Good day!
You can train the model on data in which there is text in English and in Persian (mixed data). It will be good if you will possess a balanced large set of data. (Data you also can find on the internet and change the existing text to the language you need through translators !Artifacts possible!)
The difference from the first solution lies in the fact that here you train 2 different models. Download the dictionary eng-fa. In the next stages you will need to use VecMap, which will perform the alignment of vectors(GitHub). Combine the vectors into a single model.
If at you there is another solution, then with joy I will listen to you.