In five years it will be useful but still pretty clunky. I wouldn't be surprised if within ten years the error rate will be low enough that it will be satisfactory for vernacular use. But to accurately translate the nuances and complexities of scholarly and diplomatic language, medical and court records, novels and poetry, or a simple love letter, with an error rate matching human translation, will take quite a bit longer. Arguably the most difficult problem is the fact that language shapes thought. Some sentences are challenging to translate into a second language, because merely forming that thought is awkward. So I guess my answer to your question about "perfection" is, "Much longer than twenty years."
I do not think it will ever reach "perfection" because languages are constantly and rapidly changing, so that very little that is accurate in a translation program will remain so for very long. I think they will become about as reliable as spell/grammar check, but no more.
I think you younger people will live to see some astounding advances in AI. As soon as it becomes recursive and they start adding to their own programming, you may be very surprised.
