Classification of bird syllables in noisy environments using multitapers

A method for syllable classification of the Great Reed Warbler (Acrocephalus arundinaceus) has been studied and tested. This method uses multitapers in order to calculate the ambiguity domain and extracting features. Inspired by this technique four new methods, that build on extracting features from the ambiguity function using different types of multitaper kernels and noise reduction techniques, have been developed. These methods use different kinds of kernels and multitapers in order to change the focus of the method. A smaller study is made of how the multitaper windows behave with different kernels and how to solve problems that can occur. The methods are validated in different noise levels on a simulated data set, where especially difficult cases are simulated, and on real datasets with extra noise added. From this study, it is found that kernels which retain the cross-terms and suppress the auto-terms are harder to adjust but may detect smaller differences in the signals. However, these methods lack robustness in noisy environments. On the other hand, methods that focus on the auto-terms are more noise robust but cannot detect smaller differences. In order to make the methods more robust, different noise reduction techniques are created and validated. these techniques make the methods more robust but will lose accuracy in the classification