Automatic Recognition of Bird Songs Using Time-Frequency Texture

This paper presents a new approach for identifying birds automatically from their sounds, which first converts the bird songs into spectrograms and then extracts texture features from this visual time-frequency representation. The approach is inspired by the finding that spectrograms of different birds present distinct textures and can be easily distinguished from one another. In particular, we perform a local texture feature extraction by segmenting the bird songs into a series of syllables, which has been proved to be quite effective due to the high variability found in bird vocalizations. Finally, Random Forests, an ensemble classifier based on decision tree, is used to classify bird species. The average recognition rate is 96.5% for 10 kinds of bird species, outperforming the well-known MFCC features.