Analysis Methods
Wavelets
The first idea was to use wavelets to classify parrot songs and calls. Several wavelet shapes in the Matlab wavelet toolbox were used with parrot call data collected by Liz Bradley in Irene Pepperburg's lab, and their effects examined. With several of the wavelets, the decompositions resembled the original signal, while in a few cases, the decomposition resembled the wavelet used for analysis.
Acoustic Fingerprint
The next idea was inspired by song recognition mobile applications -- if the algorithms used in these applications can recognize songs, why not birds? Shazam uses a technique called acoustic fingerprinting, which provides a digital summary of a signal based on its spectrogram. This was promising because a lot of research on bird calls is based on spectrograms of audio signals. To see if acoustic fingerprinting could be done in Matlab, a script that tested the technique on different covers of the same song was written. The test was fairly accurate, but the length of the signal needed for accurate identification was too long for useful analysis on bird calls.
Machine Learning
The next technique considered was machine learning. A conversation with Abeer Alwan from UCLA planted the seed for the idea for this approach, and a paper by Dan Stowell from Queen Mary University of London provided a detailed comparison of different feature extraction methods. A Udacity course on machine learning showed how to utilize the classifiers in the scikit learn python library. This is the current focus of research.
What's next?
The current focus of research is identifying species based on data collected from xeno-canto using a machine learning algorithm. Decision trees are the method currently being tested, but others will be used for comparison. The current long-term goal is to identify different individuals of the same species.