Classifying atypical affect, crying, heartbeat, self-assessed signal with boosted trees

In this paper we cover the four types of paralinguistic signal analysis, where the goal is to recognize the emotion of speech from disabilities, crying signals, heartbeat signals, and to predict the Likert scale of free speech. We propose simple and effective algorithm that enables the model to learn higher order interaction between features even from small size data sets, which has proven to be one of main problems that deep learning techniques and general paralinguistic tasks currently face. We first extracted emobase2010 feature and prosody feature then trained them on each corpus in a supervised scheme. Finally, we used the trained trees to deduce the appropriate label of each task. Evaluation is measured in terms of weighted average recall (WAR) and unweighted average recall (UAR) in every task.