This paper describes a novel procedure for determining a set of acoustic measurements for use in a segment-based speech recognition system. Rather than fully specifying a set of acoustic measurements, a set of generalized measurements such as the average spectral amplitude and the movement of the spectral prominences, is devised. The free parameters associated with these generalized measurements are determined through an optimization procedure, with the help of a large body of training data. Several controlled phonetic classification experiments are presented to illustrate the reduction in error rate as more complex measurements are incorporated. This technique has been successfully utilized in the development of our summit speech recognition system.