Modeling phonetic pattern variability in favor of the creation of robust emotion classifiers for real-life applications

► Phoneme level emotion classification is evaluated within cross corpora evaluations. ► Phonetic pattern dependent and independent emotion classifiers were tested within equal evaluation conditions and acoustic features. ► We showed how emotion models trained on spontaneous emotion annotated in VAD space could show quite good classification performance for acted emotion annotated with emotional categories. ► We showed importance of phonetic pattern analysis for creation of robust emotion classifiers for real life applications. ► Classification performance based on indicative vowels (7 phonemes out of 39 phonemes) analysis is close to classification performance of the original phoneme-level emotion classifier which employed a complete phoneme set.