Meta’s Google Podcast Generator Is Now Open To All

Date: October 28, 2024

Meta has released the open implementation version of the viral generate-a-podcast feature in Google’s NotebookLM.

Meta has spent time honing an AI feature that helped Google’s NotebookLM go viral. To help users simplify the absorption of complex information, Meta is releasing an open implementation model of the generate-a-podcast feature. Launched under the name NotebookLlama, it resembles Google’s NotebookLM in many features, restrictions, and potential.

Like NotebookLM, the newly launched podcast generator can back-and-forth, podcast-style digests of uploaded text files or sources you provide to the AI model. Its remarkable text-to-speech engine creates compelling and engaging audio content demonstrating the dramatized nature of human podcasts.

The process explained by Meta is simple, effective, and fast, thanks to the proprietary Large Language model, Llama. NotebookLlama first creates a PDF transcript of a news article or blog post while preserving its context using Llama 3.2-1B Instruct. Then, it feeds the interpretation to Llama 3.1-70B-Instruct, where the podcast script is generated.

To add dramatization that makes a podcast interesting, engaging, and conversational, the script draft is processed by Llama 3.2-8B-Instruct, and a crispier script is generated. Then, the AI model converts the final script into an audio podcast, orated in a conversational natural tone.

Source

The podcast generator is still in its infant stage, and that’s why the software has been released in an open version, unlike Google’s NotebookLM. While it showcases immense potential, the limitations pose a major roadblock in finding the actual use case of the product. One of the limitations is the not-so-natural conversational tone and voice used by the AI model.

“The text-to-speech model is the limitation of how natural this will sound. Also, another approach of writing the podcast would be having two agents debate the topic of interest and write the podcast outline. Right now we use a single model to write the podcast outline,” said Meta on its official NotebookLlama GitHub page.

Another limitation is the most persistent one in any AI model built so far. NotebookLlama and NotebookLM both are prone to hallucinations even if the user provides exact sources of the content to generate a podcast from. These hallucinations arise either while creating a strong context or to compensate for the lack of understanding through a new angle.

However, Meta NotebookLlama has solved one critical problem for every AI chatbot user: it can provide answers to questions they don’t even know to ask. In simple words, it takes a two-person conversational approach to explore all angles and identify important aspects in the form of a Q&A or debate. While the technology has a long way to go before becoming a flagship product, the breakthrough shows great promise for future developments.