-
Notifications
You must be signed in to change notification settings - Fork 1
擴增新的語言模型 #5
Comments
啊沒有,就修改了symbols之後就直接訓了。 |
可以參考 #1 |
|
我能看一下源碼嗎?聽上去是symbols mapping的問題 |
想問問您需要哪部分的源碼,我應該能提供。 |
|
Korean.zip |
You might want to try implementing on Style-Bert-VITS2. We found that training on it (and it's based on bert-vits 2.1) works better (can already speak some Cantonese with 100 steps) Our implement: |
你好,我使用这个工程的main branch,基于十几小时的粤语数据集进行训练,训练到235epoch,82000步了,合成的声音无法正常说话,一直‘啊啊啊啊啊啊’,请问可能是什么原因? |
Not sure what causes this issue, as you mentioned you used around 10 hours of training data and 235 epochs already You may want to try our new repo instead to see if it fix |
能否问问您是怎么训练一个新的语言的,我有尝试训练另外一种语言(韩语),并且保留中英日三个symbol但是训练不起来。
不知道您是怎么训练起来的。是不是需要把某些设定一开始先拔掉,然后好好的训练韩语的声学模型等等,例如关掉DD/SLM模组的训练这样吗? 因为看您的训练似乎是使用替代symbol的方式把中文给砍掉? 但可能是因为四川话和中文较相关,但另外一个客语的似乎就直接砍到剩下只有Hakka phoneme。想问问整體的训练方式!! 感谢您!
The text was updated successfully, but these errors were encountered: