Aerospace and Electronic Systems Magazine February 2018 - 53

SOUND SOURCE LOCALIZATION
The localization of sound sources using microphone arrays is a
well-known signal processing task with various application areas,
different requirements, and technical solutions [3]. For example,
speaker localization to improve speech acquisition [4] requires
robust real-time algorithms, ideally coping with limited technical resources such as a small number of microphones; precision
has to be sufficient for subsequent beamforming, which is often
restricted to a direction in a plane, assuming the far-field case. In
other tracking applications a camera is directed at a speaker in
a video conference [5] or a robot is navigated towards a sound
source [6].
The séance scenario poses somewhat different requirements.
Among the acoustic phenomena to be investigated are the so-called
direct voices: While the participants and the medium are assumed
to sit on their chairs, voices can be heard in the middle of the room.
They engage in an interactive dialogue with the participants and
even appear to be moving around. A widespread suspicion is that
the voices are produced by the medium (or a helper), who somehow deceives the listeners into believing the voices are coming
from somewhere else. Sound source localization promises some
objective proof, taking into account the following requirements:
1. Three-dimensional (3D) near-field localization is needed, with
high precision, ideally with centimeter accuracy throughout the
room.

balls with a maximum time resolution of 10 events per second.
Thin vertical lines have been added to indicate the x-y-position on
the floor. The 3D-model of the room has been created based on
simple yardstick measurements, including approximate position
of chairs. The characters shall give an idea about where people's
heads and bodies are allegedly located, individual size of participants, slight displacement of chairs etc. has not been taken into
account.

GCC-PHAT, SRP-PHAT, AND ADAPTATIONS
The sound source localization is based on the well-known generalized cross correlation (GCC) algorithm [7]. When applying the
phase transform (PHAT) to normalize the amplitude, the experiments showed that localization robustness significantly improved
by reducing the whitening with PHAT-β [8] (with β = 0.7). For the
microphone pairs (k, l), with k, l  {1, ..., 4}, k ≠ l, and the shortterm Fourier transform Xi(f) of recorded signal i, six cross correlation functions are obtained:
Rk ,l (τ ) = 

∞

−∞

X k ( f ) X l* ( f )
X k ( f ) X l* ( f )

e j 2π f τ df

(1)

With microphone distances of one to two meters and a high
sampling rate (up to 96 kHz), the functions span up to 500 sample

2. False measurements should be reliably detected and removed;
rather collect sparse data than unreliable data to not jeopardize
the trust in the entire investigation.
3. Computational efficiency is not an issue, calculations can be
done offline, there are no real-time requirements.
4. Audio can be recorded with high quality.
5. The scenarios vary a lot: Sometimes everybody is listening
quietly to the spirit voices, sometimes there are several people
speaking, laughing, and there might even be music playing in
the background.
6. The small room with mostly blank walls has quite strong early
reflections.
Figure 2 gives an example of measurement results for a period
of few minutes of the séance. Acoustic events are shown as dark
FEBRUARY 2018

Figure 2.

Overview: Localization of acoustic events.

IEEE A&E SYSTEMS MAGAZINE

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine February 2018