Open fadelkon opened 4 years ago
The state-of-the-art for open source MT is opennmt. We also use it but i see two possible problems, one is the need to have trained models. As far as I know, there is not a good repository of already trained models, and one needs to train each pair one by one.
The second issue is the performance, i.e. making it work for android, and also the different architectures will be important. In principle opennmt does not have serious performance issues on desktop systems, but I have no idea whether it will have issues in android architectures.
I know, Apertium is a generation behind the fancy deep learning translation engines, but, afaik, it's the best we have right now in a usable way at the libre world.
There are open source engines for neural translation, but I don't know of a finished piece of software with an open instance or even self-hostable.
So, Apertium is still good enough for many things, and has an api engine. In fact, @soimort uses the public instance in translate-shell.
However, I'm sure @collectivat know better the state of the art of machine translation with libre software and libre data, maybe they have some thoughts to share.