Low-Resource Cross-Domain Singing Voice Synthesis via Reduced Self-Supervised Speech Representations