Machine Learning based Speech Emotion Recognition in Hindi Audio

Abstract

Emotions are states of the nervous system that are connected to thoughts, sensations, behavioural reactions, and varying degrees of pleasure or annoyance. Failure to recognise emotions may result in communication failure. As a result, identifying emotion becomes crucial. Also, worldwide a total of about 577.8 million people speaks Hindi. Thus, this Speech Emotion Recognition system aims to detect emotion from Hindi audio. Here, the approach used is to extract audio-based as well as text-based features from the input audio speech to detect emotion. Experimentation is done with several machine learning algorithms like Random Forest, Support Vector Classifier and Logistic Regression applied to both audio and text datasets separately. The final model selected in each is the one that gave the highest accuracy respectively. 72% accuracy using Support Vector Classifier with GridSearchCV in the text dataset and 95% accuracy using Random Forest in the audio dataset is achieved. The combined results are used to identify four emotions which are neutral, angry, sad and happy, across a database of many different speakers.