A Performance Evaluation of Machine Learning Algorithms for Emotion Recognition through Speech

This paper aims to recognize the human expressive states from their voice samples. It intends to extract a few reliable features and combine them intelligently for the said task for effective recognition. Initially, it extracts a few sub-band spectral properties from voice samples containing emotional information. Further, the pitch and its standard deviation along with the log-energy features have been extracted to develop an efficient combinational model. The chosen features are complementary, hence expected to increase the available emotional information. To validate the combinational framework, several Machine Learning Algorithms (MLAs) have been simulated and compared. Among the classifiers, the Random Forest (RF) has outperformed all others in terms of classification accuracy whereas the Decision Tree remains computationally least expensive.