Evaluation of the usability and usefulness of automatic speech recognition among users in South Africa [electronic resource]

Abstract

An automatic speech recognition (ASR) system is a software application which recognizes human speech, processes it as input, and displays a text version of the speech as output or uses the input as commands for another application’s usage. ASR can either be speaker-dependent or speakerindependent. A speaker-dependent ASR system requires every user to perform training before its usage, while speaker-independent ASR requires no prior training before usage. The technology of ASR is based on identification and comparison of sound patterns; these sound patterns are combinations of the smallest units of sound called phonemes. The phonemes constitute fragments of uttered sounds in speech and their combination gives meaningful sound patterns in languages. There exists a set of phonemes for every language group, and associated with each group is the method of pronunciation called the accent. A language group could be identified by the accent in their speech; accent is the set of pronunciation rules of a language group. Accent reflects the cultural divide of a multi cultural society with a common language such as English. Some commercially available ASR systems are designed based on the accents of the following language groups: English, French, German, Italian, Dutch, and Spanish. These language groups are European with none having any similarities with African languages and accents, (except Afrikaans and English, which, though spoken in Africa, originated from Proto-Indo-European languages). This study involved the evaluation of commercially available English ASR systems, establishing their usability and usefulness among different language groups in South Africa which use English as a common language. Of particular interest was the effect of African accents on the performance of the ASR systems. ASR technology is widely used and researched in the developed world with reported recognition accuracy of up to 99%. However, English spoken with African accents may have adverse effect on the recognition accuracy. Despite the fact that most existing ASR systems are not designed for English spoken with South Africans’ accents, one can easily purchase them over the shelf in South Africa. The systems used in this study are: 1. Nuance Dragon NaturallySpeaking, Version10.0 (NDNS). 2. Windows Speech Recognition, Windows Vista version (WSR). The result of this study indicated that accent has influence on the ASR recognition accuracy. It also indicated that users’ satisfaction was greatly affected by the recognition accuracy obtained. The results also indicated poor performance in environments where speech cannot be loud, for example, in the library.