Phoneme Classification Using New Feature Extraction Techniques based on Mellin Transform

This paper presents a new hierarchical phoneme recognition system using the SVM classifier and different feature representations based on mellin transform. The proposed architecture uses different representations with each group of phonemes of the speech database TIMIT which are distributed in a way to reduce the confusions between phonemes having similar articulatory strcuture. The main idea of this new architecture is based on the principle that each group of phonemes has his own characteristics which requires us to choose the adequate representation for the given group. Experiments have proven the robustness of our new hierarchical phoneme recognition system (called MMP) and the use of conventional feature representations based on mellin transform explains its superior recognition performance.