Conversation clustering based on PLCA using within-cluster sparsity constraints

We propose a new method for detecting separate conversations between people. In this paper, to model the rules of turn-taking in conversation, we introduce sparsity constraints of temporal activities within each cluster into the probabilistic latent component analysis (PLCA). The proposed method can detect conversation groups by using PLCA on the within-cluster sparsity constraints although the conventional PLCA has no effectiveness in clustering. Our method has two features: First, it can be applied to the cases that more than two speakers participate in the same group, for which the within-cluster sparsity constraints can be defined. Second, it has the practical advantage that it requires no training phase. Despite the lack of any training phase, experimental results indicate that the proposed method remains effective in scenarios where three speakers participate in the same group.