Turning Research Papers Into a Podcast

I’ve found that consuming knowledge in a wide range of modalities reinforces a deeper understanding of certain topics.

My latest experiment has been listening to research papers, then reading them.

I’ve extracted the text from PDF files and piped it to OpenAI’s voice API. It was cheaper to convert Andrej Kaparthy’s PhD thesis to a 3hr mp3 file for $3.20 than to print it for 0.2c/side (~$19 for 90 pages).

A lot of the time there’s a few pages that aren’t worth transforming into audio such as references, table of contents, appendices—so they get trimmed away.

One constraint with OpenAI’s voice API is the max character count of 4096 per request. I ended up breaking up the paper into $n$ chunks of 4096 characters each. Parallelising this produced very quick results.

import os
import concurrent.futures

def process_chunk(filepath: Path, chunk_info: tuple, voice) -> None:
  # ...
  # send request OpenAI
  # stream response to mp3 file

chunks = [] # n chunks of maximum 4096 characters each
max_workers = min(32, os.cpu_count() + 4)
voice = random.choice(["alloy", "echo", "fable", "onyx", "nova", "shimmer"])

with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
	executor.map(process_chunk, repeat(filepath), chunks, repeat(voice))

The smaller chunks of mp3 files had to be merged back together using ffmpeg.org.

But now all we have is an 3hr mp3 file. How do we consume this easily? I want the convenience of having this available wherever I go, whenever I have a moment to spare. I ended up automating the mp3 creation and uploading to a GCP bucket, then hosting a Podcast RSS feed from GitHub pages for free.

After two o1-mini prompts I had a full podcast setup. Its that easy. I copy pasted the feed URL lbxa.github.io/arxiv-to-mp3/rss.xml into the free Overcast app on iOS and voilà. Podcast.

I’m excited to see if this unlocks a new dimension of knowledge dissemination. I’ll be either listening to my favourite house tracks or seminal papers in science and technology. What a dichotomy. The random narration voice selection should also keep the Podcast quite dynamic with 6 co-hosts to listen to—more than the usual Podcast.

An unsolved limitation is $\LaTeX$. PDF text extraction is a non-trivial task but I chose to extract text without OCR to get something working quickly. OCR could be more robust.

If you have any feedback, please create an issue on github.com/lbxa/arxiv-to-mp3. The Podcast home page is live at lbxa.github.io/arxiv-to-mp3. It’s pretty naked right now but we’ll see how this evolves as more content gets added.