30
68
Spécial « Congrès Acoustics 2012 »
Un nouveau procédé d’optimisation de la distance géométrique dans un système de reconnaissance automatique de chants d’oiseaux
Early Testing
Early testing (see Boucher et al [6]) was based on a CD
of parrot recordings consisting of 20 different sets of
parrot calls with about 600 instances of calls in total. It
took some years but the time came when with the initial
CD the system was 100% accurate with 0% false positi-
ves (a few that were initially found turned out to be due
to the wrong parrot making a guest appearance in some
of the tracks). This was a good result but the real world
is not CD quality and a lot more work was needed to get
similarly good results in the field.
Real World Challenges
The real world was far more challenging. The first thing
to be learnt is that birds (and we mainly worked with birds
at first) have a woeful sense of pitch. Young birds clearly
practice by both overshooting and undershooting the target
frequencies by gliding up and down through the frequency
range. We had to allow for this.
To make matters worse it seems birds do not have a
sense of scales (in the musical sense) and so in different
geographical regions they are likely to settle on a diffe-
rent “normal” frequency for a particular call.
In the search for rare species, even a poor recording is
valuable and so it is necessary to work in a high S/N envi-
ronment. Initially we aimed at 10 dB S/N but found that
number to be too conservative. Sometimes today we are
working at -20 dB S/N and getting helpful results.
To be really useful the system needs to be able to work
with multiple species at the same time. This was some-
thing that again was easy enough using CD recordings
but proved much more difficult when the calls were mixed
with significant noise.
Next we found that most species had dozens of call types
and some even more. Therefore we would need multi-
ple examples of each type of call to get good matches.
Typically we recommend that the reference files contain
at least 10 examples of each call type.
The original system was designed to handle up to 32,768
different reference calls, but this was later extended to over
1 billion. Which brings us to the next problem - where are
all these reference calls to come from? It is not so easy
to get good clean reference calls. Here it is worth noting
that although the system can work in poor S/N levels,
it is rather important to have good references as othe-
rwise the system will try to match on the call + the noise
and in very poor S/N environments can end up matching
noise with noise.
Reference File Sources
The majority of our users have had at least one attempt to
collect their own reference calls. The results are usually
disastrous. In order to find out why, we met with some
professional sound recordists to get their views on this.
It seems that patience is the main ingredient. Professional
recordists will spend days or sometimes even weeks to get
that special recording. Additionally they use top quality
equipment and have years of accumulated experience.
For the amateur the best advice is to get as close to the
target as possible and record at the lowest volume level
(to keep out extraneous sounds).
However the real answer is to get the professionals onside
and have them provide the reference files. We have found
that in Australia, at least, this is readily achieved.
The Recorder
Initially we assumed that the recorder would be some-
thing that was available off the shelf. And while it is true
that many good quality recorders are commercially availa-
ble few are designed to operate for long periods without
attention and in any kind of weather.
Two problems emerged as most salient: the need for
weatherproofing the housing and the need to power the
unit for long periods (months at a time). It was soon reali-
zed that these two factors were interdependent as lower
power consumption meant less heat, smaller batteries
and so a smaller housing.
Recorder Characteristics
Most recorders are designed to work only at close range.
That is, the user either actually holds the recorder or places
it nearby. For wildlife recording, in order to get sufficient
information, the recorder needs to operate over a wide
area. To allow for this we designed a high-gain AGC ampli-
fier, that had an effective range from 0.5 metres to seve-
ral kilometers (under the right conditions).
The recorder has a large 27 AH battery that can power the
Olympus LS-11 for about 2 months while the timer can be
activated with up to 8 settings a day. Power can also be
provided from and external socket that can double as a
solar panel charging point. The battery is designed to be
removed in seconds for easy field replacement, as is the
LS-11. The LS-11 is controlled entirely from its external
ports and there are no modifications to it.
On board is a high-gain AGC amplifier, a three frequency
filter (60, 120 and 200 Hz) and a switchable gain control.
The microphones clip into a pair of XLR sockets. The
microphones are standard low noise electrets, but ultra-
low noise ones are available if the yare needed.
Fig. 3: The recorder based on the Olympus LS-11, with large
external battery, timer and AGC card all housed in a
waterproof case.
L’enregistreur est composé d’un Olumpus LS-11
avec une grosse batterie externe, un chronomètre
et une carte AGC placés dans une valise étanche
