I received a number of responses to my post last week about comparing
bird songs. Here is a summary of the responses, followed by my original
post. I have not had a chance to follow up on these suggestions, but
many of them look promising. Thanks for the terrific advice!
Adrian del Nevo reported that he is using STATISTICA's neural network
module to perform these sorts of analyses: "We can 'train' the neural
network to recognize similar patterns of songs and thus conduct the
types of analyses you describe."
A few people (Arch McCallum, Tobias Riede, and Aisha Thorn) suggested
looking into dynamic time warping. Papers included one by Dave Bradley
and Rich Bradley in the 1970's on the Belding's Savannah Sparrow, and a
paper by John Wiens comparing Sage Sparrow songs. There is also a 1998
paper in JASA 103(4):2185-96 that compares dynamic time warping and
hidden Markov models. Another paper on dynamic time warping was:
Tchernichovski, O., Nottebohm, F., Ho, C.E., Bijan, P., Mitra, P.P.
(2000). A procedure for an automated measurement of song similarity.
Animal Behaviour. 59, 1167-1176.
Ann Chen has used a genomics technique to analyze vocalization
similarity in false killer whales. She used the Smith-Waterman
algorithm, which is the backbone of BLAST, a tool used by genomics and
proteomics researchers to compare sequences. The approach can account
for similarities between phrase types and will output a distance matrix
that can be used in a cluster analysis. She is currently working on a
techniques paper that will publish her methods.
Diego Gil said that he has used Levenshtein distance to analyze willow
warbler songs, and his analysis is reported in: Gil, Diego and Peter
Slater. 2000. Song organisation and singing patterns of the Willow
Warbler, Phylloscopus trochilus. Behaviour 137:759-782.
A couple of people (Jim Ingold and Ted Miller) suggested contacting
Donald Kroodsma at U Mass or checking out his recent book: Kroodsma,
D.E. (2005) The Singing Life of Birds. Houghton Mifflin.
Shaari Unger suggested checking out the work of Dr. John Ford at Canada
Fisheries and Oceans on Orcinus orca vocalizations.
Greg Clark wrote that he had performed a similar analysis a few years
ago, but that the technique he used was not particularly user friendly.
Jim Nollman has been using a program called Melodyne (owned by Celemony)
to analyze recordings. The software's primary purpose is to analyze
and repair music, but he thought that the software could conduct some
interesting comparative analyses.2
I'm hoping someone on this list can point me to a technique for
analyzing similarity between bird songs (within a species). Basically,
we have multiple songs recorded from about 20 bobolinks, and we want to
determine if songs recorded from individuals within a local area are
more similar than songs from different areas. Bobolink songs are
complex, but are a sequence of phrases that are put together in
different ways by different individuals.
My question is whether there is a way to analyze a series of characters
for frequency and placement. For instance we may have the following
three songs, where letters represent phrases:
1 AGHEIGHDDS
2 AGEGDSSHIIJJJ
3 CTQEUYSEKE
Songs 1 and 2 share more phrases and start with the same phrases, so I
would classify them as being more similar. Does anyone know of an
analytical technique for dealing with this sort of data? I'm thinking
that the data is probably similar to some types of genetic data, where
letters would refer to alleles, and I'm also wondering if anyone has
applied genetic analysis methods to this type of data.
--
----------------------------------------------------------
Brian R. Mitchell
Post-Doctoral Associate
University of Vermont
The Rubenstein School of Environment and Natural Resources
81 Carrigan Drive
Burlington, VT 05405-0088
(802) 656-2496
----------------------------------------------------------
|