Spectrogram segmentation for bird species classification based on temporal continuity

Abstract

This article presents an enhanced approach for bird species classification from their recorded audio signals. Observing that textures of syllables in audio spectrograms have noticeable discerning capabilities among different bird species, we adopt these texture features for bird species classification. First, we compute spectrogram from recoded audio. We propose an enhanced syllable extraction technique to identify the syllables in the spectrogram. Texture features, based on gray level cooccurrence matrix (GLCM), are computed and used for classification using an ensemble learning method. We obtain satisfactory accuracy when the approach is tested on real audio recordings of 11 different bird species.