Chapters are an increasingly popular feature for podcasts and other media, as it can help users to navigate, skim, and share the content. However, chapters can take a long time to produce as they require someone to decide how to break the content into chunks, and to write titles and descriptions for each chapter. This is especially challenging if that person is not familiar with the content, or if you are having to go through a back catalogue.
With the advent of large language models, we now have access to tools that can help with some of these tasks, such as writing titles and descriptions to summarise part of a transcript. As part of the AI4ME collaborative project, we designed and built a prototype system for efficiently chapterising media with the help of AI. Our aim was not to automate the chapterisation process, but to create a tool that can assist a human to do it very quickly. This is because LLMs can often suffer from hallucinations or misunderstand the context of what is said.
We worked with the BBC’s AI research team who had trained a machine learning model on thousands of radio programmes that had been manually chapterised by hand. This system could estimate the position of chapter points in any similar style of programme. By segmenting the transcript and running it through a LLM, we could then generate proposals for various titles and descriptions that a producer could accept, reject, or edit. We designed a simple web interface to make it quick and easy for producers to preview the recommendations for their programme, and to make adjustments to the chapter points, titles, and descriptions before publishing.