Meta has released an “open” implementation of the viral podcast generation feature in Google’s NotebookLM.
Called NotebookLamaNot surprisingly, the project uses Meta's own Llama models for much of the processing. Like NotebookLM, it could actually generate back-and-forth, podcast-like digests of the text files uploaded to it.
NotebookLlama first creates a transcript from a file – e.g. B. a PDF of a news article or blog post. “More dramatization” and interruptions are then added before the transcript is fed into open text-to-speech models.
The results don't sound nearly pretty much as good as NotebookLM. In the notebookLama Samples I listened to it, the voices are obviously robotic in nature and are likely to talk over one another in strange places.
But the meta-researchers behind the project say the standard could possibly be improved with stronger models.
“The text-to-speech model limits the naturalness of sound,” they wrote on NotebookLlama’s GitHub page. “(Also) one other approach to writing the podcast can be to have two agents discuss the subject of interest and write the podcast outline. Right now we use a single model to put in writing the podcast outline.”
NotebookLlama shouldn’t be the primary attempt to copy NotebookLM's podcast feature. Some projects were more successful than others. But nobody – not even NotebookLM itself – has managed to unravel the hallucination problem that plagues all AI. That means AI-generated podcasts are It's definitely going to contain some made up stuff.
