Finding two-level interpersonal context: proximity and conversation detection from personal audio feature data

We propose a method to detect adhoc meeting based on crosscorrelation between audio feature data, which are collected from personal mobile terminals. This method can detect whether there is conversation between each pair of users without raw audio data. Through a two-day evaluation with eight users, we found our method could detect meeting contexts with 0.9 F-measures on average. We also introduce example applications such as a document search application in which detected meeting context is used as a file annotation.