Abstract

In this paper, we propose deep neural network (DNN) architecture with attention mechanism (DNN-WA) for\nemotion recognition (ER). DNN-WA is an utterance level classification mechanism which accounts for the long-term\ndependencies within an utterance unlike the conventional frame level classification. Mel-frequency cepstral coefficients\n(MFCC) are used to represent the emotion information within the spoken utterance. To incorporate additional temporal\ninformation at feature level, shifted delta cepstra (SDC) operation is performed on frame based MFCC features. The studies\non ER are carried out using Berlin database. The results of our studies show that, the DNN outperforms the baseline GMM\nsystem indicating a better representation capability. Further, DNN-WA outperforms the DNN based system. From this, it is\nevident that DNN-WA indeed captures the contextual information which is essential for ER.