The IBM 2016 Speaker Recognition System

摘要

In this paper we describe the recent advancements made in the IBM i-vectorspeaker recognition system for conversational speech. In particular, weidentify key techniques that contribute to significant improvements inperformance of our system, and quantify their contributions. The techniquesinclude: 1) a nearest-neighbor discriminant analysis (NDA) approach that isformulated to alleviate some of the limitations associated with theconventional linear discriminant analysis (LDA) that assumes Gaussianclass-conditional distributions, 2) the application of speaker- andchannel-adapted features, which are derived from an automatic speechrecognition (ASR) system, for speaker recognition, and 3) the use of a deepneural network (DNN) acoustic model with a large number of output units (~10ksenones) to compute the frame-level soft alignments required in the i-vectorestimation process. We evaluate these techniques on the NIST 2010 speakerrecognition evaluation (SRE) extended core conditions involving telephone andmicrophone trials. Experimental results indicate that: 1) the NDA is moreeffective (up to 35% relative improvement in terms of EER) than the traditionalparametric LDA for speaker recognition, 2) when compared to raw acousticfeatures (e.g., MFCCs), the ASR speaker-adapted features provide gains inspeaker recognition performance, and 3) increasing the number of output unitsin the DNN acoustic model (i.e., increasing the senone set size from 2k to 10k)provides consistent improvements in performance (for example from 37% to 57%relative EER gains over our baseline GMM i-vector system). To our knowledge,results reported in this paper represent the best performances published todate on the NIST SRE 2010 extended core tasks.