Specific environmental sounds recognition using time-frequency texture features and random forest

Traditional approaches to environmental sounds recognition used acoustic features merely based on time domain or frequency domain. In this paper, a new feature descriptor that uses image texture information is proposed to identify specific environmental sounds based on the recognition of fixed-duration sounds segments where their corresponding spectrums are viewed as gray-level images. The proposed specific environmental sounds recognition system firstly conducts short-time spectrum estimation algorithm to the noisy sounds segments, and then extracts 5 time-frequency texture features descriptors(TFD) from the enhanced spectrum using sum and difference histogram (SDH), in the last place, applies random forest(RF) to make classification and recognition. The average recognition rate is 92.5% for 51 kinds of environmental sounds, outperforming the well-known MFCC features; meanwhile, it is robust to noise.