Assessing the Effectiveness of Voice-Translation Tools for Translating Medical Information from English to Arabic

Abstract

In this study, the performance of Gemini and DeepSeek is evaluated in translating spoken medical information from English into Arabic. Specifically, special attention is paid to medical and linguistic aspects of translation. A descriptive, analytical, and comparative approach is adopted in examining a corpus of 40 spoken medical excerpts, which was collected from public sources of medical information on YouTube. In general, the corpus covers various areas of medicine including diseases and symptoms, medicines and doses, medical tests, warnings, and instructions. Each spoken medical excerpt was translated into Arabic independently by Gemini and DeepSeek, and the resulting translation outputs were assessed with the help of the MQM approach paying special attention to such criteria as Accuracy and Fluency. Errors in medical translation were classified into several groups, namely mistranslation, omission, addition, and linguistic errors. In order to complement error analysis, ten representative excerpts were chosen and carefully analyzed qualitatively. Results showed that both systems produce clear, fluent, and natural Arabic translations; nevertheless, the systems differ in the ability to preserve medical terms, clinical information, degree of probability, and urgency. It should be emphasized that the analysis demonstrated that fluency and medical accuracy in translation do not coincide as fluent and natural translations may still be semantically different from the source information. Therefore, it is crucial to make the distinction between linguistic and semantic characteristics of translations produced by AI. The current study proves that Gemini and DeepSeek have great potential for becoming valuable tools for supporting English-Arabic medical translation especially concerning spoken medical information. However, their employment in the sphere of medicine is possible only with the specialized human evaluation.