Comment by beanlog

It sounds useful to me because you can use tone information to help with the translation, which text-to-text translation can't do. But I'm not sure if that's how this model actually works.