A Comparative Study of Human and Machine: A Critical Appraisal of Simultaneous Interpreting Output in Human and AI-Driven Performance by using deep learning algorithms

الملخص

This study presents a quantitative descriptive analysis comparing the performance of human simultaneous interpreters (H-SI) and AI-driven interpreting systems (AI-SI) using deep learning architectures specifically Transformer-based models (e.g., mBART, Wav2Vec 2.0 + NMT pipelines) in real-time multilingual contexts. Drawing on a corpus of 1,248 interpreted segments extracted from TED Talks (English–Spanish, English–Mandarin), we evaluate output quality across six empirically derived metrics: lexical accuracy, syntactic fluency, temporal alignment, semantic fidelity, discourse cohesion, and error density. Data were collected from 30 professional human interpreters and three state-of-the-art AI systems (Google Translate Live, DeepL Pro, and a fine-tuned mBART-50 model). Statistical analyses (descriptive statistics, Mann-Whitney U tests, and Kruskal-Wallis H tests) reveal that while AI-SI outperforms H-SI in lexical accuracy (M = 92.4%, SD = 3.1) and temporal alignment (M = 0.87s lag, SD = 0.21), human interpreters demonstrate significantly superior performance in semantic fidelity (M = 86.7% as well as. 71.3%, p < .001) and discourse cohesion (M = 84.1% as well as. 63.9%, p < .001). Error density was significantly lower in human output (Mean Rank = 47.3 as well as. 29.1, p = .002). Findings suggest that current AI systems, despite algorithmic advancements, remain deficient in pragmatic and contextual adaptation core competencies rooted in human cognitive-linguistic processing. The study contributes a validated metric framework for evaluating AI interpreting performance and calls for hybrid human-AI paradigms in professional settings. Implications for interpreter training curricula and machine translation pedagogy are discussed.