Sentiment and Emotion Analysis for Effective Human-Machine Interaction during Covid-19 Pandemic

With the onset of Covid-19, interactions between humans and machines have increased at a rapid rate. Helping the machine identify the emotion and sentiment of the user plays a key role in making these interactions feel more natural. To do so, existing models for Speech Emotion Recognition (SER) and Sentiment Analysis (SA) focus on the detection of either only emotion or sentiment on acted databases. Unlike these existing works, this work presents a simple model with a comparatively small speech feature vector, to detect both emotion and sentiment from the spontaneous database, Multimodal Emotion Lines Dataset (MELD). This contains voice samples similar to those in a real-time environment. Speech features such as Mel Frequency Cepstral Coefficients (MFCC), Entropy, Teager Energy Operator have been extracted from the voice samples and are classified using Logit Boost, Logistic and Multiclass classifier. The performance of the model is improved by using feature selection techniques such as Backward elimination and Gaussian distribution coefficients. The proposed model is simple, and the results are comparable to existing work on the MELD database.