Generalized Discriminant Methods for Improved X-Vector Back-end Based Stress Speech Recognition
Barlian Henryranu Prasetio, Hiroki Tamura, Koichi Tanno · IEEJ Transactions on Electronics Information and Systems · 2019
In this paper, we discuss a system that can recognize and classify stress based on speech. The system was built using a deep neural network (DNN) embedding to improve the baseline system (i-vector), called x-vector. The i-vector system usually requires Linear Discriminant Analysis (LDA) for dimension reduction, followed by Probability Linear Discriminant Analysis (PLDA) for scoring. In this work, the Generalized Discriminant Analysis (GDA) based Gaussianized cosine kernel and Joint-PLDA was used to replace the LDA and PLDA, respectively. The SUSAS database was used for training, testing, and enrollment data of our proposed system. We assessed our proposed system's effectiveness compared to the baseline system using equal error rate (EER). The evaluation and experiment results show that our proposed system outperforms the baseline system.