Improving Text-To-Audio Models with Synthetic Captions