SEVERAL FEATURES FOR DISCRIMINATION BETWEEN VOCAL SOUNDS AND OTHER ENVIRONMENTAL SOUNDS (FriAmOR6)

Author(s) : Yuan-Yuan Shi (HCI Lab, Samsung Advanced Institute of Technology, China) Xue Wen (HCI Lab, Samsung Advanced Institute of Technology, China) Bin She (HCI Lab, Samsung Advanced Institute of Technology, China)

Abstract : Several features are found to discriminate between the vo-cals sounds and other environmental sounds. The vocal sounds include speaking, laughter, etc., 23 kinds of non-verbal and verbal sounds; and the environmental sounds are recorded in domestic environments. The discriminative features are selected from 22 kinds of features. They are the speech recognition features of LPCC and MFCC, time-spectral features from FFT, statistics of pitch values and contour, ratio of voiced and unvoiced segments, and spectrum of pitch contour. The 9 features cal-culated from pitch contours perform much better than the features calculated from spectrums, which show no dis-criminability. The classification is performed simply by a neural net-work to evaluate the performance of the 9 features. They are tested on a 21CDs environmental sound database. And the hit rate of 98.73% with the false alarm rate of 11% are ob-tained. The classification result confirms the effectiveness and efficiency of the features.

Menu