Japanese sounds unnatural

[michaellin99999]

I have combined the phoneme sets for all three langauges,
English, Chinese, Japanese and started fine tuning using a datset comprised of all three speech languages
The base model I use is the chinese and english base.
However after 500 epochs, the result I get, chinese is good, english is good, however japanese sounds unnatural .
My udnerstanding is that the phonemes are correct but the tone is just not how japanese is spoken.
What can I do to improve this?

Here is a sample data of the japanese output. https://soundcloud.com/michael-lin-674069136/japanese-test

[eliteexod]

Are you using it on Docker?

[michaellin99999]

i have tried on docker and also onnx runtime both sound like this

[baishouwujianfei]

Hello, may I ask which pre-trained model you used for fine-tuning? How long did you train? How is the config set up? The model I trained cannot produce complete sentences, and the speech is very strange.