UCT CS Research Document Archive

The Impact of Accents on Automatic Recognition of South African English Speech: A Preliminary Investigation

Mbogho, Audrey and Michelle Katz (2010) The Impact of Accents on Automatic Recognition of South African English Speech: A Preliminary Investigation. In Proceedings Annual Conference of South African Institute for Computer Scientists and Information Technologists (SAICSIT) 2010, pages 187-192, Bela Bela, South Africa.

Full text available as:
PDF - Requires Adobe Acrobat Reader or other PDF viewer.

Abstract

The accent with which words are spoken can have a strong effect on the performance of a speech recognition system. In a multilingual country such as South Africa where English is not the first language of most citizens, the need to address this issue is critical when building speech-based systems. In this project we trained two sets of hidden Markov Models for isolated word English speech. The first set of models was trained with native English speakers and the second set was trained with non-native speakers from a representative sample of major South African accent groups. We compared the recognition accuracies of the two sets of models and found that the models trained with accented English performed better. This preliminary research indicates that there is merit to committing resources to the task of accented training.

EPrint Type:Conference Paper
Keywords:Speech Recognition
Subjects:I Computing Methodologies: I.5 PATTERN RECOGNITION
I Computing Methodologies: I.2 ARTIFICIAL INTELLIGENCE
ID Code:635
Deposited By:Mbogho, Audrey J. W.
Deposited On:07 November 2010