This study aimed to design a model for automatic identification of emotion from spontaneous communication using the acoustic characteristics of human speech. An experimental setup to collect and annotate call center Amharic telephone dialogs containing natural emotions is presented. These dialogs, involve 35 subjects (18 male and 17 female), are first manually decomposed into speaker turns and then segmented into intermediate chunks to be used as the analysis unit for feature calculation. Open class annotation is carried out by 3 professional psychologists and the various emotional states are mapped onto 4 cover classes, and a Majority Voting (MV) technique is applied to decide perceived emotion in each chunk. A total of 170 acoustic features consisting of prosodic, spectral and voice quality features are extracted from each chunk. An optimal feature set representing emotion (i.e. 33 all together) are selected through the use of generic algorithm and used to train Multilayer Perceptron Neural Network (MLPNN) classifier. A prototype application has been developed and the classification performance has been evaluated based on extracted features. Our preliminary speech emotion recognition model exhibits an average accuracy of 72.4% in identifying Anger, Fear, Positive and Sadness emotions.