Speaker Independent Speech Recognition Using Neural Network

In spite of the advances accomplished throughout the last few decades, automatic speech recognition (ASR) is still a challenging and difficult task when the systems are applied in the real world. Different requirements for various applications drive the researchers to explore for more effective ways in the particular application. Attempts to apply artificial neural networks (ANN) as a classification tool are proposed to increase the reliability of the system. This project studies the approach of using neural network for speaker independent isolated word recognition on small vocabularies and proposes a method to have a simple MLP as speech recognizer. Our approach is able to overcome the current limitations of MLP in the selection of input buffers’ size by proposing a method on frames selection. Linear predictive coding (LPC) has been applied to represent speech signal in frames in early stage. Features from the selected frames are used to train the multilayer perceptrons (MLP) feedforward back-propagation (FFBP) neural network during the training stage. Same routine has been applied to the speech signal during the recognition stage and the unknown test pattern will be classified to one of the nearest pattern. In short, the selected frames represent the local features of the speech signal and all of them contribute to the global similarity for the whole speech signal. The analysis, design and the PC based voice dialling system is developed using MATLAB®.

Preview PDF 1000548949_t_FK_2004_90.pdf Download (171kB)

Actions (login required)

View Item