S-vector: A discriminative representation derived from i-vector for speaker verification

Representing data in ways to disentangle and factor out hidden dependencies is a critical step in speaker recognition systems. In this work, we employ deep neural networks (DNN) as a feature extractor to disentangle and emphasize the speaker factors from other sources of variability in the commonly used i-vector features. Denoising autoencoder based unsupervised pre-training, random dropout fine-tuning, and Nesterov accelerated gradient based momentum is used in DNN training. Replacing the i-vectors with the resulting speaker vectors (s-vectors), we obtain superior results on NIST SRE corpora on a wide range of operating points using probabilistic linear discriminant analysis (PLDA) back-end.