Automatic Bird Vocalization Identification Based on Fusion of Spectral Pattern and Texture Features

Abstract

Automatic bird species identification from audio field recordings is studied in this paper. We first used a Gaussian mixture model (GMM) based energy detector to select representative acoustic events. Two different feature sets consisting of spectral pattern and texture features were extracted for each event. Then, a ReliefF-based feature selection algorithm was employed to select distinguishing features. Finally, classification was performed using support vector machine (SVM). The main focus of the proposed method lies in the fusion of a spectral pattern feature with several texture descriptors, which extends our previous work. Experiments used an audio dataset comprised of field recordings of 11 bird species, containing 2762 bird acoustic events and 339 detected “unknown” events (corresponding to noise or unknown species vocalizations). Experimental results demonstrate superior classification performance compared with that of the state-of-the-art method, which renders the proposed method more suitable for real-field recording analysis.