Description

The analysis of animal sounds or even communication is an emerging research topic, e.g. in biodiversity research, climate studies or digital farming. Considering animal sounds in a natural environment, it becomes clear, that the underlying signal processing may be quite challenging, e.g. by a low signal-to-noise ratio due to a large microphone distance or other acoustic peculiarities, e.g. additional sound sources. Furthermore, the classification of the signals depends on the availability and interpretability of appropriate (and annotated) sound data, e.g. representative recordings of dog barking in our contribution. We investigated, whether specific dog barking can be distinguished from silence or other sounds, like animal or traffic noise, to control a window-closing mechanism in a smart home scenario. The sound recordings have been collected and improved with a wavelet de-noising technique and notch filters. The analysis included varying analysis frames between 21 and 168 ms, and up to 8,239 temporal or spectral features that are reduced to a set of 51 features by a Linear Discriminant Analysis (LDA). Additionally, we applied a Correlation-based Feature Selection (CFS) method. We then classified the samples by various methods, namely AdaBoost, Random Forest, Support Vector Machine (SVM), Multi-layer Perceptron (MLP) and decision tree C4.5. Preliminary results show the best performance for a selection of all 51 features (after LDA) without any CFS, based on analysis frames of 21 ms. The described methods are useful to detect barking of one specific dog.