I just can't stand how non-deterministic many deep learning TTSes are. At least the classical ones have predictable pronunciation which can be worked around if needed.