Comparison of low-dimension speech segment embeddings: Application to speaker diarization

Segment clustering is a crucial step in unsupervised speaker diarization. Bottom-up approaches, such as, hierarchical agglomerative clustering technique are used traditionally for segment clustering. In this paper, we consider the top-down approach to clustering, in which a speaker sensitive, low-dimensional representation of segments (speaker space) is obtained first, followed by Gaussian mixture model (GMM) based clustering. We explore three methods of obtaining the low dimension segment representation: (i) multi-dimensional scaling (MDS) based on segment to segment stochastic distances; (ii) traditional principal component analysis (PCA), and (iii) factor analysis (i-vectors), of GMM mean super-vectors. We found that, MDS based embeddings result in better representation and hence result in better diarization performance compared to PCA and even i-vector embeddings.