28
68
Spécial « Congrès Acoustics 2012 »
eginning in 2002 we set out to design a recording and PC
analysis package that could allow 24/7 acoustic surveillance
of areas that were suspected to be prime sites for the irre-
gular appearance of a rare Australian parrot, the Coxen’s Fig
Parrot (Cyclopsitta diophthalma coxeni). From the outset we
decided to build the system to be a general purpose sound
detection and recognition system, with the intention that it
could be used for other acoustics purposes.
An initial survey of the technology available at the time
indicated that, although many had attempted to build such
a system before, there was little evidence that anything
existed that could meet our requirements, either for the
recording devices or the PC analysis. In particular, our
recording device needed to operate for months at a time
in a sub-tropical rainforest and the PC recognition system
needed to be capable of distinguishing our target parrot
species from 12 other co-habiting parrot species, some
of which were known to be challenging to separate by ear
even to an experienced human listener.
Past failures by others convinced us quite early that we
needed to be innovative and explore new technologies up
to and including new mathematics if necessary.
The Recognition Concept
The basic concept of the system was to transform a sound
into a spectral image and then use pattern matching tech-
niques. The matching was to be mathematical and not
using AI or similar techniques.
Ultimately it was decided to use a library of reference calls,
stored as mathematical images of their spectrums and
then to find the best matching reference to each sound.
Initially the Euclidian distance was used but, while it proved
adequate for perfect or near perfect matches, it was not
good at handling similarity. At this point we adopted a
method pioneered by one of us (Jinnai) which he called
the Geometric Distance (GD).
A new optimization method of the geometric
distance in an automatic recognition system for
bied vocalisations
Un nouveau procédé d’optimisation de la distance géométri-
que dans un système de reconnaissance automatique
de chants d’oiseaux
Michihiro Jinnai
Kagawa National College of Technology
355 Chokushi-cho
761-8058 Takamatsu
Japan
E-mail :
Neil Boucher
SoundID
PO Box 649
Maleny, 4552
Queensland
Australia
E-mail :
Minoru Fukumi
University of Tokushima
2-1 Minami-josanjima
770-8506 Tokushima
Japan
E-mail :
Hollis Taylor
Wissenschaftskolleg zu Berlin
Institute for Advanced Study
Wallotstrasse 19
14193 Berlin
Germany
E-mail :
Abtract
We describe a fully automated, PC based wildlife monitoring and survey system that
is used for diverse species studies. The system as described is fully functional and
operational at this time (download it at
). It uses a wide-area recorder
that can record over areas of up to several square kilometres. The recorder can run,
unattended, for more than a month. The recordings can either be analysed in real time
to produce a particular response (e.g. send an SMS if a rare parrot is detected), or can
be analysed post-recording on a PC. Any number of different species can be analysed
simultaneously. In survey mode, calls can be counted and recognised with a summary
of species and calling rate produced. The system has been successfully tested with the
dawn chorus (which itself can be used for censuses) and against human surveys with
impressive results. The software is equally effective for animal and non-animal sound
sources. It can analyse calls at a rate of more than 100,000 per second.
(This is equivalent to comparing 100,000 spectrograms per second.)
Résumé
Nous décrivons ici un système de surveillance de la faune et d’enquête, entièrement
automatisé sur PC, qui est utilisé pour étudier diverses espèces d’oiseaux. Il utilise un
capteur qui permet d’enregistrer sur des zones de plusieurs kilomètres carrés. L’appareil
peut fonctionner, sans surveillance, pendant plus d’un mois. Les enregistrements peuvent
être analysées en temps réel pour produire une réponse particulière (par exemple envoyer
un SMS si un perroquet rare est détecté), ou peuvent être analysés plus tard sur un PC.
N’importe quel nombre d’espèces différentes peuvent être analysés simultanément. En
mode enquête, les appels peuvent être pris en compte et reconnus avec un résumé des
espèces et le taux d’appels produits. Le système a été testé avec succès lors du chant de
l’aube (qui lui-même peut être utilisé pour les recensements) et malgré une surveillance
humaine avec des résultats impressionnants. Le logiciel est également efficace pour les
sources d’origine non-animale. Il peut analyser plus de 100 000 appels par seconde.
(Ceci est équivalent à comparer 100 000 spectrogrammes par seconde.). Ce système
est entièrement fonctionnel et opérationnel (le télécharger à
).
B
