Malay Speaker Dependent Digits Recognition with Improved Backpropagation
Authors
Ummu Salamah Mohamad
Faculty of Computer Science and Information Technology
Universiti Putra Malaysia
Ramlan Mahmod
Faculty of Computer Science and Information Technology
Universiti Putra Malaysia
Siti Mariyam Shamsuddin
Faculty of Computer Science and Information System
Universiti Teknologi Malaysia
Abstract
This paper presents a study of a Malay speaker dependent recognition using improved Neural Network (NN). The performances are evaluated for recognition of the isolated Malay digits of "0" through "9". The Error Backpropagation (BP) and an improved error signal of the BP are used in this study. Experiments are carried out by comparing the recognition rates and convergence time of the standard BP and improved BP, as well as the effects of normalization techniques on Malay speaker dependent data. The utterances are represented using the Linear Prediction Coding (LPC) method. The results show that the improved BP outperforms the standard BP in terms of its convergence with better recognition rates for unnormalized data. For the effects of normalization data, the unit simple method gives better result compared to unit range and unit variance with improved BP gives faster convergence and higher recognition rates.
