To a degree there is some learning on both sides. I don't know if I've seen it mentioned in detail in papers that I've read, but it was a general tone at a few confs when dealing with coarse motor functionality.

As per why the same approach is not well suited for speech, one (of many) components is the data bandwidth you need for speech to happen at a reasonable rate. For many applications the ML involved in the brain computer interface is only able to extract a few bits per second. It works for applications where the output is coarse (e.g. up/down/left/right) or where large delays are acceptable, but it's not the best fit for speech.