Audio description and machine learning: accessibility innovations in audiovisual translation
Аннотация
The purpose of this study was to evaluate the effectiveness of modern machine learning architectures for the automatic generation of audio descriptions in the field of audiovisual translation. The study was theoretical and applied in nature and was conducted in January–August 2025. The study used the You Described, We Archived (YuWA) corpus of more than 1.2 million segments (2013–2022) and a control subsample of 200 segments according to European Broadcasting Union standards. The corpus analysis revealed syntactic deviations: attribute omissions (35–42%), term reduction (40%), and simplification of structures (45%). The collocation analysis revealed stable pairs characterising qualitative shifts in the generated descriptions, including ‘gaze – intense’ (8.2 × 10 3 ), ‘hands – clench’ (6.7 × 10 3 ), ‘panel – holographic’ (5.4 × 10 3 ), and ‘hero – stops’ (5.9 × 10 3 ). Automatic evaluation by BLEU, METEOR, ROUGE, CIDEr, SPICE, and BERTScore metrics showed the following dynamics: the classical Convolutional Neural Network and Recurrent Neural Network (CNN-RNN) reached BLEU 0.28–0.42 and SPICE 0.18–0. 24; the specialised Cinema Audio Description (CineAD) featured increased METEOR (up to 0.44) and moderate SPICE (up to 0.33); the modern Large Language Model for Audio Description (LLM-AD) provided the highest results – BLEU up to 0.53, ROUGE up to 0.52, CIDEr up to 0.67, SPICE up to 0.41, and BERTScore up to 0.71. Mean Opinion Score (MOS) confirmed this hierarchy: CNN-RNN – 2.8–3.4 points, CineAD – 3.6–3.9, LLM-AD – 4.1–4.3. The technical parameters were characterised by an average description length of about 17 words with occasional exceedances of the 25-word standard (≈ 1.1–3.2%), and the accuracy of hitting the ‘silence windows’ was 88–93% for LLM-AD, 77–84% for CineAD, and 61–72% for CNN-RNN. The innovation of this study lies in the combination of corpus, machine, and acoustic analysis, while the practical significance lies in the possibility of improving automated translation systems and inclusive content standards in Ukraine and the EU.