If this is the same technique I’ve read about before, they need to start with a training set from the particular keyboard. You map the sounds of the keystrokes against the produced text.
I haven’t read about this approach yet, but my guess is that there’s enough of a difference between the sound of each key that they can tell with some percent accuracy which key was hit. It will probably be higher confidence for some than for others, and higher with a mechanical keyboard than one with butterfly switches. Combine that with the same kind of technology you use for autocomplete, and just keep iterating as more text is entered.