This paper presents a case study of cross-lingual transfer learning applied for affective computing in the domain of spoken dialogue systems. Prosodic features of correction dialog acts are modeled on a group of languages and compared with languages excluded from the analysis.

Speech from different languages was recorded in carefully staged Wizard-of-Oz experiments, however, without the possibility to ensure balanced distribution of speakers per language. In order to assess the possibility of cross-lingual transfer learning and to ensure reliable classification of corrections independently of language, we employed different machine learning approaches along with relevant acoustic-prosodic features sets.

The results of the experiments with mono-lingual corpora (trained and tested on a single language) and cross-lingual (trained on several languages and tested on the rest) were analyzed and compared in the terms of accuracy and F1 score.