I'm going to try training it on a codebook to see if such a small model would work for a TTS.