JiaYK
JiaYK
strict=False，会显示正确的韵母[n2]（虽然个人觉得可能n4比较常用）
嗯嗯，按标准来说，的确没有 ```n``` 这个韵母，才发现自己用了这么多年的 ```嗯嗯``` 其实都是错的 试了一试，感觉 好像也不是声母的 ```n``` ，但是这么做应该比较方便，不然 ```嗯``` 这个字大多数情况下都要想办法单独再写处理逻辑 百度给的是 ```ng```，但是也没有 ```ng```这个韵母
@kan-bayashi Thank you very much for your reply. alignment loss in first step is generator_align_loss=29.134, in 989 epoch 601-650 batch is generator_align_loss=24.842. I think it's a good idea to scaling...
Thank you very much for your experience. I changed alignment loss and retrained it. I hope the model can converge well.
To be honest, the training was not successful. I changed the parameter of ```lambda_align``` several times, but in the end, the alignment effect was very poor. There was no semantic...
The batch bins I use is 1200000 by one 3090. I look forward to your good news
经过不严谨测试，选择 -DCMAKE_BUILD_TYPE=Release 并把之前build的文件夹删除，重新编译，就可以win+cuda11.6在编译成功
You can use this replace convert.py and you can get test.wav. That is the result. `# -*- coding: utf-8 -*- # /usr/bin/python2 from __future__ import print_function import argparse from data_load...
You don't have Timit datasets?
It's better to work under Linux. Tensorflow-gpu does not support Python2.7.