Developing “voice care”: Real-time methods for event recognition and localization based on acoustic cues

Abstract

This paper presents methods for sound recognition in a living space and ways to track the location of the sound sources. Algorithms were developed so sound recognition and localization can both be performed in real time. The sound recognition method is based on Gaussian mixture modeling with outlier rejection. The sound source localization method is based on multiple signal classification (MUSIC) and it borrows the idea of particle filtering to confine the estimation error. Estimates of the sound source location can be successively refined by Kalman filtering. The recognition method was tested with real recordings and achieved > 90% of accuracy in distinguishing 8 classes of sounds while keeping both the false-acceptance and the false-rejection rates below 20%. The localization method was tested in real time and demonstrated the capabilities to track a sound source moving at about 0.3 m/s. These results indicate that the methods, when integrated, can be deployed to the home for acoustic event detection purposes.