Robust scream sound detection via sound event partitioning

摘要

This paper proposes a robust scream-sound detection scheme for acoustic surveillance applications. To enhance the discriminability between scream and non-scream sounds, a sound-event partitioning (SEP) method that facilitates the extraction of multiple acoustic vectors from a single sound event is developed. Regularized principal component analysis (PCA) and normalization are applied to the acoustic vectors, which are then classified by support vector machines (SVMs). Experimental results based on 1000 sound events show that the proposed scheme is effective even if there are severe mismatches between the training and testing conditions. The experimental results also show that the proposed scheme can reduce the equal error rate (EER) by up to 60 % when compared to a classical approach that uses mel-frequency cepstral coefficients (MFCC) as features. Extensive analyses on different processing stages of the proposed sound detection scheme also suggest that sound partitioning and feature normalization play important roles in boosting the detection performance.