Evaluation of the I-vector system for text-dependent speaker verification

• Published in: 2017 11th IEEE International Conference on Anti-counterfeiting, Security, and Identification (ASID) pp. 60 - 63
• Main Authors: Li, Lin, Guo, Huiyang, Shang, Fengyi, Hong, Qingyang, Liu, Kai
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-10-2017
• Subjects: Feature extraction; GMM-SVM; Hidden Markov models; i-vector; Integrated circuits; Mel frequency cepstral coefficient; speaker verification; Support vector machines; text-dependent; tied mixture model(TMM); Training; Viterbi algorithm
• ISSN: 2163-5056
• DOI: 10.1109/ICASID.2017.8285744

Text-dependent speaker verification systems can be deployed for more and more applications of security authorization, in which both speaker identity and lexical content should be verified together. There are several methods to model the utterances with short duration, such as Gaussian mixture model-universal background model (GMM-UBM), tied mixture model-UBM(TMM-UBM) and GMM-support vector machines(GMM-SVM). In this paper, we build an efficient i-vector based system and compare it with other systems based on the RSR2015 and King-ASR-L-057 corpus. Experimental results showed that the EER value of imposter-correct(IC) could be reduced by 40.3% for King-ASR-L-057 and 63.8% for RSR2015 in contrast to the traditional GMM-UBM system.