A Comparative Analysis of Pretrained Language Models for Text-to-Speech