Varta Rasa - A Simple and Accurate System for Emotion Recognition in Conversations

Abstract

Emotion Recognition in Conversation (ERC) has become one of the most explored topics in Speech Technology. Unlike Emotion Recognition in a single utterance, context plays an important role in Emotion Recognition in Conversation. In literature complex architectures are proposed for ERC which makes the developement of real-time applications difficult. In this paper, we propose a simple architecture called Varta Rasa (in Sanskrit, Varta means conversation and Rasa means emotions) for ERC, which is a stacked ensemble model. Extensive experiments on the IEMOCAP dataset generated results comparable to state-of-the-art models while providing significantly lower complexity.