Right so under the hood do you normalize to romajii? How do you deal with kunyomi vs onyomi - do they even matter? Even if you had a stream of kana, how do you

Right so under the hood do you normalize to romajii? How do you deal with kunyomi vs onyomi - do they even matter? Even if you had a stream of kana, how do you tokenize when they're not delimited? (don't know much about NLP beyond surface level software experience but have an amateur interest in Japanese)