Acoustic Comparison of Malaysian and Nigerian English Accents

Abstract

This study examines the differences in spectral and cepstral acoustics features between Malaysian and Nigerian English accents with the aim of determining the effect of accents on spectral and cepstral features of speech. Accent has received a great attention from ARS researchers due to the fact that it is a major source of ASR performance degradation. Most ASR applications were developed with native English speakers speech samples disregarding the fact that majority of its potential users speaks English as a second language with a marked accent. Malaysia and Nigeria were both colonized by Britain and speaks English as an official or second language despite being multi-ethnic nations. The results of the study revealed that formants values can be used to differentiate between ME and NE accents, most especially F1 and F2. Cepstral (MFCC) performs better in accents recognition than formants features. While the combination of both formants and MFCC yields a better classification performance. However, the effect of the formants is non-uniform and depends on the vowels and accents under consideration. This is evident as each of the formants has different predictive values. Classification rate shows that Multi-Layer Perceptron (MLP) performs better than K-nearest neighbors (KNN).