A LIGHTWEIGHT DOWNSCALED APPROACH TO AUTOMATIC SPEECH RECOGNITION FOR SMALL INDIGENOUS LANGUAGES

Abstract

Development of fully featured Automatic Speech Recognition (ASR) systems for a complete language vocabulary generally requires large data repositories, massive computing power, and a stable digital network infrastructure. These conditions are not met in the case of many indigenous languages. Based on our research for over a decade in West Africa, we present a lightweight and downscaled approach to AI-based ASR and describe a set of associated experiments. The aim is to produce a variety of limited-vocabulary ASRs as a basis for the development of practically useful (mobile and radio) voice-based information services that fit needs, preferences and knowledge of local rural communities.