Bi-lingual Model Training

[bely66]

I'm trying to Train the Recognition model for Two Languages English and a Non-Latin Language.
I can see that it worked with Chinese with a huge number of characters.
However I'm facing a problem with Language Confusion, Some times some of the words in the dataset are classified in English instead of the non-latin language and the number of times that happen isn't low.

So any tips on how to optimize the Training Dataset to get an optimal performance for the two languages?

Regards,