UNESCO PDF-to-Podcast is an open-source project that turns scientific PDFs into podcast-style conversations. It extracts structured content from a PDF, creates a conversational script, and synthesizes the audio. Scripts can be tailored for experts, the general public, young people, or novices, and generated in English, French, Chinese, Spanish, Arabic, or Russian. English and Chinese have included native voices; the other four languages use English voices unless users supply their own samples. Users can choose a target duration and select two to four speakers; the defaults are 10 minutes and two speakers. The documented pipeline uses Docling for PDF extraction, Ollama with Granite4 for script generation, and VibeVoice for audio synthesis. It runs from the command line and supports Linux, macOS, Windows, and self-hosted use. Setup requires Python 3.9 or later, Poetry, and Ollama. A GPU is recommended, while CPU use is also possible. It handles one PDF at a time, and long documents may need to be split into chunks. Audio generation can require substantial GPU memory, especially with the optional 7B model. Audio export is WAV, and the project supports voice cloning.
Who it is for
This project suits readers who want to turn scientific PDFs into conversational audio, with scripts tailored to different audiences and support for six languages. It is also suited to people comfortable setting up Python, Poetry, and Ollama and running a command-line pipeline.
What is good
- Tailors scripts for experts, the general public, young people, or novices.
- Generates scripts in UNESCO’s six official languages.
- Lets users set duration and choose two to four speakers.
- Exports audio as WAV and supports voice cloning.
- Open source under the MIT license.
What to know first
- Processes one PDF at a time.
- Long PDFs may need to be split into chunks.
- Audio generation can require substantial GPU memory.
- Four languages use English voices unless users add samples.
Verdict
Choose UNESCO PDF-to-Podcast if you want a local, open-source pipeline for turning scientific PDFs into audience-tailored audio conversations. It offers language and speaker controls, but the one-PDF workflow and GPU memory demands may be limiting for larger workloads.
Get started with UNESCO PDF-to-Podcast
- Open the project website on GitHub.
- Install Python 3.9 or later, Poetry, and Ollama.
- Follow the project’s installation instructions for macOS or Linux, using Poetry or pip in a Python environment.
- Run the pipeline from the command line with a PDF document.
- Set the target duration and choose two to four speakers.
Limits to know first
The project processes one PDF at a time, and long documents may need chunking. Audio generation requires substantial GPU memory, especially with the optional 7B model; CPU use is also possible. The project is free and open source.
Questions about UNESCO PDF-to-Podcast
How much does UNESCO PDF-to-Podcast cost?
It is free and open source under the MIT license.
Which systems does it support?
Listed platforms are Linux, macOS, Windows, and self-hosted use.
Which languages can it generate?
It supports English, French, Chinese, Spanish, Arabic, and Russian. English and Chinese have included native voices; the other languages use English voices unless users add samples.
What does it use to create a podcast?
The pipeline uses Docling for PDF extraction, Ollama with Granite4 for script generation, and VibeVoice for audio synthesis.
What does setup require?
It requires Python 3.9 or later, Poetry, and Ollama. A GPU is recommended, with a CPU fallback.
What audio controls are available?
Users can set a target duration and select two to four speakers. The defaults are 10 minutes and two speakers.
Compared on AI podcast generators
- Free plan
- No
- Host dialogue
- Yes
- Source imports
- PDF documents
- Voice cloning
- Yes
- Audio export
- wav
