Comparison of Large–scale SVM Training Algorithms for Language Recognition