For ASR and many other applications, feature extraction transforms the signal to a feature vector that is modeled by the subsequent stochastic classifier. One of the shortcomings of the current feature extraction methods is that they are susceptible to recording channel and speaker characteristics. This has been addressed by learning a transformation in the initial feature vector t o suppress the detrimental factors. Linear and nonlinear trans- formations are studied. This thesis explores the transformation of features to functions of posterior probabilities of phonemes which are modeled by the subsequent stochastic classifier.