Ecological Sound Events Classification Based on Time-Frequency Features

In order to improve the accuracy of environmental sound events recognition under non-stationary noise, a sound events recognition method based on time-frequency analysis was proposed. In this method, power spectrum of the signal was firstly output in the form of time-frequency diagram. Then, sum and difference histograms(SDH) was adopted to calculate texture features. Afterwards, the matching pursuit(MP) algorithm was employed to obtain effective time-frequency features to supplement the texture features to yield higher recognition accuracy. Two experiments were conducted to demonstrate the effectiveness of these joint features for unstructured environmental sound events classification. The results show that the method can produce good performance, and is robust to noise.