What it does Feed it a research paper PDF and it spits out a ~10-minute audio discussion: a host, a curious learner, and an expert arguing over your commute. The pipeline plans, scripts, and voices the whole thing via OpenAI’s API, then hands you an MP3.

The interesting bit The three-person format isn’t just theater—it’s structural. The “planning chain” maps the paper first, then a retrieval-augmented “discussion chain” keeps the dialogue anchored to actual source content, and an “enhancement chain” scrubs redundancies. It’s basically a tiny assembly line for synthetic Socratic method.

Key highlights

• Three fixed personas (Host / Learner / Expert) with distinct voice roles
• RAG-based discussion generation to limit hallucinations
• OpenAI TTS with realistic voice differentiation per character
• ~$0.16 to convert a 19-page paper into 9 minutes of audio
• Sample outputs live in ./sample_podcasts

Caveats

• Currently OpenAI-only; local LLM/TTS support is on the roadmap but not implemented
• Generation is slow; the README explicitly flags runtime as a known pain point
• Single PDF input, no batch processing or URL fetching mentioned

Verdict Great for researchers who want to audit papers while walking the dog, or educators building accessible content. Skip it if you need real-time generation, open-source stack purity, or anything beyond the hardcoded three-host format.