I know some readers wish they could listen to articles – particularly the longer ones – while they’re doing other things. I cannot take the time to do a proper spoken version myself, nor can I afford to pay someone else to do it. So when I heard from one reader that they did it this way, I figured I might try offering that to everyone who is interested.

This is so, so awkward for me, because the results, as you can hear, are nowhere near the level of quality I would prefer.1¹ Role adherence is bad, intonation is off, conversation is stiff, etc. Even frontier tts models occasionally generate words with weird artifacts. But even with the crappiness of the outputs of these models, they may be better than nothing. People who prefer quality (or reading) can still read the original article and ignore the experiment.

The obvious way to generate a spoken version would be a plain text-to-speech reading of the original article. I actually use that when I self-edit2² Some issues are easier found when one hears someone else read the words as written., but they are a terrible listening experience in general. The articles on this site are written to be read. They include code, plots, equations, tables, and I expect readers to jump back and forth and cross-reference as they read. Not possible in the more linear format of sound.

I have tried a few ways to get language models to compress articles into a comprehensible listening experience3³ Both for this experiment, but also for myself when I want to learn something while vacuuming, or whatever., and the podcast format is the one I’ve had most success with.

I’d much appreciate feedback on this experiment, so I know if it’s something I should bother keeping up with. I’m primarily interested in hearing from people who like listening to technical stuff: are the ai fake podcast versions of articles useful to you? If yes, why? If not, why not?

If you have experience in this area and know of ways I could improve this, I’d love to hear more from you as well.