Cosine distance features for improved speaker verification
Authors:
Open Access
- 1 June 2015
- journal article
- research article
- Published by Institution of Engineering and Technology (IET) in Electronics Letters
- Vol. 51 (12) , 939-941
Abstract
Similarities are used with people known already as a means to enhance speaker verification accuracy. Motivated by this, experimental work has been conducted regarding the use of cosine distance (CD) similarity with respect to a set of reference speakers, CD features, with a back‐end support vector machine (CDF‐SVM) classifier for speaker verification. A state‐of‐the‐art i‐vector with CD scoring (i‐CDS) is used as the baseline system for the experiments and for the computation of CD similarity. Experimental results on the telephone speech of the core short2‐short3 conditions of NIST 2008 speaker recognition evaluation (SRE), for female, male and both‐gender trials, show that the proposed CDF‐SVM outperforms the baseline i‐CDS system. The CDF‐SVM achieved an absolute improvement of 1.16% in equal error rate (EER) and 0.38% in minimum DCF over the baseline i ‐CDS for female trials. Similar performance improvements were also obtained for the male and all‐gender trials of the SRE. Finally, fusing the CDF‐SVM with i ‐CDS gave the best overall performance, an absolute improvement of 4.19% in EER and 1.99% in minimum DCF, over the individual CDF‐SVM system performance for the all‐gender trials. Similar performance improvements were also achieved for male and female trials.Keywords
This publication has 4 references indexed in Scilit:
- Boosting the Performance of I-Vector Based Speaker Verification via Utterance PartitioningIEEE Transactions on Audio, Speech, and Language Processing, 2013
- LIBSVMACM Transactions on Intelligent Systems and Technology, 2011
- Utterance partitioning with acoustic vector resampling for GMM–SVM speaker verificationSpeech Communication, 2011
- Front-End Factor Analysis for Speaker VerificationIEEE Transactions on Audio, Speech, and Language Processing, 2010