Your scenario introduce 'translation' as an additional task. I wouldn't be surprised if that ends up being the bottleneck, rather than anything specific to either the input or output language processing.