Selection of negative pool for PR-SVM in language recognition

This paper presents a simple approach of negative pool selection to improve the performance of SVM-based phonotactic language recognition system. In this work, how to find a better negative pool of SVM is explained and developed. For evaluating the approach, some open software (HTK, SRILM) and a phone decoder of English are used and experiments were carried out on the NIST-LRE2011 database. The proposed approach outperforms the baseline phonotactic systems considering n-grams up to n=3. In those cases, average of the greatest 24 costs were obtained by generalized sequence forward selection (GSBS) algorithm: 4.02%, 10.45%, 20.34% for the 30s, 10s and 3s closed-set test conditions (meaning 24.01%, 12.97%, 14.82% relative reduction to baseline system, respectively), proving that the method of negative pool selection can improve the performing of SVM.