Open-Source Podcast Generation from PDFs/URLs using Llama 3

Methodology: An open-source application replicates Google's NotebookLM functionality by extracting content from a source PDF or URL and using Meta's Llama 3.3-70B to generate a podcast script with two hosts in lively discussion, based on a prompt crafted by Gabriel Chua.
Implementation Details: The generated script is then converted to speech using Kokoro-82M, with Llama 3.3-70B running at 1,000 tokens/second on Cerebras Systems inference. Audio generation is performed in streaming mode by Kokoro, running on HF's H200s without GPUs.
Performance: The system achieves near-instant podcast generation due to the speed of Llama 3.3-70B and the real-time audio generation capabilities of Kokoro-82M.
Cost Efficiency: The solution rivals the quality of closed-source solutions at close to no cost, leveraging open-source models and free GPU resources.
Additional Insight: According to additional sources, a separate Hugging Face Space called "Open Notebook LM" exists for querying Jupyter notebooks, using a likely large language model to answer questions about notebook content after users upload `.ipynb` files. The specific language model used in this separate application is not identified.
Reactions: User feedback indicates that the results are very good and crazy fast.