Gender and accent identification for Malaysian English using MFCC and Gaussian mixture model

Speaker and speech variability are a challenge in speaker and speech recognition. In the context of Malaysian English speakers, the variability is highly complex due to sociolinguistic and cultural background. Past researches focused on vowel classification of Malaysian English on a small dataset wi...

Full description

Saved in:
Bibliographic Details
Main Author: Goh, Eng Lyn
Format: Thesis
Language:English
Published: 2013
Subjects:
Online Access:http://eprints.utm.my/id/eprint/35850/5/GohEngLynMFSKSM2013.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Speaker and speech variability are a challenge in speaker and speech recognition. In the context of Malaysian English speakers, the variability is highly complex due to sociolinguistic and cultural background. Past researches focused on vowel classification of Malaysian English on a small dataset with limited number of speakers and text corpus. Among speaker specific characteristics, gender is the most prominent features, which is followed by accents. This project approached the issues of speaker and speech variability in Malaysian English by proposing identifier that combined the gender and accent aspects. This method is fulfilled by training the classifier with gender-dependent data and accent-prone text corpus. The genderaccent database collected is comprised of 120 speakers categorized into four genderaccent groups namely Malay Female (CF), Malay Male (MM), Chinese Female (CF) and Chinese Male (CM). MFCC algorithm is used to extract the features, while GMM is the algorithm used to model the identifier. Findings from the test results show that female and Chinese speakers have higher degree of distinctiveness compared to other accent-gender groups. Chinese female is the best recognized accent-gender group, meanwhile Malay Male is the least recognized, due to codemixing of Malay language and English. Optimum configuration of GMM is also studied across 3 different numbers of Gaussians (12, 24, and 32). It is found that 24 is the most optimal configuration of MFCC-GMM. Meanwhile, it is also known that MFCC-GMM performs better than LPC-KNN on noisy dataset. Overall, the MFCCGMM identifier scored 99.34% gender identification rate, 67.5% accent identification and 65.83% for accent-gender identification task.