It just dawned on me using my Apple Watch that you could use a simple tweak to improve dictation software, probably really easily. You just create a dataset targeted at people that have speech impediments, probably broken out by type (e.g., a lisp versus a stammer), where the training dataset is generated by people correcting text generated by the existing dictation software. That is, you use whatever dictation software you have, and then build a new dataset mapping inputs to corrected outputs. I wouldn’t even call this machine learning, it’s just a simple added layer that accounts for the fact the user has a speech impediment. In fact, you should probably allow this to be a feature in any dictation device, since someone could just have a funny voice that doesn’t map well to the training dataset or more generally, simply doesn’t perform well for whatever reason. This is a trivial adjustment, that could have real utility.

I did a quick Google search, and it looks like a simple version of this product does not exist, and instead people are using Machine Learning. There’s no way this is a hard problem, especially using machine learning. A lot of people stutter, and so there should be demand for exactly this software. I don’t have time to implement this (in particular because it requires people to generate the dataset), so you can keep this one, there it is, public domain.