Meta has introduced an “open” version of its generate-a-podcast feature, called NotebookLlama, which utilizes the company’s Llama models. This tool generates podcast-style digests from uploaded text files by creating a transcript and adding dramatic elements before using open text-to-speech models.
Initial results indicate that the generated podcasts have a robotic tone and may contain awkward overlaps between voices. However, researchers acknowledge that model quality can be improved to enhance natural speech. They suggest strengthening the text-to-speech model and exploring alternative approaches, such as having two agents debate topics for a podcast outline.
NotebookLlama is not the first attempt to replicate NotebookLM’s podcast feature but faces challenges like hallucination, which leads to inaccuracies in AI-generated podcasts.
Source: https://techcrunch.com/2024/10/27/meta-releases-an-open-version-of-googles-podcast-generator/