In this thesis I designed and built a desktop application that automatically turns
raw footage of hunting into an edited, narrated, and pedagogically annotated educational
video. The application is built on a multi-agent architecture. Four agents, a Director, a
Narrator, a Critic, and an Editor, each a separate call to the Gemini large multimodal
language model, cooperate in a sequential pipeline: automatically trimming redundant
footage, generating time-aligned expert narration in Slovenian, converting it to speech
with Edge TTS, and assembling the final video with FFmpeg. At the first mention of each
cut of meat, the video pauses on a freeze-frame instructional board, later complemented
with an anatomical diagram. The thesis describes the tools and technologies used, the system design, and implementation details of the individual modules, from windowed narration generation for longer videos to synchronizing synthesized speech with the video. I also present the working application and its user interface. I tested the system on real, unprocessed footage covering several game species and several types of hunting content, confirming the extensibility of the underlying design.
|
