Robustness Analysis of Feature Extractors for Ethnic Identification of Malaysian English Accents Database

Abstract

Accent is fascinating human speech behavior that can be used to mark personal identity and social characteristics of its bearer. However, it also has potential to bias social interaction which includes prestige, job competency and ethnic discrimination. Albeit many successful methods have been deployed in the past to identify a speaker accent, the success rates are most likely database-dependent. This study aims to inquire about identification of Malaysian English (MalE) accents caused by ethnic diversities in this country. Robustness analysis was conducted using seven noisiness levels by corrupting the speech signals with additive white Gaussian noise (AWGN) to investigate the performance of four different schemes of feature extractors under clean and noisy conditions. These methods are filter bank analysis consists of mel-frequency cepstral coefficients (MFCC) and a new set of formulated features named as descriptors of mel-bands spectral energy (MBSE). Principle component analysis (PCA) was utilized to transform to another new features called PCA-MBSE. Second, vocal tract analysis consists of linear prediction coefficients (LPC) and formant frequencies (formants). Third, hybrid analysis consists of discrete wavelet transform (DWT) and LPC. The last scheme is fusions of spectral features (SFFs) of MFCC with formants and LPC with formants. Experimental results showed that SFFs techniques possess more sturdy noise resistivity than MFCC, LPC, MBSE, and DWT-derived LPC features. Similarly, PCA-transformed MBSE was just moderately affected as compared to the original features. While PCA-MBSE only caused a performance drop of 15 % in average and the SFFs were just slightly affected by the AWGN from 8 to 13 % drop, the percentage drop of other feature sets were fairly above 30 %.