Abstract: A recognition method based on a multi-scale subband energy feature set for muzzle blast is proposed to detect gunshot in real complicated acoustic environment. After muzzle blast signal is enhanced with a group of low-frequency subband filters, the muzzle blast detection followed by a crest search procedure is applied to determine the start position of candidate muzzle blast. Based on the start position, a set of nested data fragments with different lengths are segmented from the candidate muzzle blast and its subband components, and the short-time energies of data fragments and the corresponding energy ratios are then obtained to form the multi-scale subband energy set. The energy set is used as the feature vector for either support vector machine (SVM) or k-nearest neighbor (kNN) classifiers for the classification of muzzle blast and non-muzzle blast. The experimental results with 372 field recordings of gunshot demonstrate that both the precision and recall ratios of the proposed method are larger than 93%, and F1 score is better than 0.95. The proposed method shows less difference in recognition performances in using both SVM and kNN classifiers, with feature dimension and time consumption being much lower than those of the traditional wavelet-based means, indicating that it is more approaching practical applications. Key

Key words: gunshotdetection, muzzleblastrecognition, multi-scalesubbandenergyset, supportvectormachine, k-nearestneighbor