Atanda, Abdulwahab F. and Mohd Yusof, Shahrul Azmi and Husni, Husniza (2017) Accent identification of Malaysian and Nigerian English based on acoustic features. In: Sintok International Conference on Social Science and Management (SICONSEM 2017), 4-5 December 2017, Adya Hotel, Langkawi Island, Kedah, Malaysia.

PDF Restricted to Registered users only Download (367kB) | Request a copy

Abstract

Purpose - This paper studies acoustics features of energy, pitch and formants of Malaysian and Nigerian English vowels with the aim of effective accents identification using multi liner regression (MLR) and linear discriminant analysis (LDA) classifiers for performance improvement of ASR when exposed to accented speech.Accent being a foremost source of ASR performance degradation has received a great attention from ASR researchers.Majority of ASR applications were developed with native English speakers speech samples without considering fact that most of its potential users speaks English as a second language with a marked accent, hence its poor performance when exposed to accented speech. Previous studies on accent has shown that the ability to correctly recognized accent has greatly enhanced the recognition performance of ASR when exposed to accented speech data.In a study of 14 regional accents of British, (Hanani, Russell, & Carey, 2013) achieved a performance increase of 5.58%.A study by (Vergyri, Lamel, & Gauvain, 2010) using six different regional accented English shows an average of 41.43% WER.Which was reduced to 27% on the incorporation of accent identification module.Several studies have explored several acoustic features of speech such as energy, pitch, formants, MFCC, and LPC to establish the differences between regional or cross ethnics accent aimed at better understanding of the differences in the acoustic features to enhance ASR performance.Apparently from the previous studies reviewed above, it is evident that accent constitute a hurdle to the performance of ASR. Hence, consequently serves as a barrier to ASR wide reception and usage in real life situations. Consequently, it becomes pertinent that accent should be given adequate research attention with the view of enhancing ASR performance to accented speech which will inherently promotes its wide acceptability and applicability globally.

Item Type:	Conference or Workshop Item (Paper)
Additional Information:	eISBN 978-967-2064-65-7 Organized by: UUM PRESS Unversiti Utara Malaysia
Uncontrolled Keywords:	Automatic speech recognition; Accent recognition; Multi Liner Regression (MLR) ; Linear Discriminant Analysis (LDA); Formants analysis.
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions:	School of Computing
Depositing User:	Mrs. Norazmilah Yaakub
Date Deposited:	30 Jul 2018 01:09
Last Modified:	30 Jul 2018 01:09
URI:	https://repo.uum.edu.my/id/eprint/24494

Actions (login required)

View Item