Meta recently released an "open" version of Google’s NotebookLM podcast generator, calling it NotebookLlama. Built on Meta’s own Llama models, NotebookLlama mimics NotebookLM by turning uploaded text files into conversational, podcast-style summaries.
Here’s how it works: NotebookLlama first transcribes the text — like a PDF of an article or blog post. It then dramatizes the dialogue a bit, adding in pauses and interruptions before using open-source text-to-speech software to create the audio. However, the output is noticeably robotic; the voices occasionally overlap awkwardly, falling short of NotebookLM’s smoother delivery.
Meta researchers acknowledge the limitations, pointing out that stronger text-to-speech models would improve the quality. They also suggest that instead of having a single model script the podcast, future versions could involve two AI "speakers" to debate and craft a more dynamic outline.
NotebookLlama joins several other attempts to replicate NotebookLM’s podcasting feature, each facing similar challenges. Even NotebookLM hasn't resolved the persistent issue of "hallucinations" in AI-generated content, meaning some inaccuracies are bound to slip in.
