Title:Unsupervised learning human's activities by overexpressed recognized non-speech sounds

Abstract:Human activity and environment produces sounds such as, at home, the noise produced by water, cough, or television. These sounds can be used to determine the activity in the environment. The objective is to monitor a person's activity or determine his environment using a single low cost microphone by sound analysis. The purpose is to adapt programs to the activity or environment or detect abnormal situations. Some patterns of over expressed repeatedly in the sequences of recognized sounds inter and intra environment allow to characterize activities such as the entrance of a person in the house, or a tv program watched. We first manually annotated 1500 sounds of daily life activity of old persons living at home recognized sounds. Then we inferred an ontology and enriched the database of annotation with a crowed sourced manual annotation of 7500 sounds to help with the annotation of the most frequent sounds. Using learning sound algorithms, we defined 50 types of the most frequent sounds. We used this set of recognizable sounds as a base to tag sounds and put tags on them. By using over expressed number of motifs of sequences of the tags, we were able to categorize using only a single low-cost microphone, complex activities of daily life of a persona at home as watching TV, entrance in the apartment of a person, or phone conversation including detecting unknown activities as repeated tasks performed by users.