Acoustic group feature selection using wrapper method for automatic eating condition recognition

In this paper, we present a wrapper-based acoustic group feature selection system for the INTERSPEECH 2015 Computational Paralinguistics Challenge (ComParE) 2015, Eating Condition (EC) Sub-challenge. The wrapper-based method has two components: the feature subset evaluation and the feature space search. The feature subset evaluation is performed using Support Vector Machine (SVM) classifiers. The wrapper method combined with complex algorithms such as SVM is computationally intensive. To address this, the feature space search uses Best Incremental Ranked Subset (BIRS), a fast and efficient algorithm. Moreover, we investigate considering the feature space in meaningful groups rather than individually. The acoustic feature space is partitioned into groups with each group representing a Low Level Descriptor (LLD). This partitioning reduces the time complexity of the search algorithm and makes the problem more tractable while attempting to gain insight into the relevant acoustic feature groups. Our wrapper-based system achieves improvement over the challenge baseline on the EC Sub-challenge test set using a variant of BIRS algorithm and LLD groups.