Towards Selection of Text-to-speech Data to Augment ASR Training