Domain adaptation and classification on bird noises in the SINGA:PURA urban polyphonic dataset

Abstract

This paper makes use of the SINGA:PURA Urban Polyphonic Dataset to study the effectiveness of different methods of audio data classification in relation to the domain sensitivity of classifier performance. Audio files were classified according to the label taxonomy in the SiNGA;PURA dataset. The approach taken compares the performance of a logistic regression classifier to that of a Convolutional Neural Network (CNN) classifier, as well as to a Domain-Adversarial Neural Network (DANN) model on classification tasks in situations where domain data is available and vice versa. Some other factors affecting classification performance are also discussed.