A Comparative Cross Language View On Acted Databases Portraying Basic Emotions Utilising Machine Learning

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1917–1924 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1917 A Comparative Cross Language View On Acted Databases Portraying Basic Emotions Utilising Machine Learning F. Burkhardt1,2, A. Hacker1, U. Reichel2, H. Wierstorf 2, F. Eyben2, B.W. Schuller 2,3,4 1Technical University of Berlin, 2audEERING GmbH, 3University of Augsburg, 4Imperial College London Germany, Germany, Germany, UK {f.burkhardt, anabell.hacker}@tu-berlin.de {ureichel, hwierstorf, fe, bs}@audeering.com Abstract For several decades emotional databases have been recorded by various laboratories. Many of them contain acted portrays of Darwin’s famous “big four” basic emotions. In this paper, we investigate in how far a selection of them are comparable by two approaches: on the one hand modeling similarity as performance in cross database machine learning experiments and on the other by analyzing a manually picked set of four acoustic features that represent different phonetic areas. It is interesting to see in how far specific databases (we added a synthetic one) perform well as a training set for others while some do not. Generally speaking, we found indications for both similarity as well as potential language-specific differences. Keywords: emotional, database, language resource evaluation 1. Introduction The way we speak is influenced by many factors. One of them is the emotional expression, a field that has been studied by numerous researchers over the past four decades (Schuller and Schuller, 2021; Gangamo- han et al., 2016; Schr¨oder, 2001). Many of these studies recorded their own emotional speech data, often acted versions of the “big four basic emotions”: anger, hap- piness, sadness and fear. In the theory of a “basic emotion” concept, these are the emotions that should be interpretable in a cultur- ally universal manner and thus should be expressed in a comparable manner. In reality emotional expressions are influenced by culture and context (Scherer et al., 1999; Burkhardt et al., 2006; Barrett et al., 2019). In parts the outcomes of this investigation may answer the question, in how far the language (and underlying cul- ture) of the speech resources influenced the acoustical correlates of the emotional portrayals, as has been done in, for example, (Tamuleviˇcius et al., 2020; Neumann and Vu, 2018; Feraru et al., 2015; Schuller et al., 2010). To investigate the matter, we analysed several databases from the literature that included at least the basic emotions neutral, anger, happiness, and sadness with respect to a set of hopefully meaningful acoustic parameters that we extract algorithmically. In a first analysis step, we performed cross data super- vised machine learning. Each database was used once as a training set and once as a test set. The underlying idea is that databases that portray culturally universal basic emotions should be able to act as a training or test set in cross database machine learning. We can then estimate the performance to learn emotional ex- pression as portrayed in the other databases, thus re- sulting in an estimate of similarity. We added one syn- thesised database …

Content truncated to 500 words