GenAI for speech writing in the training of Maltese conference interpreters for the European Union

Abstract

The present paper explores the potential of GenAI tools in generating speeches to prepare for the European Union’s interinstitutional accreditation test. A small-scale experimental empirical study was conducted in which interpreting students were instructed to annotate, critically assess and compare English and Maltese speeches generated by three GenAI tools, viz., Gemini, Copilot and ChatGPT, to be used for beginner consecutive interpretation practice. The GenAI tools were prompted to generate three English and three Maltese speeches modelled on those in the European Commission’s Speech Repository. The analysis focuses on compliance with the prompt, suitability for purpose and linguistic output quality. The results indicate that, upon initial analysis, the speeches in both languages satisfy many of the criteria in the prompt. However, more thorough scrutiny reveals that the speeches may prove challenging for trainees to interpret, primarily due to their poor argumentative structure, low factual density, lack of clear links and intent, and low terminological complexity. In addition, the speech topics are excessively simplistic, not well-researched and insufficiently nuanced. The differences between English, a high-resource language, and Maltese, a low-resource language, are minimal. The main discrepancy between the two is the higher number of linguistic errors in Maltese. Overall, the results indicate that the speeches in both languages require extensive post-editing to meet their intended use.