Emotional State Prediction Using Speech Signals

Abstract

In this work, a methodology to estimate the emotional state of humans through the voice is presented using the anger, boredom, disgust, anxiety, fear, happiness, sadness, surprise and neutral emotions. The operation of our system is based on the Partially Observable Markov Processes better known as POMDP. POMDP main problem is to decide what policy-action to take in one iteration in order to approach from a source state to a goal while obtaining the maximum reward in each iteration. POMDP were used, because it allows us to know through states of beliefs how the emotions expressed in a conversation can influence the emotional state of a different person. To know the emotional levels of a person, we start with a uniform distribution on emotions. Each time new information is obtained an therefore the belief state must be updated. Finally, this paper presents two strategies to update the state of belief, which are divided into exact and approximate algorithms to obtain the best performance based on precision and speed in real time.