Sparse feature extraction and sound source classification for impact sounds

The sparse representation is used in source recognition of impact signals, aiming at classification more efficiently and exactly for signals under low SNR environment. It is shown that the sparse features perform better than the traditional audio features, MFCCs. The Orthogonal Matching Pursuit (OMP) is applied on three types of predefined dictionaries as well as a learned dictionary based on training signals respectively to obtain sparse representation of recorded impact sounds. The sparse features are extracted for recognition of the sound sources with different SNRs, and the performances are compared with that of MFCCs. The experimental results on the classification of 12 types of materials show that sparse features perform better than MFCC in almost all cases. Specifically, sparse features have better anti-noise performance under low SNR environment.