Using Pre–trained Models for Phoneme Representation in Czech Speech Synthesis