The main problem is, music in your brain isn't physical sound. The only way that would work is to invent a neuron reader.
It seems to be mainly a problem with resolution and neural net density. Presumably those are going to increase over time. The neural nets themselves, I'd put money on. Bloody rapidly if the history of transistors on silicon is anything to go by. FuckifIknow with the brain scanners but it wouldn't surprise me. Maybe some
So once we have the tech in place, we scan at max resoution, umpteen images per second. So you play a known note or sound and then you tell the trainer guy to imagine it just so. Then spit the scanner output in the markov and crosslink the sound/note/tune with the scan.
Eventually you're catching notes and waveforms, then whole sequences of those, then it'll be polyphonic, multi tibral, symphonic... Remember the Moog? First stage - mental Stylophonetm Did you have those in the states?
It'd get interesting when you were applying enough horsepower