Cover Image for Meta launches an 'open' version of Google's podcast generator.
Mon Oct 28 2024

Meta launches an 'open' version of Google's podcast generator.

Meta has launched an 'open' version of the popular podcast generation feature found in Google’s NotebookLM, called NotebookLlama.

Meta has launched an "open" implementation of the popular podcast generation feature from NotebookLM, known as NotebookLlama. This new project uses Meta's own Llama models to handle much of the processing. Like NotebookLM, NotebookLlama has the ability to create podcast summaries from text files that are provided to it. First, the system generates a transcription from a file—such as a PDF of an informative article or a blog post. It then adds "more dramatization" and interruptions before sending it to open-source text-to-speech models.

Despite its functionality, the results do not sound as polished as those from NotebookLM. In the samples of NotebookLlama that have been analyzed, the voices have a notably robotic quality and, at times, tend to interrupt each other at unusual moments. However, Meta researchers acknowledge that this quality can be improved with more advanced models. On the NotebookLlama GitHub page, they remark: “The text-to-speech model limits how natural this will sound. [Additionally,] another way to structure the podcast could be to have two agents debate the topic of interest and draft the podcast outline. Currently, we use a single model to write the podcast outline.”

NotebookLlama is not the first attempt to replicate the podcast feature of NotebookLM. Some projects have been more successful than others, but none—even NotebookLM—have been able to solve the hallucination problem that affects all artificial intelligences. This means that AI-generated podcasts are likely to contain fabricated information.