First, think about what you want to do with the language. If you were to learn classical Arabic, you could understand news broadcasts, or the Koran, but little spoken language in any Arab country you were to visit.
That's because most regions speak a dialect of Arabic which can be quite different from classical Arabic. Apparently, the translators trained by the US Army for Iraq service learned Modern Standard Elementary Arabic, and when they got to Iraq they couldn't understand what the locals were saying. Thus the army had to hire local translators.
GPUs are a bit silly, even though they work well, since the power consumption per unit of computation is higher than FPGAs , and orders of magnitude higher than the human brain. The problem with deep learning is that real brains don't do multiply-accumulates and so contain no power-sucking floating point hardware. So there is a lot of room for improvement here.