@article{ART003383723},
author={Bae Munjung},
title={Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks},
journal={The Journal of Translation Studies},
issn={1229-795X},
year={2026},
volume={27},
number={3},
pages={369-397},
doi={10.15749/jts.2026.27.3.012}
TY - JOUR
AU - Bae Munjung
TI - Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks
JO - The Journal of Translation Studies
PY - 2026
VL - 27
IS - 3
PB - The Korean Association for Translation Studies
SP - 369
EP - 397
SN - 1229-795X
AB - This study investigates error patterns in AI-based real-time speech translation by comparing AI outputs with human interpreting in a live Korean-to-English lecture setting. The analysis draws on a corpus comprising source speech (SS), automatic speech recognition transcripts (TR), machine-translated outputs (TT), and human interpreter renditions (HI). Rather than applying a predefined evaluation framework, the study inductively categorizes TT errors according to their underlying causes. The analysis identifies four major categories of errors: errors in processing speech and colloquial language, target-language inadequacy, information omission and over-generation, and discourse-context errors. Of these, discourse-context errors account for the largest proportion, closely followed by errors related to speech and colloquial language processing. This highlights the limitations of current AI speech translation systems in handling spoken-language variability and integrating multimodal and contextual cues. In contrast, human interpreting performs effectively in these areas through strategies such as compensating for disfluencies, context-sensitive reformulation, multimodal integration, and pragmatic adjustment. The findings suggest that evaluation frameworks for AI speech translation that draw on human interpreting assessment may fail to capture the distinctive error profile of AI speech translation. They therefore point to the need for frameworks that incorporate AI-specific error sources such as ASR errors (both speaker disfluency-induced and system-induced) and discourse-level misinterpretation.
KW - AI-based speech translation;error patterns;evaluation frameworks;multimodal communication;human interpreter strategies
DO - 10.15749/jts.2026.27.3.012
ER -
Bae Munjung. (2026). Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks. The Journal of Translation Studies, 27(3), 369-397.
Bae Munjung. 2026, "Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks", The Journal of Translation Studies, vol.27, no.3 pp.369-397. Available from: doi:10.15749/jts.2026.27.3.012
Bae Munjung "Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks" The Journal of Translation Studies 27.3 pp.369-397 (2026) : 369.
Bae Munjung. Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks. 2026; 27(3), 369-397. Available from: doi:10.15749/jts.2026.27.3.012
Bae Munjung. "Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks" The Journal of Translation Studies 27, no.3 (2026) : 369-397.doi: 10.15749/jts.2026.27.3.012
Bae Munjung. Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks. The Journal of Translation Studies, 27(3), 369-397. doi: 10.15749/jts.2026.27.3.012
Bae Munjung. Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks. The Journal of Translation Studies. 2026; 27(3) 369-397. doi: 10.15749/jts.2026.27.3.012
Bae Munjung. Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks. 2026; 27(3), 369-397. Available from: doi:10.15749/jts.2026.27.3.012
Bae Munjung. "Error patterns in real-time AI speech translation in an authentic lecture setting: Implications for evaluation frameworks" The Journal of Translation Studies 27, no.3 (2026) : 369-397.doi: 10.15749/jts.2026.27.3.012
