Signal Processing - May 2017 - 51

with a gradually diminishing accuracy. Since first-order early
reflections are perceptually more important than the higher-order
reflections, the resulting reverberation is perceptually realistic and
statistically very similar to that of an actual room. A by-product of
this design is that the SDN does not require separate modules for
early reflections and late reverberation, while still allowing precise
and explicit control of the room geometry, source and receiver directivity patterns, and wall absorption characteristics. Furthermore, it
enables a straightforward implementation of virtual multichannel
recordings and binaural auralization. A block diagram of the SDN
reverberator is depicted in Figure 14, showing that it has a structure
very similar to the FDN, while allowing direct and explicit control
over all the physical characteristics of the space it simulates.

S
Sound
Source
S
S

Microphone

Audio source culling

Complex virtual environments typically include many sound
sources, which makes synthesizing their acoustics a challenging
task in terms of the associated computational cost. This difficulty is especially pronounced when rendering such audio content
over devices with limited computational power, such as mobile
phones. In a typical scenario involving many concurrent sources, it may be necessary to select and render only a few.
State-of-the-art game engines typically use volumetric
culling of sound sources. Each sound source has an associated
culling volume (cube, sphere, or cylinder), and when the listener is within this volume, the sound is rendered. This is a
simple approach that does not incur any significant computational cost, apart from the relatively simple comparison operations between the bounding boxes of the listener and each of
the sound sources. However, this approach does not take into
account the relative levels of the sound sources. It also does not
limit the number of sound sources that can be simultaneously
active. This proximity makes the only determinant of whether
or not a sound source will be rendered, completely disregarding its perceptual salience. Sound sources are dynamically
activated in response to user-generated events in interactive
applications such as games and VR. For scenes comprising
multiple concurrent sound sources, many of these sources will
be masked by the others. This makes it redundant to process
these inaudible sources.
Tsingos et al. [95], [96] provide a perceptually based sourceculling approach. The approach is based on ranking the sources

Scattering Node
Bidirectional Delay Line
Monodirectional Delay Line

FIGURE 13. A conceptual depiction of the SDN reverberator. The figure
shows a rectangular room as observed from above, with the associated delay lines interconnected at scattering junctions on the wall. Other
delay lines associated to the floor and ceiling are also present but are not
shown here for clarity. (Figure used courtesy of [12].)

in the scene using their binaural loudness at different frequency
bands as a measure of perceptual salience. Loudness values are
used to calculate a masking threshold from a time-frequency
representation of the sound sources and stored for use during
runtime. As a new sound event occurs, the decision to render
the new sound source is made at the audio-frame level. Each
frame is compared with the existing mix for evaluating whether
the mix can mask it. If it can, the frame is culled. As a result of
the frame-level temporal resolution, several frames from a single sound source can be culled while others are rendered. This
results in each sound source being only partially culled. A similar algorithm was proposed by Metan and Hacıhabibog˘lu [97].
The audibility calculation in this algorithm is slightly different
from that of Tsingos et al. As a new sound event is generated, a
look-ahead algorithm checks for the audibility of each frame of
a sound source, given the current mix being played. The decision to render a sound source is based on the ratio of audible

Delays Between
Nodes

Input

Source to Nodes:
Delays, Attenuation,
and Source Directivity

Nodes to Microphone:
Delays, Attenuation,
and Microphone Directivity

Scattering and
Wall Absorption

Output

Direct Path: Delay, Attenuation,
Source and Microphone Directivity

FIGURE 14. A conceptual block diagram of the SDN reverberator.
IEEE Signal Processing Magazine

May 2017

Table of Contents for the Digital Edition of Signal Processing - May 2017