Speaker Diarization aims at inferring who spoke when in an audio stream and involves two simultaneous unsupervised tasks: (1) the estimation of the number of speakers, and (2) the association of speech segments to each speaker. Most of the recent efforts in the domain have addressed the problem using machine learning techniques or statistical methods (for a review see [11]) ignoring the fact that the data consists of instances of human conversations. When humans want to use language to communicate orally with each other, they are faced to a coordination problem. "Avoidance of collision is one obvious ground for this coordination of actions between the participants. In order to coordinate efficiently and successfully, they will therefore have to agree to follow certain rules of interaction" [8]. One such rule is that no one monopolizes the floor but the participants take turns to speak. This concept is called turn-taking. The computational linguistic literature is rich on the analysis of human conversations; the seminal work of [9] shows that conversations obey to predictable interactions pattern between participants and a speaker turn is related in predictable ways to the previous and next turn and follows a structure similar to a grammar. In between the social phenomena that regulates the turns in a conversation, lot of attention has been devoted to roles. In fact people interact in different ways depending on the context of the environment but "Their interactions involve behaviors associated with defined statuses and particular roles. These statuses and roles help to pattern our social interactions and provide pre-dictability" [10].
展开▼
