This work demonstrates that, under limited data condition, it is indeed possible to improve the speaker recognition performance by controlled noise addition. The problem of limited data (<15 s) for training and testing is overcome to some extent by adding noise at very high Signal to Noise Ratio (SNR) values. The noise added versions may be viewed as different instances of the given data. Hence put together increases the number of feature vectors. The speaker identification study is conducted using randomly selected 100 speakers from TIMIT database, Mel-Frequency Cepstral Coefficients (MFCC) features and Gaussian Mixture Model (GMM)-Universal Background Model (UBM). The method provides performance of 78.20% using only limited data and 80% using both limited and noisy data. © 2010 Elsevier Ltd. All rights reserved.