Model close to getting aligned, but still jumbled

[Scrollkeeper]

Hi again, I've trained the model up to 89,000 steps. It is close to getting aligned, but is still not quite there. The problem is that it has been pretty much the same since 38,000 steps. Synthesis sounds flawless when exporting from train.py but is very poor from eval.py Some examples:


It could just be that the dataset isn't cohesive enough, but I'm not certain.
Thank you for your time and input! :)

[Scrollkeeper]

Some more graphs: