A DCT based nonlinear predictive coding for feature extraction in speech recognition systems

Speech representation strategies play a key role in automatic speech recognition systems. In this study, a nonlinear procedure has been proposed to overcome the complexities of speech sequence representations. The proposed method may be considered as an extension of nonlinear predictive coding representation procedure in cosine transform domain. The best results belong to classification of nonlinear behaved stop phonemes (i.e. /b/, /d/, /g/) in TIMIT database which show good performance while reducing the computational complexity in comparison to standard NPC.