Significance of Corpus Quality for Direct Speech-to-Text Translation Systems

Abstract

Performance improvement in Direct Speech-to- Text Translation systems are mainly attributed to its training using a larger corpora which led to the development of many larger corpora containing data scrapped from various online sources, This resulted in quality issues and impracticality in the long run. Hence, this work investigates the role of quality in the corpora to determine whether size or quality has more contribution in the performance of these systems. Experimental results indicate that a corpus containing a richer vocabulary with better translation and audio quality is more effective and has a greater contribution in the performance. © 2024 IEEE.