VOICE-BASED SADNESS AND ANGER RECOGNITION WITH CROSS-CORPORA EVALUATION

摘要

Real-life scenarios often require detection of few target emotional categories under a high mismatch between training and operation conditions. We present results of a study on sadness and anger detection with cross-corpora evaluations using two publically available databases. We demonstrate the influence of the mismatch on the detection accuracy comparing cross-corpora results to a single test corpus cross-validation results. We introduce the methodology of representing the broad complementary category by a number of hidden classes. We show performance improvements in sadness and anger detection by using the hidden-classes approach in both cross-corpora and single-corpus evaluations. We explore feature subset selection achieving further improvement in the cross-corpora settings.