Random discriminative projection-based feature selection for computational paralinguistics

Computational paralinguistics deals with the underlying meaning of the verbal messages. Understanding the meaning of verbal messages provides interpreting spoken content and behaving accordingly like humans. It allows us to develop human like machines. Hence, paralinguistic area is attracting increasing attention for research. Paralinguistic analysis involves extracting features from raw speech data, chunking, selecting relevant features and training the model. In this thesis, the focus is on the feature selection step. Feature selection aims at finding a relevant and necessary set of features to train generalizable models. The main challenge for feature selection methods is the greedy-search nature of them. One major motivation for this study to develop an efficient feature selection technique is the success of a recently developed discriminative projection based feature selection method. Here, the method is enhanced by applying the power of stochasticity to overcome traps in local minimum while reducing the computational complexity. The proposed approach assigns weights both to groups and to features individually in many randomly selected contexts and then combines them for a final ranking. The efficacy of the proposed method is shown in two recent challenge corpora to detect level of depression severity and conflict.