32
68
Spécial « Congrès Acoustics 2012 »
Un nouveau procédé d’optimisation de la distance géométrique dans un système de reconnaissance automatique de chants d’oiseaux
Accuracy also increases as the number of reference
calls in the template is increased. For example if there
are 100 calls of a certain species, for each signal on the
recording every one of the 100 recordings is compared
in turn. Each one will be assigned a GD and the smallest
GD is the one that is declared the match.
This goes even further when there are multiple species.
Assuming each species has its own template then as
each of the templates are run the lowest GD becomes
the “best match”. Some species sound similar and as an
example the Australian Currawong sometimes imitates
the Grey Shrike Thrush (both species considered above).
So on a first pass the software may assign a match of
a call to the Currawong (when it is running the corres-
ponding template) . However, later it might run the Grey
Shrike Thrush (which we assume here is the calling bird).
Now to all but the most expert human ear these are the
same. But the software will not be fooled and will assign a
lower GD to the Grey Shrike Thrush which at the exact time
of the call will now over-ride its first “guess” of a Currawong
with the Grey Shrike Thrush. When the run is completed
the correct assignment will have been made.
So the software accuracy improves not only with more
examples of the target species, but also with more exam-
ples of other species that might be calling in the area.
Trade-offs
As the software developed it became clear that it was
both possible and desirable to trade CPU time for grea-
ter accuracy. Increasing the number of templates (each
with their own settings) certainly improved the accuracy
but increased CPU usage almost in direct proportion to
the number of templates.
The 2-D analysis for most calls closely approximates the 3-D
and since it runs faster it is the preferred mode. The 3-D
mode is suited best to those situations where the tempo-
ral signature is important (for example in estimating the
number of frogs calling in a chorus).
Noise performance of the system is good and it is possi-
ble to trade off accuracy for the ability to work in a noisy
environment. The system will perform well at S/N levels
of 10 dB, but can still give useful results in conditions
as noisy as -20 dB S/N if required. In this instance it
is found that by matching just the peak energy part of
the call (and hence by setting the parameters to focus
on the peak energy section of the call) it is possible
to get good matching. However, by doing this, a lot of
information about the rest of the call is not used and
the uniqueness of the call is totally searched for in a
small portion of the energy peak, so that false positi-
ves will increase.
After a lot of testing we found that the PC clock speed
was the most important indicator of the total run-time. In
recently years clock speeds seem to have saturated and
3.8 GHz seems to be about as fast as a PC will run without
over-clocking. Modern PCs tend to be adding processors
rather than ramping up clock speed and this is a minor
dilemma for the software. While most 64 bit code has
powerful parallel implementations the older 32 bit code
usually does not. We therefore decided to limit the 32
bit code to one processor (although on a multiprocessor
machine multiple instances of the code can be run at any
time and this is recommended for processing very large
file collections).
Field Applications
It is well established that employing bioacoustic methods
for animal surveys in ecological studies offer a number
of advantages over using visual-type surveying methods.
These advantages include: a reduced need to disturb or
handle animals, an ability to survey visually cryptic species,
achieving effective monitoring during inclement weather,
and cost savings through reducing the amount of times a
site needs to be visited by highly skilled specialists.
Despite its clear utility and widespread use (see for exam-
ple work by Rountree et al[1],Payne et al[2], Riede[3]), the
application of bioacoustic monitoring can by somewhat
constrained by “back end” data analysis requirements.
That is, studies using bioacoustics can generate large
amounts of acoustic data, often thousands of hours of
recordings [1]. Processing large amounts of data that is in
an acoustic form can be laborious and time consuming. As
such, the time, resource and cost savings gained through
implementing bioacoustics monitoring over other forms
of survey method may be completely eroded.
Automated analysis of bioacoustic data is promoted as
the answer to circumventing manual data processing and
analysis obstacles. In practice however, the development
of automated sound recognition is challenging, primarily
because the vocalisations of many species can be complex.
For example, some bird species can sing in duets, while
others can intentionally mask their calls, perform vocal
mimicry, have regional dialects, have large song repertoires
and can perform improvisational songs [4]. Furthermore,
despite widespread reporting of successful automated
sound recognition in the literature (including for birds
species [5]), the utility, practicality and accessibility of
these systems to field ecologists seemed limited.
Field Implementation
As discussed, the recorder and automated sound recogni-
tion system was initially designed as an acoustic surveillance
tool for rare parrots, and in this regard is an advancement
in acoustic survey techniques for the conservation of rare
and threatened fauna species. To illustrate, the system can
be deployed at key sites on a long term basis (e.g. over the
fruiting season of certain food source trees within the known
range of the target species) to make high quality recor-
dings over a distance of hundreds of metres. Recordings
can then either be accurately analysed by the software in
real-time to produce a particular response, for example
send an SMS notification over a mobile phone network if
a rare parrot is detected, or analysed post-recording on a
PC to extract information on timing and frequency of site
visitation and potentially species abundance.
Clearly, however, the system has additional utility across the
fields of natural resource management. For example, in most
Australian jurisdictions a Development Application requires
the completion of an Environmental Impact Assessment
(EIA). In these cases, and particularly for large scale deve-
lopments such as mines, a comprehensive survey is requi-
red of local fauna to determine the impact the development
will have on wildlife populations. Because the automated
sound recognition system can be used to recognise the
sound/vocalisation of any species or group of animals it
can be used in conjunction with conventional techniques
to enable a more accurate census of wildlife to be underta-
ken in the EIA process with minimal additional resourcing.
