Animal sound classification using Machine Learning

Kurzfassung

This paper describes the development of a monitoring solution utilizing acoustics data for the automated recognition and classification of wild animals. We implemented machine learning (ML) algorithms for classification fed with features extracted from audio file streams. Mel-frequency cepstrum coefficients (MFCCs), Low-level, Psychoacoustic, and Auditory filterbank temporal envelope features were applied as four subset features. The data used in this study comes from different sources: the data science community Kaggle, Macauley Library, xeno-canto, and the Dian Fossey Gorilla Fund. Together, these sources provided 2944 audio files on five taxa, including Birds (7.2% of files), cats (5.6% of files), domestic dogs (3.9% of files), Pan Troglodytes (chimpanzee) (66.4 % of files), Gorilla beringei (mountain gorilla) (1.6% of files), and wind (15.3 % of files) as background noise. Methodically, the results of Support Vector Machine (SVM), Random Forest, Decision Tree, and Naive Bayes are presented. ML algorithms were compared, considering the use of the macro average (F1 score) as a performance metric instead of accuracy due to the imbalanced nature of the dataset. SVM performed best (93%), followed by Random Forest (92%). The MFCC feature set performed well against all other feature sets combined.