Drip detection through its acoustic signature with variable background noise

In this paper, the problem of detection of an specific event- a drip-in the domestic acoustic scene is addressed. The detection of noises generated by water can be of great interest in domestic scenes because it can help to prevent domestic floods. In the case of drip sounds, it is quite di cult to get real sounds covering the di erent sound qualities that di erent drops may have. Their sound depends on many factors as their size, the kind of surface they hit, etc. In order to approach real life as much as possible, a database including real and synthesized sounds have been created. A training set has been set up using the speed of variation of the MFCC, the kurtosis and the probability density function of the high frequencies as features to train a SVM classifier.