Cascade of Ordinal Classification and Local Regression for Audio-Based Affect Estimation

摘要

Affective dimensions (i.e. valence, arousal, etc.) are continuous, real variables, bounded on [-1,+1]. They give insights on people emotional state. Literature showed that regressing these variables is a complex problem due to their variability. We propose here a two-step process. First, an ensemble of ordinal classifiers predicts the optimal range within [-1,+1] and a discrete estimate of the variable. Then, a regressor is trained locally on this range and its neighbors and provides a finer continuous estimate. Experiments on audio data from AVEC'2014 and AV+EC'2015 challenges show that this cascading process can be compared favorably with state of art and challengers results.