Resumen
Automatic speech recognition (ASR) has recorded enormous development in both research and implementation such as voice commands to control electronic appliances, video games, interface to voice dictation, assistive leaving for the elderly, and dialogue systems. Rapid development on ASR can be seen on the English language, while duplicating the ASR framework for Malay language is possible, but the work demands endlessly efforts. One of common tools that is able to classify Malay vowels is Multinomial Logistic Regression (MLR). However, careless on estimating the parameters of MLR may lead to producing biased classifier which inappropriate for future classification. Besides, the used on huge number of features for classification sometimes hinder MLR to perform well. This paper outlines a new idea for estimating the unknown MLR parameters with less number of features using a double-stage features extraction based on MLR (DSFE-MLR). The proposed DSFE-MLR extracted 39-MFCC from speech waveform and constructed an MLR using training set. Next, the MLR output of class membership probabilities were further extracted through MLR and evaluated using test set. Empirical evidence on Malay sample of students shows that the DSFE-MLR recorded the highest accuracy compared to other classifiers. Besides, the method is able to recognize each of five Malay vowels correctly. In general, DSFE-MLR provides an increment of accuracy for Malay speech recognition.