In this paper, a novel approach for the design of cohort models for word spotting in continuous speech is presented. This new approach is based on modifying the probability density function of a conventional filler so that regions in the feature space that are related to the keyword will be reduced or removed. By modifying these regions, the filler and keyword models become more orthogonal in the sense that they represent different areas in the feature space, making the filler appropriate to be used as a cohort model. The algorithms, named Gaussian Subtraction (GS) and Gaussian Removal (GR), may be considered discriminative training algorithms.