Comparison of feature performance in gunshot detection depending on noise degradation

Abstract

This paper compares three different features and various feature orders for the purpose of determining the best feature for gunshot detection under adverse noise condition. Compared features cover LPC, LPCC and MFCC with orders from 8 to 30. All features were extracted from sounds with the sound-to-noise ratios 30, 20, 10, and 0 dB. The background noise was simulated by white noise. Experimental results indicate that LPC coefficients are the most efficient features, especially for low noise. On the other hand, MFCC performed well in noisy environments at 10 dB and 20 dB.