At present, most speaker identification systems use cepstral or linear prediction (IP) based features. The performance of these systems degrades significantly with the presence of noise in the training and/or the testing speech. To improve this performance, higher order statistics (HOS) or cumulant-based LF features are proposed. Using these features and singular value decomposition (SVD), it is shown that the performance of a speaker identification system can be improved considerably in the presence of additive white or colored Gaussian noise.