THU-EE System Description for NIST LRE 2015

This paper describes the systems developed by the Department of Electronic Engineering of Tsinghua University for the NIST Language Recognition Evaluation 2015. We submitted one primary and three alternative systems for the fixed training data evaluation and didn't take part in the open training data evaluation for our limited data resources and computation capability. Both the primary system and three alternative systems are fusions of multiple subsystems. The primary system and alternative systems are identical except for the training, development and fusion data. The subsystems are different in feature, statistical modeling or backend approach. The features of our subsystems include MFCC, PLP, TFC, PNCC and Fbank. The statistical modeling of our subsystems can be roughly categorized into four types: i-vector, deep neural network, multiple coordinate sequence kernel (MCSK) and phoneme recognizer followed by vector space models (PR-VSM). The backend approach includes LDA-Gaussian, SVM and extreme learning machine (ELM). Finally, these subsystems are fused by the FoCal toolkit. Our primary system is presented and briefly discussed. Post-key analyses are also addressed, including comparison of different features, modeling backend approaches and a study of their contribution to the whole performance. The processing speed for each subsystem is also given in the paper.