It's a great site, but not really the sort of thing that would be useful for FSO. Generating these voices takes massive amounts of computing power and can't be done in real-time. That site would be more suitable for modders who generate voice files ahead of time and include them in their mod release.
The training part is the computationally expensive one but the author said that he's open to add new voice sets if the proper material is given (and it's notoriously hard to find quality datasets) so FS2 clips could be used as training sets and with that you can get in return new machine-generated audio with little effort (as a pre-trained model).
As for feasibility I wasn't considering training in-engine as that would be way too intensive but instead using a pre-trained model.
With a pre-trained model one could either generate audio files in batches from submitted text (similar to the way the site works now, only generating files as output) or, with an higher effort, the pre-trained model could in theory be integrated in FRED for example and generate the audio clips on mission save.
The nice thing about neural networks of all sorts is that while training is crazy expensive from a computational side, running the model is relatively trivial and is routinely done by low-power devices like smartphones.
That said, this is merely a speculative thought on how AI content generation could be useful to stretch source material, not a feature request of course so I will leave any value consideration to you guys.