I think you might have misunderstood the article. In one case they used the sound input from a Zoom meeting and as a reference they used the chat messenges from set zoom meetings. No keyloggers required.

I haven’t read the paper yet, but the article doesn’t go into detail about possible flaws. Like, how would the software differentiate between double assigned symbols on the numpad and the main rows? Does it use spell check to predict words that are not 100% conclusive? What about external keyboards? What if the distance to the microphone changes? What about backspace? People make a lot of mistakes while typing. How would the program determine if something was deleted if it doesn’t show up in the text? Etc.

I have no doubt that under lab conditions a recognition rate of 93% is realistic, but I doubt that this is applicable in the real world. Noboby sits in a video conference quietly typing away at their keyboard. A single uttered word can throw of your whole training data. Most importantly, all video or audio call apps or programs have an activation threshold for the microphone enabled by default to save on bandwith. Typing is mostly below that threshold. Any other means of collecting the data will require you to have access to the device to a point where installing a keylogger is easier.