I’m out of reviews to test this on, but you can try this fix for Keisei Phonetic Semantic Composition: