Comparison of machine learning techniques for the automatic recognition of soundscapes

Rather than recognising single environmental sound sources, research on soundscape recognition aims at identifying general unstructured auditory scenes composed of multiple simultaneous sound sources. Nowadays, soundscape recognition has multiple applications, such as: i) contextawareness for mobile robots and portable devices, so as to provide information about the surrounding acoustic environment, enabling the automatic reaction of the device without human intervention; ii) background noise identification for speech recognition systems, so as to improve the robustness of such systems by facilitating their adaptability to any indoor or outdoor environment; iii) support tool in urban planning and noise annoyance assessment. In this paper, pattern recognition techniques specially designed to automatically recognise soundscapes are implemented. Specifically, the study focuses on finding the most appropriated learning paradigm. A set of machine learning techniques, commonly employed in similar sound recognition tasks, are considered: decision trees, K-Nearest Neighbours algorithm, Gaussian Mixture Models and Neural Networks. Extensive experiments are carried out so as to empirically test and compare the performance attained by the different learning techniques. The experiments are performed employing an audio data corpus composed of 15 different soundscenes, such as park, traffic street, restaurant or stadium.