = A cos , (15)
cial reverberation [62]. i =1
Li
At low frequencies, the acoustic characteristics of a room
can be modeled with sufficient accuracy by approximating where A is an arbitrary amplitude coefficient.
the main dimensions of a basic cuboid model and simulating Examples of modal distributions calculated using (15) are
the corresponding resonant modes. In real rooms there are shown in Figure 10(a). Note that the additional advantage of
typically additional architectural features that lead to fre- a hyper-dimensional DWM reverb is the resultant high-densi-
quency-dependent irregularity in these predicted trajectory ty distribution of high-frequency modes while simultaneous-
lengths. Hence, at higher frequencies, a typical RIR will ly avoiding potentially problematic (in terms of perceived
demonstrate a large number of densely and irregularly dis- sound quality) low-frequency resonances. This is due to the
tributed modal peaks that are not determined by the basic trajectories being kept shorter compared to a similar model
geometry of the space alone. However, for high-quality artifi- with the same number of junctions but lower dimensionality
cial reverberation not all of these modes need to be simulated and is demonstrated in Figure 10(b) comparing 2-D and four-
directly, with approximately 1,500 modes distributed evenly dimensional DWMs. Hence, hyper-dimensional DWM reverb
or along a logarithmic scale between 80 Hz and 10 kHz being satisfies the requirement for a densely distributed high fre-
sufficient for diffuse and natural sounding artificial reverber- quency modal response while giving the freedom for the
ation [63]. In a DWM, the number of primary resonant modes more precise and sparsely arranged low-frequency modes to
is equal to the number of dimensions, with higher dimen- be modeled with any other appropriate technique without
sions leading to a more irregular arrangement of modal frequency overlap. Further work for a more natural reverber-
peaks. It is possible to extend these equations describing the ant effect requires the implementation of frequency-depend-
resultant wave propagation to the hyper-dimensional case, ent losses to simulate air and boundary absorption as used in
where for each mesh dimension xi the primary mode has a RIR synthesis with standard 2-D and 3-D DWM models.
frequency corresponding to (c/2Li ) with Li defined as the
trajectory length. In an N-dimensional space, standing waves CONCLUSIONS
occur at the following frequencies: The digital waveguide mesh has been an active area of
music acoustics research for over ten years. Although


 N  2  founded in 1-D digital waveguide modeling, the principles
c  ni
fn1 n2 n3 ...nN =  , (14) on which it is based are not new to researchers grounded in
2 i =1
Li
numerical simulation, FDTD methods, electromagnetic
simulation, etc. This article has attempted to provide a con-
where ni is the integer index of the current mode for each siderable review of how the DWM has been applied to
dimension and c is the speed of sound. Furthermore, at a specif- acoustic modeling and sound synthesis problems, including

2-D
50
Magnitude (dB)

1-D
1,000 2-D
3-D 0
4-D
5-D
Number of Modes

800
−50
600 0 0.05 0.1 0.15 0.2 0.25
4-D
50
Magnitude (dB)

400

0
200

0 −50
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Relative Frequency (1/fs) Relative Frequency (1/fs)
(a) (b)

[FIG10] Frequency response information for 2,310 node DWMs of varying dimension: (a) theoretical modal distribution varying with
increasing dimensionality; (b) frequency response of a 2-D and 4-D DWM.

IEEE SIGNAL PROCESSING MAGAZINE [64] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
new 2-D object synthesis and an overview of recent research Antti Kelloniemi (antti.kelloniemi@tkk.fi) received his M.Sc.
activities in articulatory vocal tract modeling, RIR synthe- from the University of Helsinki, Finland, in 2003. He is current-
sis, and reverberation simulation. The extensive, although ly finalizing his doctoral thesis, “Room acoustics modeling with
not by any means exhaustive, list of references indicates the digital waveguide mesh—Boundary structures and approxi-
that though the DWM may have parallels in other disci- mation methods,’’ at the Helsinki University of Technology. He
plines, it still offers something new in the field of acoustic has recently been appointed as the technology manager for
simulation and sound synthesis. Perhaps one reason for the loudspeaker manufacturer Panphonics Oy, where his research
continued interest in this area is the natural and intuitive concentrates mainly on electrostatic loudspeakers and active
complex emergent behavior that results from such simple, noise control. His interests include the modeling and design of
locally defined scattering equations. However, despite this acoustics and audio electronics.
perceived simplicity, it is also clear that there are still many Jack Mullen (jm220@ohm.york.ac.uk) received the
nontrivial problems to be solved. There are few current M.Eng. in electronic engineering with music technology
examples of useful and playable virtual instruments using from the University of York, York, U.K., in 2002. His final
DWM based sound synthesis (although a virtual drum in the year M.Eng. project involved research into digital waveguide
London Science Museum that can be played in real-time mesh boundary implementations for room acoustics model-
and is a realization of the work presented in [28] is a ing. He has recently been awarded a Ph.D. degree from the
notable exception), and this is mainly due to the computa- University of York in 2006 for his thesis on multidimensional
tional resources required for such a real-time model. waveguide vocal-tract modeling. His research interests
Nonreal-time operation is not a problem when simulating include speech and acoustical modeling.
static, linear time-invariant systems such as a representa- Simon Shelley (sbs107@ohm.york.ac.uk) received the
tion of an acoustic space, and hence most recent DWM work M.Eng. in electronic engineering with music technology sys-
has focused in this area. The real-time convolution of an tems from the University of York, U.K., in 2003. He is now work-
audio input signal with the impulse response generated ing toward a Ph.D. on the accurate modeling of acoustic
from such a model is trivial to implement on a modern boundaries in the digital waveguide mesh, which is to be com-
computing platform and is now a commonly used sound pleted in 2007. Research interests include acoustic modeling,
processing operation. Offline, only DWM-based virtual sound spatialization, and human–computer interaction for
instruments for sound synthesis prohibit user interaction musical applications.
and severely limit playability. However, some of the recent
developments presented in this article, particularly those
relating to dynamic, real-time vocal tract simulation, are
REFERENCES
beginning to make significant inroads in this area and will [1] S.A. Van Duyne and J.O. Smith, “Physical modeling with the 2-D Digital
hopefully lead to new DWM implementations that can be waveguide mesh,’’ in Proc. Int. Computer Music Conf., Tokyo, Japan, 1993,
pp. 40–47.
applied more generally—and more successfully—to the [2] V. Välimäki, J. Pakarinen, C Erkut, and M. Karjalainen, “Discrete time modeling
diverse range of possible sound synthesis applications. of music instruments,’’ Rep. Prog. Phys., vol. 69, pp. 1–78, Jan. 2006. [Online]
Available: http://www.iop.org/EJ/abstract/0034-4885/69/1/R01/
[3] S. Bilbao, Wave and Scattering Methods for Numerical Simulation. Chichester,
ACKNOWLEDGMENT UK: Wiley, 2004.
This work is supported in part by the U.K. Engineering and [4] J.L. Kelly and C.C. Lochbaum, “Speech synthesis,’’ in Proc. Fourth Int. Congr.
Acoustics, Copenhagen, Denmark, Sept. 1962, pp. 1–4.
Physical Sciences Research Council (EPSRC), grant number
[5] P.B. Johns and R.L. Buerle, “Numerical solution of 2-dimensional scattering
GR/S01481/01, and in part by the Swedish Foundation for problems using a transmission line matrix,’’ Proc. IEE, vol. 118, pp. 1203–1208,
International Cooperation in Research and Higher Education Sept. 1971.
[6] A. Fettweis, “Digital filters related to classical structures,’’ AEU: Archive für
(STINT), contract IG2002-2049. Antti Kelloniemi has Elektronik und” Ubertragungstechnik, vol. 25, pp. 78–89, Feb. 1971.
received funding from the Academy of Finland (project [7] J.O. Smith, “A new approach to digital reverberation using closed waveguide
201050) and the Nokia Foundation. networks,’’ in Proc. Int. Computer Music Conf., Vancouver, Canada, 1985, pp.
47–53.
[8] J.O. Smith, “Music applications of digital waveguides,’’ CCRMA, Stanford Univ.,
AUTHORS Stanford, CA, Tech. Rep. STAN-M-39, 1987.
Damian T. Murphy (dtm@ohm.york.ac.uk) received the B.Sc. [9] J.O. Smith, “Physical modeling using digital waveguides,’’Computer Music J.,
vol. 16, no. 4, pp. 74–87, 1992.
(Hons) in mathematics in 1993; the M.Sc. in music technolo-
[10] Stanford University News Service, “Music synthesis approaches sound quality
gy in 1995; and the D.Phil. in music technology in 2000, all of real instruments,’’ Stanford University News Release, July 1994. [Online]
Available: http://www.stanford.edu/dept/news/pr/94/940607Arc4222.html
from the University of York, York, U.K. He is currently a lec-
[11] J.O. Smith, “Principles of digital waveguide models of musical instruments,’’
turer in the Department of Electronics, University of York, in Applications of Digital Signal Processing to Audio and Acoustics, M. Kahrs and
and a visiting lecturer in the Department of Speech, Music K. Brandenburg, Eds. Boston, MA: Kluwer, pp. 417–466, 1998.
[12] M. Karjalainen, P. Huang, and J.O. Smith, “Waveguide networks for room
and Hearing, KTH, Stockholm, Sweden. His research inter- response modeling and synthesis,’’ in Proc. 118th AES Conv., Barcelona, Spain,
ests include physical modeling, spatial sound, and applica- preprint 6394, May 2005.
tions of the digital waveguide mesh. Dr Murphy is a member [13] M. Karjalainen and C. Erkut, “Digital waveguides versus finite difference
structures: Equivalence and mixed modeling,’’ EURASIP J. Applied Signal
of the Audio Engineering Society. Processing, vol. 7, pp. 978–989, June 2004.

IEEE SIGNAL PROCESSING MAGAZINE [65] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.
[14] C. Erkut and M. Karjalainen, “Finite difference method vs. digital waveguide [38] L. Savioja, T. Rinne, and T. Takala, “Simulation of room acoustics with a 3-D
method in string instrument modeling and synthesis,’’ in Proc. ISMA 2003, Mexico finite difference mesh,’’ in Proc. Int. Computer Music Conf., Aarhus, Denmark,
City, Dec. 2002. Sept. 1994, pp. 463–466.
[15] A. Kelloniemi, “Frequency dependent boundary conditions for the 3-D digital [39] P.R. Cook, “Identification of control parameters in an articulatory vocal tract
waveguide mesh,’’ in Proc. 9th Int. Conf. Digital Audio Effects (DAFX06), model with applications to the synthesis of singing,’’ Ph.D. dissertation, Stanford
Montréal, Canada, Sept. 2006, pp. 161-164. [Online]. Available: Univ., Stanford, CA, 1990.
http:www.dafx.ca/proceedings/papers/p_161.pdf
[40] V. Välimäki and M. Karjalainen, “Improving the Kelly-Lochbaum vocal tract
[16] D.T. Murphy and M.J. Beeson, “The KW-boundary hybrid digital waveguide model using conical tube sections and fractional delay filtering techniques,’’ in
mesh for room acoustics applications,’’ IEEE Trans. Speech Audio Processing, to Proc. Int. Conf. Spoken Language Processing (ICSLP), Yokohama, Japan, 1994, pp.
be published. 615–618.
[17] M.J. Beeson and D.T. Murphy, “RoomWeaver: A digital waveguide mesh based [41] S. Mathur and B.H. Story, “Vocal tract modeling: Implementation of continu-
room acoustics research tool,’’ in Proc. 7th Int. Conf. Digital Audio Effects ous length variations in a half-sample delay Kelly-Lochbaum model,’’ in Proc.
(DAFX04), Naples, Italy, Oct. 2004, pp. 268–273. [Online]. Available: IEEE Intl. Symp. Signal Processing and Information Technology, Darmstadt,
http://dafx04.na.infn.it/WebProc/Proc/P_268.pdf Germany, 2003, pp. 753-756.
[18] F. Fontana and D. Rocchesso, “Signal-theoretic characterization of waveguide [42] H.W. Strube, “Are conical segments useful for vocal-tract simulation?’’ J.
mesh geometries for models of two-dimensional wave propagation in elastic Acoust. Soc. Amer., vol. 114, no. 6, pp. 3028–3031, 2003.
media,’’ IEEE Trans. Speech Audio Processing, vol. 9, no. 2, pp. 152–161, Feb. 2001.
[43] J. Mullen, D.M. Howard, and D.T. Murphy, “Waveguide physical modeling of vocal
[19] F. Fontana and D. Rocchesso, “A new formulation of the 2-D waveguide mesh tract acoustics: Flexible formant bandwidth control from increased model dimension-
for percussion instruments,’’ in Proc. XI Colloq. Musical Informatics, Bologna, ality,’’ IEEE Trans. Speech Audio Processing, vol. 14, no. 3, pp. 964–971, May 2006.
Italy, Nov. pp. 27–30, 1995.
[44] J. Mullen, D.M. Howard, and D.T. Murphy, “Real-time dynamic articulations in
[20] J. Laird, P. Masri, and N. Canagarajah, “Modeling diffusion at the boundary of the 2D waveguide mesh vocal tract model,’’ IEEE Trans. Speech Audio Processing,
a digital waveguide mesh,’’ in Proc. Int. Computer Music Conf., Beijing, China,, submitted for publication, 2006.
Oct. 1999, pp. 492–495.
[45] B.H. Story, I.R. Titze, and E.A. Hoffman, “Vocal tract area functions from mag-
[21] S.A. Van Duyne and J.O. Smith, “A simplified approach to modeling dispersion netic resonance imaging,’’ J. Acoust. Soc. Amer., vol. 100, no. 1, pp. 537–554, 1996.
caused by stiffness in strings and plates,’’ in Proc. Int. Computer Music Conf.,
Aarhus, Denmark, 1994, pp. 407–410. [46] D.G. Childers, Speech Processing and Synthesis Toolboxes. New York: Wiley, 2000.

[22] D.T. Murphy and D.M. Howard, “2-D digital waveguide mesh topologies in [47] J. Mullen, D.T. Murphy, and D.M. Howard, “Physical modeling of the vocal
room acoustics modeling,’’ in Proc. 3rd Int. Conf. Digital Audio Effects (DAFX00), tract with the 2-D digital waveguide mesh,’’ [Online]. Available: http://www-
Verona, Italy, Dec. 2000, pp. 211–216 [Online]. Available: http://profs.sci.univr.it/∼ users.york.ac.uk/∼ dtm3/vocaltract
dafx/DAFx-final-papers.html [48] S. Shelley, “Acoustic modeling—The digital waveguide mesh,’’ [Online].
[23] L. Savioja, T.J. Rinne, and T. Takala, “Simulation of room acoustics with a 3-D Available: http://www-users.york.ac.uk/∼dtm3/vocaltract.html
finite difference mesh,’’ in Proc. Int. Computer Music Conf., Aarhus, Denmark, [49] S.A. Van Duyne, J.R. Pierce, and J.O. Smith, “Traveling wave implementation
1994, pp. 463–466. of a lossless mode-coupling filter and the wave digital hammer,’’ in Proc. Int.
[24] S.A. Van Duyne and J.O. Smith III, “The 3-D tetrahedral digital waveguide Computer Music Conf., Aarhus, Denmark, 1994, pp. 411–418.
mesh with musical applications,’’ in Proc. Int. Computer Music Conf., Hong Kong, [50] D. Rocchesso and F. Fontana, Eds., The Sounding Object. Florence, Italy:
1996, pp. 9–16. Mondo Estremo, 2003.
[25] S.J. Miklavcic and J. Ericsson, “Practical implementation of the 3D tetrahedral [51] L. Savioja, J. Backman, A. Järvinen, and T. Takala, “Waveguide mesh method
TLM method and visualization of room acoustics,’’ in Proc. 7th Int. Conf. Digital for low-frequency simulation of room acoustics,’’in Proc. Int. Congr. Acoustics,
Audio Effects (DAFX-04), Naples, Italy, Oct. 2004, pp. 262–267. [Online]. Available: Trondheim, Norway, vol. 2, pp. 637–641, June 1995.
http://dafx04.na.infn.it/WebProc/Proc/P_262.pdf
[52] D.T. Murphy and M.J. Beeson, “Modeling spatial sound occlusion & diffraction
[26] F. Fontana, D. Rocchesso, and E. Apollonio, “Using the waveguide mesh in effects with the digital waveguide mesh,’’ in Proc. Audio Eng. Soc. 24th Int. Conf.,
modeling 3-D resonators,’’ in Proc. 3rd Int. Conf. Digital Audio Effects (DAFX-00), Banff, Canada, June 2003, pp. 207–216.
Verona, Italy, Dec. 2000, pp. 229–232 [Online]. Available: http://profs.sci.univr.it/
∼dafx/DAFx-final-papers.html [53] U. Stephenson, “Comparison of the image source method & the particle simu-
lation method,’’ Appl. Acoust., vol. 29, no. 1, pp. 35–72, 1990.
[27] G.R. Campos and D.M. Howard, “On the computational efficiency of different
waveguide mesh topologies for room acoustic simulation,’’ IEEE Trans. Speech [54] A. Kelloniemi, V. Välimäki, and L. Savioja, “Simulation of room acoustics
Audio Processing, vol. 13, no. 5, pp. 1063–1072, Sept. 2005. using 2-D digital waveguide meshes,’’ in Proc. IEEE Int. Conf. Acoustics, Speech,
and Signal Processing (ICASSP), Toulouse, France, May 2006,
[28] M.L. Aird, J. Laird, and J. Ffitch, “Modeling a drum by interfacing 2-D and 3-D pp. 163-168. [Online]. Available: http://www2.hsu-hh.de/ant/dafx2002/papers
waveguide meshes,’’ in Proc. Int. Computer Music Conf., Berlin, Germany, Aug. /DAFX02_Murphy_Mullen_waveguide_mesh_modeling.pdf
2000, pp. 82–85.
[55] A. Kelloniemi, D.T. Murphy, L. Savioja, and V. Välimäki, “Boundary conditions
[29] F. Fontana, “Physics-based models for the acoustic representation of space in in a multi-dimensional digital waveguide mesh,’’ in Proc. IEEE Int. Conf.
virtual environments,’’ Ph.D. dissertation, Dept. Computer Science, Univ. of Acoustics, Speech, and Signal Processing (ICASSP), Montreal, Canada, May 2004,
Verona, Italy, Apr. 2003. [Online]. Available: http://profs.sci.univr.it/∼ pp. IV-25–IV-28.
fontana/paper/20.pdf
[56] A. Kelloniemi, V. Välimäki, and L. Savioja, “Spatial filter-based absorbing
[30] D. Rossiter, A. Horner, and G. Baciu, “Visualization and manipulation of 3D boundary for the 2-D digital waveguide mesh,’’ IEEE Signal Processing Lett., vol.
digital waveguide structures,’’ in Proc. Int. Computer Music Conf., Beijing, China, 12, no. 2, pp. 126–129, Feb. 2005.
1999, pp. 43–46.
[57] A. Kelloniemi, “Improved adjustable boundary condition for the 3-D digital
[31] P. Huang, S. Serafin, and J.O. Smith, “A 3-D waveguide mesh model of high- waveguide mesh,’’ in Proc. 2005 IEEE Workshop Applications of Signal Processing
frequency violin body resonances,’’ in Proc. Int. Computer Music Conf., Berlin, to Audio and Acoustics, New Paltz, NY, Oct. 2005, pp. 191-194.
Germany, Aug. 2000, pp. 86–89.
[58] J. Huopaniemi, L. Savioja, and M. Karjalainen, “Modeling of reflections and air
[32] S. Serafin, P. Huang, and J.O. Smith, “The banded digital waveguide mesh,’’ in absorption in acoustical spaces: a digital filter design approach,’’ in Proc. 1997
Proc. MOSART Workshop Current Research Directions Computer Music, IEEE Workshop Applications of Signal Processing to Audio and Acoustics), New
Barcelona, Spain, Nov. 2001. [Online]. Available: http://www.imi.aau.dk/∼sts/ Paltz, NY, Oct. 1997.
publications/mosart.pdf
[59] K. Lee and J.O. Smith, “Implementation of a highly diffusing 2-D digital
[33] A. Kelloniemi, V. Välimäki, and L. Savioja, “Simulation of room acoustics waveguide mesh with a quadratic residue diffuser,’’ in Proc. Int. Computer
using 2-D digital waveguide meshes,’’ in Proc. IEEE Int. Conf. Acoustics, Speech, Music Conf., Miami, Florida, Nov. 2004. [Online}. Available: http://ccrma.stan-
and Signal Processing (ICASSP), Toulouse, France, May 2006, pp. 313-316. ford.edu/~kglee/pubs/klee-icm04-pdf
[34] L. Savioja and V. Välimäki, “Reducing the dispersion error in the digital wave- [60] S. Shelley and D.T. Murphy, “Measuring diffusion in a 2-D digital waveguide
guide mesh using interpolation and frequency warping techniques,’’ IEEE Trans. mesh,’’ in Proc. 8th Int. Conf. Digital Audio Effects (DAFX-05), Madrid, Spain,
Speech Audio Processing, vol. 8, no. 2, pp. 184–194, Mar. 2000. Sept. 2005. [Online]. Available: http://dafx05.ssr.upm.es/Proc_DAFx05/P_249.pdf
[35] L. Savioja and V. Välimäki, “Interpolated rectangular 3-D digital waveguide [61] D. Rocchesso and J.O. Smith, “Circulant and elliptic feedback delay networks
mesh algorithms with frequency warping,’’ IEEE Trans. Speech Audio Processing, for artificial reverberation,’’ IEEE Trans. Speech Audio Processing, vol. 5, no. 1, pp.
vol. 11, no. 6, pp. 783–789, Nov. 2003. 51–63, Jan. 1997.
[36] F. Fontana and D. Rocchesso, “Online correction of dispersion in 2D wave- [62] A. Kelloniemi, V. Välimäki, P. Huang, and L. Savioja, “Artificial reverberation
guide meshes,’’ in Proc. Int. Computer Music Conf., Berlin, Germany, Aug. 2000, using a hyper-dimensional FDTD mesh,’’ in Proc. 13th European Signal
pp. 78–81. Processing Conf. (EUSIPCO 2005), Antalya, Turkey, Sept. 2005.
[37] F. Fontana, “Computation of linear filter networks containing delay-free loops, [63] J. Merimaa, T. Lokki, T. Peltonen, and M. Karjalainen, “Measurement, analysis,
with an application to the waveguide mesh,’’ IEEE Trans. Speech Audio and visualization of directional room responses,’’ presented at AES 111th Conv.,
Processing, vol. 11, no. 6, pp. 7743–782, Nov. 2003. New York, NY, Nov. 2001, preprint 5449. [SP]

IEEE SIGNAL PROCESSING MAGAZINE [66] MARCH 2007

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on July 26,2024 at 13:12:25 UTC from IEEE Xplore. Restrictions apply.