Feature learning for bird-call segmentation using phase based features

In this paper, we extend an existing algorithm for the segmentation of bird calls from the background. By utilizing features which has information about the magnitude and phase of the Fourier transform, we demonstrate an improvement in segmentation performance as compared to magnitude-only features. The proposed method utilizes a dictionary learnt from the time-frequency representation. The coefficients obtained by projecting a recording on to this dictionary is used to estimate Reney entropy between bird vocalization and background. The proposed method obtains an improvement of 25 percent as compared to similar features derived from conventional magnitude-based spectrogram.