请教作者大佬，训练模型问题

继续作者75K训练到390，LOSS从350左右就一直在025，现在390还是025
用起来效果一般，有些地方吞音，有些地方发音奇怪
是不是LOSS还不够低？
现在是继续训练模型，还是训练vocoder？
有什么优化模型效果的方法吗？

The issue mentions continuing 75K training to step 390, a loss around 0.25, and deciding between more model training and vocoder training. No file, test, or entry point is identified, so first clarify the relevant training configuration and expected output quality. Done would require an agreed diagnosis and reproducible improvement criteria.

Written by the indexing model from the issue text.