The purpose of this thesis is to describe a biologically motivated approach for phoneme recognition by using a self-organized neural network and sequence learning algorithm. Phoneme recognition in continuous speech is a tough task with a low accuracy rate. By using the sequence learning algorithm to add sequential information of individual phonemes, recognition performance can be improved.
This thesis includes three parts. A self-organized neural network is the first stage, which classifies the input sound waves into forty two different phoneme categories. The 42 output neurons of the neural network are sent to the Sequence Learning block which is composed of Long Term Memory cells. Finally each LTM cell sends a unique feedback strength signal to each output of the neural network to predict the next phoneme, hence, to improve the phoneme recognition based on the sequential information.