A Multimodal Narrative Analysis Framework for University Ceremony Live Streaming Based on Deep Vision and Speech Models

This study presents a multimodal narrative analysis framework for live-streaming scenarios of university ceremonies. The system has three core components: a visual analysis module, a speech-language processing module, and an acoustic-emotion modeling module. YOLOv8 is used for shot-scale classification in the visual component, while scene boundary detection (SBD) identifies shot transitions. In the language component, speech is transcribed using the Whisper model and then processed with TF-IDF and LDA for keyword extraction and topic clustering. The acoustic component focuses on pitch contour tracking and sentiment orientation estimation to characterize emotional fluctuations. By aligning outputs from the three modalities on a unified temporal axis, the framework enables the examination of narrative rhythm, semantic emphasis, and emotional progression within the event. This approach contributes to the structured analysis of institutional livestreams and may inform future research in educational media understanding and event-based multimodal content processing.