A Method for Creating a Criterion-referenced Assessment of Bird Vocalization Recognition

This report presents an analysis of the feasibility of using a method to recognize bird vocalizations. The automated extraction of information from sound recordings collected over extended periods of time is reviewed and a method proposed to improve the scope of biodiversity monitoring. For this purpose, birdsong is automatically segmented into acoustic events and subsequently, for each segment, the areas of increased energy density are determined. After framing the resulting areas, the coordinates and values of the local energy density maxima are determined. Different acoustic events of the same bird song species have varying contour lengths referring to the same syllable type. Therefore, the most compressed information representation of the cycle through a given parametric model is sought. An interpolation model based on GMM with M components is chose. The model parameters are then fed to a classifier to recognize the syllable type of the bird vocalization. The 78 syllables obtained during signal segmentation are classified with ninety-eight percent accuracy, the recognition method and the type of acoustic event are the determining factor.