Scaling Effect of Self-Supervised Speech Models