The pretrained network is calibrated for English, but if you trained it on non-English data it should still work fine since the entire network is retrained (assuming there is enough overlap with characters in the vocabulary; e.g. Spanish would work fine but CJK languages would not).
