Variational Gaussian Mixture Models for Speech Emotion Recognition

Abstract

In this paper applicability of variational methods for estimation of parameters of models used for speech emotion recognition is discussed.When the amount of data available is not adequate for training complex models, variational Bayesian method helps in training models with less amount of data. It also helps in determining the optimal complexity of the model. Our studies on Berlin emotional speech database show that variational methods perform better than maximum likelihood approach to estimate parameters of Gaussian mixture models used in speech emotion recognition.