RNN with Improved Temporal Modeling for Speech Emotion Recognition

Emotions are a natural part of daily human activities and play a key role in human decision-making, interactions, and cognitive processes. In recent years, an automatic speech emotion recognition has aroused great interest due to its wide potential use in environments where machines need more natural human interaction or monitoring (i.e., detection of stress levels in pilots). In this work, we focused on exploring new ways of temporal modeling of recurrent neural network (RNN) architectures. The proposed solution leverages contextual information from the speech sequence using recurrent model inspired by Siamese RNN. Compared to a baseline LSTM model, the proposed solution achieved a relative improvement of accuracy of approx. 5% on the IEMOCAP database.