In this paper, an evolutionary algorithm is used to select an optimal set of acoustic features for emotional speech recognition. A new algorithm that combines differential evolution (DE) optimization and linear discriminant analysis (LDA) is proposed to design an effective feature selection and classification model. An original acoustic feature framework based on auditory modeling is also presented. The auditory-based features are provided as inputs to the DE-LDA based emotional speech recognition system. To evaluate the effectiveness of the DE-LDA approach, a subset of the Emotion Prosody Speech and Transcript corpus covering five emotional states (happiness, anger, panic, sadness, and interest) is used throughout the experiments. The results show that the proposed DE-LDA model performs significantly better than the baseline systems. It achieves a classification rate of 91.6% using only 50 input parameters that are optimally selected from 128 original acoustic features.