Personal name pronunciation for South African automatic speech recognition: The effect of speaker primary language.
Kgampe, Betty Mongalo Mpho
Kgampe, Betty Mongalo Mpho
Citations
Altmetric:
Abstract
Proper names are often pronounced in very different ways across language boundaries. This study investigates the consistency with which speakers of the different primary languages– Setswana, English, Afrikaans and isiZulu- are able to pronounce personal names. We sought to determine whether some of these pronunciation errors are systematic and if so, in which ways. In particular, we analyse some of the typical errors made by speakers when pronouncing the same names in the four languages. We gather data in a controlled research study and analysed cross-lingual pronunciation effects. We find that speakers with a similar primary language tend to agree on the correct pronunciation of a name originating from their own language community, and that the ability of speakers from other language communities to approximate this pronunciation is highly dependent on the speaker-word language pair. We also find that there are systematic ways in which names are mispronounced by different language communities. Understanding such systematic ways is of consequence when extending electronic pronunciation dictionaries (used in spoken dialogue systems) to the most significant variants that occur in practice, in order to increase the accuracy of name recognition. The study then compares the human respondents’ results with the pronunciations generated by the four language-specific grapheme-to-phoneme (G2P) predictors which focused on generic words from those four languages. We find that the G2P predictors are able to predict at least some of the typical errors humans make and, that these errors are more predictable than the correct pronunciations themselves. The results found in the analysis could form the basis for the development of a G2P/P2P approach that is applicable to the prediction of South African personal names. In order to develop such a system, a significantly larger data set would be required.
Description
Submitted in fulfilment of the requirements for the Magister Technologiae: Language practice in the Department of Applied Languages.
Date
2013-03-05
Journal Title
Journal ISSN
Volume Title
Publisher
Tshwane University of Technology
Research Projects
Organizational Units
Journal Issue
Keywords
Pronunciation, Automatic Speech Recognition (ASR), Personal Names, Primary Language, South Africa