Automating VOD Chapter Generation with AWS Elemental MediaConvert and Amazon Bedrock

This solution extracts shot transition candidates using the Per-frame metric reports feature of AWS Elemental MediaConvert, and Amazon Bedrock analyzes them alongside the audio script to generate chapters. The core concept is obtaining scene transition information during the existing transcoding process without requiring separate video analysis tools.

By including the timecoded script obtained from Amazon Transcribe and the shot transition timestamps from MediaConvert in the prompt, the AI is restricted to selecting chapter start points only from the candidate list. The AI output is validated by code to ensure alignment with shot transition points, and in case of failure, evenly divided default chapters are generated to ensure pipeline stability.

Additionally, it leverages MediaConvert's image-based trick play feature to generate representative thumbnails and preview tiles for each chapter in a single pass. It uses actual interval values from the HLS manifest to prevent thumbnail positioning errors, and automates the entire pipeline with a Step Functions workflow.