[eghad]

The problem with adapting this to other forms of atypical speech is that their recommendation system likely relies on a catalog the phonemes L2 speakers have issues with (the example most people know is the Japanese "L" "R" swap) so it's much easier to create courses with specific focuses and solutions.

If Google's Project Euphonia [0] is actually still ongoing and they release their dataset/methodology of training models with that sparse dataset I can see your idea as approachable; even accented speech is a tough problem to work on considering how many variants exist worldwide (but their approach looks good!).

[0] https://sites.research.google/euphonia/about/