This proves that it's a real speaker rather than a text-to-speech engine
The robot voices can read anything you throw it, including new sentences if the course maintainers decide to add some later on. (Not sure whether that's possible.) They can also read individual words if you hover over them in a sentence discussion.
But with real speakers (e.g. the Irish course, the Ukrainian one, Esperanto, a couple of others), there's only a budget for a certain number of sentences -- 3000, I believe. Any above that number, or any new ones added after the recording was done, will have no audio.
I sympathise very much; this was one of the things I found frustrating as well at the time! Especially given how differently Irish maps letters to sounds compared to English.
I don't know whether they replaced the 3000 (or whatever it was) sentences by the old speaker with 3000 recordings of the same sentences by a new one, or added 3000 to the existing 3000, but either way, you're not going to have audio for all sentences unfortunately -- it's a money thing mostly, I think (and partly time, perhaps).