Feature Selection Applied to G.729 Synthesized Speech for Automatic Speaker Recognition
Kawthar Yasmine Zergat, Sid‐Ahmed Selouani, Aissa Amrouche · 2018 IEEE 5th International Congress on Information Science and Technology (CiSt) · 2018
Performance of Automatic Speaker Recognition (ASR) over wireless communications knows a clear degradation due to the effect of the synthesized speech. In order to enhance the performance accuracy, we propose to apply a feature selection using Linear Discriminant Analysis (LDA) in the front end part of the ASR system. For simulation the G.729 codec is used. The speaker and session dependent i-vector distribution is modeled using a Gaussian probabilistic LDA (GPLDA). The obtained results, carried out by using the TIMIT corpus and MFCCs input parameters demonstrate that the LDA technique mainly dismissed the MFCC's derivatives, which enhances the recognition accuracy in both clean and synthesized speech.