Hi HackerNews! I wrote an ML pipeline to create a customized podcast that summarizes other podcasts. It uses the latest and greatest in transcription (whisper), summarization (gpt3), and narration (tortoise) models.
Of these models, I am most impressed with Tortoise. It's a near-perfect audio deepfake.
I'm playing with making this into an app/service where you can select your favorite podcasts and get an RSS feed. I'm not sure how to make that sustainable since the compute requirements are hefty. I am very open to feedback!