Speaker-Independent Malay Syllable Recognition Using Singular And Modular Neural Networks

Abstract

The paper investigates the use of Singular and Modular Neural Networks in classifying the Malay syllable sounds in a speaker-independent manner. The syllable sounds are initialized with plosives and followed by vowels. The speech tokens are sampled at 16 kHz with 16-bit resolution. Linear Predictive Coding (LPC) is used to extract the speech features. The Neural Networks utilize standard three-layer Multi-Layer Perceptron (MLP) as the speech sound classifier. The MLPs are trained with stochastic Back-Propagation (BP). The weights of the networks are updated after presentation of each training token and the sequence of the epoch is randomized after every epoch. The speech training and test tokens are obtained from 25 (17 females and 8 males) and 4 (all females) Malay adult speakers respectively. The total training and test token number are 1600 and 320 respectively. The result shows that modular neural networks outperform singular neural network with a recognition rate of about 92%.