Voice Emotion Recognition in Real Time Applications

摘要

This paper reports the results of voice emotion recognition in real time using machine learning models. The models are trained with some commonly used and well-known audio emotion datasets together with a custom dataset. This custom dataset was recorded from non-actor and non-expert people who were trying to imagine themselves in scenarios leading to arise of the related emotion. The reason for considering this important dataset is to make the model proficient in recognizing emotions in people who are not perfect in reflecting their emotions in their voices. The results from several machine learning classifiers while recognizing five emotions like anger, happiness, sadness, neutrality and surprise are compared. Models were evaluated with and without considering the custom data set to show the effect of employing an imperfect dataset. Our experiments showed that without using our custom dataset. the ensemble machine learning models such as gradient boosting, begging and random forest reach validation accuracies 89.82%, 88.58% and 84.83% respectively, which are higher than other evaluated models. After considering our custom dataset, again these ensemble methods obtained better accuracies of 87.34%. 86.71% and 82.98% respectively. This shows that although considering our custom dataset lowers the overall accuracy but empowers the model for predicting the emotions in everyday scenarios.