Challenges of Idiom Translation in Speech-to-text Systems

This research article examines the challenges associated with idiom translation in speech-to-text (SLT) systems, specifically in comparison to conventional machine translation (MT) systems. The study evaluates the performance of various systems, including end-to-end SLT architectures and traditional MT methods, across two language pairs: German to English and Russian to English. It highlights that SLT systems often struggle with idiomatic data, resulting in frequent literal translations that fail to convey the intended meaning. In contrast, MT systems and large language models (LLMs) show superior handling of idioms. The research also discusses the necessity for idiom-specific strategies and improved representations within SLT architectures. The findings reveal a notable performance discrepancy between SLT and MT systems when dealing with idioms, emphasizing the ongoing need for advancements in speech translation methodologies. The article also outlines the evaluation methodology and datasets used to assess the translation capabilities of the systems being studied.