Ensemble learning methods and feature selection techniques for the development of a bioacoustic identification system
2016
Jangit, R.G.
Bioacoustic identification is a machine learning problem which employs audio recordings of animal vocalizations to identify a particular animal species. Specifically, it is a classification problem where the target class is presented by an animal species and the feature vector is composed of acoustic attributes of an animal's vocalization. In this study, an evaluation of three ensemble learning methods (bagging, boosting and random forest) using three-based component classifiers and features selection techniques (gain ratio, random forest and principal component analysis was conducted. The goal is to determine effective methods for the development of a bioacoustic identification web application. Tests were conducted on both segments and full recordings datasets with different number of classes and feature sets. Mel Frequence Cepstral Coefficient (MFCC) and other spectro-temporal features commonly employed as parametric representation of an acoustic signal on various automatic sound classification were examined. Results showed that higher classifier performances were achieved on segments dataset. Among the learning algorithms employed, random forest emerged as the best performing ensemble learning method with the highest recorded accuracy 0f 97.09 percent on a dataset with 16 classes of 70 MFCC.
显示更多 [+] 显示较少 [-]