Inspiration & Overview
We want to listen to more podcasts, but don't have enough time! With PodLight, you can input a topic of your choice, and our AI engine will splice together the most salient sound bytes from podcasts about this topic. An AI emcee mediates the final product, which contains the most relevant clips from podcasts.
Technical Overview
We downloaded transcripts and audio files for several reputable, information-dense podcasts from Youtube, splitting them up by "chapters" within the podcasts. We calculated and stored the embeddings for each chapter in Pinecone. Upon receiving a user query, we perform a similarity search to find the most relevant embeddings in vector space. Our AI stitcher then retrieves the audio clips for these embeddings, and stitches them together. An intro, outro, and clip-specific transitions are created by AI, and used to pull the previously-retrieved clips into a single, coherent podcast. This podcast is then presented to the user on our webpage, where they can listen to the podcast and download it. More details about the technical aspect of our project may be found on our architecture diagram in our slides here.
Log in or sign up for Devpost to join the conversation.