Activity recognition using a spectral entropy signature

Context identification is one of the key challenges in Ubi- comp. An application example is providing contextual in- formation to caregivers of person with dementia to iden- tify assistance needs. Environmental audio provides sig- nificant and representative information of the context and the challenge is to automatically identify audio cues com- ing from overlapping sound sources without sophisticated microphone arrangements. My thesis proposes a succinct representation of the audio, based on the spectral entropy of the signal, and we show experimentally its robustness to source overlap and noise. This would permit ubiquitous applications that perform sound-based activity identification directly in mobile phones.