SUT Submission for NIST 2016 Speaker Recognition Evaluation: Description and Analysis

SUT Submission for NIST 2016 Speaker Recognition Evaluation: Description and Analysis Hossein Zeinali, Hossein Sameti and Nooshin Maghsoodi Department of Computer Engineering Sharif University of Technology, Tehran, Iran zeinali@ce.sharif.edu, sameti@sharif.edu, nmaghsoodi@ce.sharif.edu Abstract In this paper, the most recent Sharif University of Technology (SUT) speaker recognition system devel- oped for NIST 2016 Speaker Recognition Evaluation (SRE) is described. The major challenge in this evaluation is the language mismatch between training and evaluation data. The submission is related to the fixed condition of NIST SRE 2016 and features a full description of the database and the systems which were used during the evaluation. Most of the systems are developed in the context of the i-vector framework. Preconditioning the i-vectors, score normalization methods and the classifier used are dis- cussed. The performance of the systems on the development and evaluation parts of the NIST 2016 SRE16 dataset are reported and analyzed. The best obtained minimum and actual DCF are 0.666 and 0.740, respectively. This is achieved by score fusion of several systems and using different methods for mismatch compensation. Keywords: Speaker verification, NIST SRE 2016, SUT, i-vector, PLDA The 2017 Conference on Computational Linguistics and Speech Processing ROCLING 2017, pp. 276-286  The Association for Computational Linguistics and Chinese Language Processing 276 1. Introduction During the past two decades, National Institute of Standards and Technology (NIST) has organized sev- eral speaker recognition evaluations (SRE). The goals of these evaluations are exploring new ideas in speaker recognition and optimizing speaker recognition systems. Like all SREs, in the SRE16 some challenges are followed. One of them is the mismatch between training and evaluation datasets. Due to attention that most of the provided training data is in English while the evaluation data is in Cantonese and Tagalog, efficient methods are required for reducing the effects of this mismatch. The second chal- lenge is short duration enrollment and test utterances. This challenge more happens for test utterances where their duration varies from 10 to 60 seconds. The last challenge is imbalanced multi-session train- ing. In fact, there are two enrollment conditions for SRE16: three segments available for training some speaker models while only one segment for others. The focus of SRE16 is on the telephone speech in Cantonese and Tagalog languages. In this paper, we provide the description of our system and analyze the results of using different features sets, different Voice Activity Detection (VAD) systems and methods for preconditioning the i-vectors. Our contrastive system is constructed by combining 5 subsystems that each of them is an i- vector based system. The subsystems differ from each other in terms of input features (i.e. MFCC, PLP, SBN or Perseus) or applied VAD method (i.e. FVAD or EVAD). We have developed two sets of these 5 systems with and without labeled data (i.e. Contrastive 1 and 2). The final system is constructed by fusing these two sets. The first version of our system description without any analysis on the evaluation data can be found in [1]. The rest of …

Content truncated to 500 words