The quadratic discrimination function (QDF) is a common method widely used in pattern recognition. It is optimal when the sampled features are normal with known means and covariance matrices. In this paper, a new transform is proposed and a method named modified QDF (MQDF) is deduced aim to improve the non-normality of the features during the speech emotion recognition, both prosodic and voice quality features are considered in the experiment. The results show that our method is effective and robust.