Speaker identification using diffusion maps

In this paper we propose a data-driven approach for speaker identification without assuming any particular speaker model. The goal in speaker identification task is to determine which one of a group of known speakers best matches a given voice sample. Here we focus on text-independent speaker identification, i.e. no assumption is made regarding the spoken text. Our approach is based on a recently developed manifold learning technique, named diffusion maps. Diffusion maps enable embedding of the recording into a new space, which is likely to capture the speech intrinsic structure. The algorithm is tested and compared to common identification algorithms. Experimental results show that the proposed algorithm obtains improved results when few labeled samples are available.