Adaptive energy detection for bird sound detection in complex environments

A new bird sound classification approach based on adaptive energy detection was proposed to improve the recognition accuracy of bird sounds in noisy environments. In this paper, the bird sounds with background noises were divided into three linear frequency bands according to their frequency distribution in spectrogram. The noise spectrum of each band was estimated and the existent probability of the foreground bird sound for each band was computed to serve for the adaptive threshold of energy detection. These foreground bird sound signals were detected and selected via adaptive energy detection from the bird sounds with background noises. Then, the features of Mel-scaled Wavelet packet decomposition Sub-band Cepstral Coefficient (MWSCC) and Mel-Frequency Cepstral Coefficient (MFCC) were extracted from the above signals for classification by using the classifier of Support Vector Machine (SVM), respectively. Moreover, the differences of recognition performance were implemented on 30 kinds of bird sounds at different Signal-to-Noise Ratios (SNRs) under different noisy environments, before or after adaptive energy detection. The results show that MWSCC has better noise immunity function, and the recognition performance after adaptive energy detection improves more significantly, indicating that it is a very suitable approach for the bird sound recognition in complex environments.