Analysis of Posterior Estimation Approaches to I-vector Extraction for Speaker Recognition

The i-vector approach to speaker recognition requires estimating Sufficient Statistics (SS) (i.e., zerothand firstorder statistics) for a given utterance of speech with respect to a Universal Background Model (UBM) usually represented by Gaussian Mixture Models (GMM). To estimate SS, alternate approaches have also been experimented. Studies suggest that using acoustic phone posteriors estimated from a Deep Neural Network (DNN) based Automatic Speech Recognition (ASR) system can be useful in estimating accurate speaker representations with ivectors. In this paper, we analyze and compare the UBM-GMM and several versions of DNN approaches together with subspace Gaussian Mixture Models to estimate i-vectors for a speaker. We show that better alignments of speech frames can lead to superior speaker verification performance. This is achieved through the use of the decoded output from the ASR system, whereas existing systems only use posteriors at the output of the DNN directly. The posteriors from the decoding lattices are rescaled suitably to deal with its sparse nature that can affect SS computation. We show that a direct correlation exists between senone recognition accuracy of the system generating the posterior and the performance of corresponding speaker recognition systems. The posterior estimation methods are compared on standard NIST 2010 SRE dataset. Significant improvements are obtained when using the ASR decoder, thereby confirming that with better frame-level alignments speaker verification performance improves. Equal Error Rate (EER) as low as 0.9% is achieved on the telephone condition of the evaluation set.