Using Oralstats for prosodic characterisation of speakers in different discourse genres

Abstract

This paper presents the prosodic and morphosyntactic analysis of 437 intonational phrases, 2777 words and 12520 phones registered from the speech of four Spanish politicians in four different discourse genres (TV gathering, press conference, political rally and interview). Based on methods from forensic linguistics, sociolinguistics, phonetics and computational linguistics, a set of R scripts called Oralstats has been developed, which allows to carry out a multivariate, dynamic and interactive analysis of acoustic and textual data aligned with PRAAT. By doing so, a new computational approach is suggested for finding individual forensic marks, combining both phonetic and linguistic factors, such as lexical or morphosyntactic values. Data exploration comes from the visualization of prosodic and morphosyntactic variables collected through decision trees, heatmaps and boxplots. The results show how speakers behave in a significantly similar way when it comes to differentiating themselves by genre; however, in confrontational genres some unexpected phonic marks were found, such as low intensity or low pitch, (the expected behavior being the opposite), thus enabling us to single out some of the speakers.