Mean shift algorithm for exponential families with applications to speaker clustering

Themos Stafylakis, Vassilis Katsouros, Patrick Kenny and Pierre Dumouchel

 

---

Abstract

This work extends the mean shift algorithm from the observation space to the manifolds of parametric models that are formed by exponential families. We show how the Kullback-Leibler divergence and its dual define the corresponding affine connection and propose a method for incorporating the uncertainty in estimating the parameters. Experiments are carried out for the problem of speaker clustering, using both single Gaussians and i-vectors.

Keywords

Text-Independent Speaker Recognition
Commercial Applications