Robot Emotional State through Bayesian Visuo-Auditory Perception

Abstract

In this paper we focus on auditory analysis as the sensory stimulus, and on vocalization synthesis as the output signal. Our scenario is to have one robot interacting with one human through vocalization channel. Notice that vocalization is far beyond speech; while speech analysis would give us what was said, vocalization analysis gives us how was said. A social robot shall be able to perform actions in different manners according to its emotional state. Thus we propose a novel Bayesian approach to determine the emotional state the robot shall assume according to how the interlocutor is talking to it. Results shows that the classification happens as expected converging to the correct decision after two iterations.

Synthesis of Emotions on a <em>Human-Robot-Interactive</em> Platform

“…Progresses were made in [23] and [10] however they claim to have a high percentage of accuracy, they are not real-time. In this paper we are using our own model previously proposed on [15] witch is a real time system for auditory emotional classification, though it is well suitable for our purposes.…”

“…Most state-of-the-art in the area of emotive robots do not run in real-time, being then still inapplicable to real cases of human-robot-interaction applications. However recent researches like [20] and [15] shows result about two Bayesian classifiers inside a structure for human-robot-interaction that are applicable to HRI in real-time. These classifiers have both the purpose to classify human emotional state among the scope {anger,fear,sad,neutral and happy}.…”

“…This paper builds on [19,15] where the used classifiers were defined; and also builds on [4] where the five-factor model personality traits was defined. Based on this mentioned model, we reinterpret the "openness to experience" SBP presented on [4] as Humorous personality.…”

“…For the detection of features and classification of facial expressions we are using our previously proposed methods from [18]. For the classification of vocal expressions we are using a Bayesian model previously proposed by us in [15]. Both classifiers give a confidence score for emotions among the scope of {Anger, Fear, Sad, Neutral, Happy}.…”

“…Most state-of-the-art in the area of emotive robots do not run in real-time, being then still inapplicable to real cases of human-robot-interaction applications. This paper builds on the work of [19,15] where real-time Bayesian classifiers were presented for visual signal and to auditory signal; independently and respectively. Both of these classifiers give output among the scope {happy, sad, fear, neutral, anger}.…”