Методи оцінки якості синхронізації аудіо та відео на основі нейронних мереж

Abstract

This thesis analyses the impact of audiovisual synchronisation on the realism of digital avatars used in film dubbing, video conferencing and digital assistants. Accurate lip-audio alignment is highlighted as a key factor in achieving natural, Comparative attention is given to specialised lip-sync models (e.g. SyncNet) and advanced multimodal approaches (e.g. AV-HuBERT). The results show that stable synchronisation significantly improves audience engagement and reduces visual inconsistencies. Practical applications in multilingual face-dubbing illustrate the potential and limitations of current methods.