It's probably impossible to detect ALL languages without training for them specifically, but there's good generalization happening. Our model is a unified model rather than a separate model per language. We started out with language-specific models but found that the unified approach yielded slightly better results in addition to being more efficient to train.