SEMI-SUPERVISED CLASSIFICATION OF SPEAKER'S PSYCHOLOGICAL STRESS

摘要

It is well known that speech signal is affected by speaker's stress. Some of the recent works have evaluated different acoustic features individually for the detection of stress from speech. In our previous work, a new mixed feature (TEO-Pch-LFPC) was proposed for this purpose. Here, this feature is evaluated for the task of stress classification using simulated domain of SUSAS database. Although, we have used more simple classifiers than HMM, and the Round Robin Method is exerted, the classification accuracy rates are improved. Also, we present a semi-supervised approach which can efficiently employ unlabeled data in the structure of supervised classifiers. Experiments using this method result in greater classification rates with the same labeled data set.