To be fair, most of the original recordings (with caveats listed there) were linked at the bottom, so I presume that the author still has them around - that's hardly "throwing" them away.

I don't think the goal of using Google speech-to-text was to solely use the transcriptions for the rest of the project, but you've gotta find some way to sift through those recordings and pull out the interesting bits. I think that was the right choice providing additional context and picking out the good stuff. Imagine having to listen to ~28 hours of recordings (60 sec * 1724 answered calls) when there's a service that can easily turn those recordings into a more easily consumable format, and then you can go back and listen to the neat stuff.