Audio analysis for multimedia retrieval from a ubiquitous home

We present a system for video retrieval based on analyzing audio data from a large number of microphones in a home-like environment. Silence elimination on individual microphones is followed by noise reduction based on regions consisting of multiple microphones, to identify audio segments. An algorithm based on the energy distribution of sounds in the house is used to localize sound sources, thereby removing sounds heard in regions other than they are generated. A set of time domain features are used to classify these sounds for video retrieval. The algorithms were evaluated with 200 minutes of audio data from each microphone, gathered during an experiment where a family lived in the ubiquitous home. It was possible to achieve an overall accuracy of above 80% from all algorithms.