How can we train 2 languages at once?

[rooshil-bhatia]

I tried Pre-training with English and one other language data which is roughly around 4000 hours of data for 2200000 steps. But i am getting weird outputs when Reference Audio and Reference Text is in English and the gen text is of the other language.
By weird output i mean to say that the output i am getting is also speaking some part of the reference audio I am suspecting the duration is not being properly calculated if there is a change in the languages of reference and generation prompt.

Kindly share some solutions or if somebody experiencing this we can discuss few ideas in this thread as well.