Sound-event partitioning and feature normalization for robust sound-event detection

The ubiquitous of smartphones has opened up the possibility of mobile acoustic surveillance. However, the continuous operation of surveillance systems calls for efficient algorithms to conserve battery consumption. This paper proposes a power-efficient sound-event detector that exploits the redundancy in the sound frames. This is achieved by a sound-event partitioning (SEP) scheme where the acoustic vectors within a sound event are partitioned into a number of chunks, and the means and standard deviations of the acoustic features in the chucks are concatenated for classification by a support vector machine (SVM). Regularized PCA-whitening and L2 normalization are applied to the acoustic vectors to make them more amenable for the SVM. Experimental results based on 1000 sound events show that the proposed scheme is effective even if there are severe mismatches between the training and test conditions.