Self-Supervised Learning for Speech Enhancement through Synthesis