Publication

Abstract : Sentiment analysis is a natural language processing (NLP) technique used to classify a statement into three polarities, namely, positive negative and neutral. Thus, speech sentiment analysis invariably becomes a NLP task rather than a direct speech processing task, making the available speech recognition model just a tool to carry out NLP. Nevertheless, we do not have a way to directly use the pre-trained speech models to carry out sentiment analysis on speech utterances. The present study proposes to achieve this by directly using the speech signal to carry out sentiment analysis. We evaluate two such approaches on our custom made dataset consisting of movie and political reviews. The first approach used a fully connected neural network (FCNN) model and the second one used a 3-shot few shot learning (FSL) framework. For the FCNN model, the proposed framework provides a classification accuracy of 61.53%, whereas we get an accuracy of 99.83% for the 3-shot FSL framework. The performance is comparable to the current state-of-the-art (SOTA) system in place.