AI-Powered Voice Translation in Medical Settings: An Evaluation of English–Arabic Translation Accuracy

Abstract

In the current paper, the translation of spoken medical information from English into Arabic using Gemini and DeepSeek is studied. Specifically, the linguistic and medical aspects of translation are analyzed in detail. A descriptive, analytical, and comparative approach is used to analyze 40 spoken medical excerpts retrieved from publicly available medical information on YouTube. The medical information under analysis includes various medical areas such as diseases and symptoms, medications and doses, medical tests, warnings, and instructions. Both Gemini and DeepSeek translate the selected spoken medical excerpts into Arabic, and the translations were assessed with the use of the MQM framework, paying special attention to the accuracy and fluency. In case of errors, they have been classified into mistranslation, omission, addition, and linguistic errors. In order to provide additional analysis, ten excerpts have been chosen for the qualitative examination. The results of the analysis showed that both models provided fluent and natural in Arabic translations, but each model translates medical information in different ways, namely, the medical terminology, information, probabilistic nature, and urgency of situations are preserved differently. Moreover, it has been proven that fluency and medical accuracy cannot be considered sufficient as the fluent translation may differ in semantics from the source information. Thus, there should be a distinction between the linguistic and semantic qualities of AI translation. It has been proved that the Gemini and DeepSeek models have a great potential as the tools for English–Arabic medical translation, especially of the spoken information.