Scream detection for home applications

Audio signal is an important clue for the situation awareness. It provides complementary information for video signal. For home care, elder care, and security application, screaming is one of the events people (family member, care giver, and security guard) are especially interested in. We present here an approach to scream detection, using both analytic and statistical features for the classification. In audio features, sound energy is a useful feature to detect scream like audio. We adopt the log energy to detect the energy continuity of the audio to represent the screaming which is often lasting longer than many other sounds. Further, a robust high pitch detection based on the autocorrelation is presented to extract the highest pitch for each frame, followed by a pitch analysis for a time window containing multiple frames. Finally, to validate the scream like sound, a SVM based classifier is applied with the feature vector generated from the MFCCs across a window of frames. Experiments of screaming detection is discussed with promising results shown. The algorithm is ported and run in a Linux based set top box connected by a microphone array to capture the audio for live scream detection.