Gammatone Wavelet features for sound classification in surveillance applications

Sound can deliver highly informative data about the environment, which can be of particular interest for home-teleassistance and surveillance purposes. In the sound event recognition process, the signal parameterisation is a crucial aspect. In this work, we propose Gammatone-Wavelet features (GTW) by merging Wavelet analysis, which is well-suited to represent the characteristics of surveillance-related sounds, and Gammatone functions, which model the human auditory system. An experimental evaluation that consists of classifying a set of surveillance-related sounds employing Support Vector Machines has been conducted at different SNR conditions. When compared to typical Wavelet analysis with Daubechies mother function (DWC), the GTW features show superior classification accuracy both in noiseless conditions and noisy conditions for almost any SNR level. Finally, it is observed that the combination of DWC and GTW yields the highest classification accuracies.