Sports audio classification based on MFCC and GMM

Audio segmentation and classification can provide useful information for multimedia content analysis. In this paper, we present a approach to segment and categorize the sports audio into speech, music and other environmental sounds for sports video classification and highlight detection. We investigate the performance of mel frequency cepstral coefficients (MFCC) in a Gaussian mixture model frame work, and compare it to traditional short-time energy and zero-crossing rate feature. We achieve a correct identification close to 90% on MFCC with its first and second derivatives.