Scaling Properties of Speech Language Models