Google Translate now with real-time audio translation | Hacker News Reader

Google Translate now with real-time audio translation

wired.com

37 pointsjustnearme8 comments

Such a thing is useless without half-decent translation. Google Translate, as far as I know, currently uses a phrase-based statistical model which is beyond terrible: it isn't aware of context or grammar. It's basically operating on the theory that if you stuff enough training data into a bad model, it'll magically stop being bad.

You're lucky if you can even understand the output of a Google translation from Chinese or Japanese, let alone the translation being correct. Among machine translators, it's widely considered to be one of the worst in existence (e.g. see the "Translation Experiment" at http://amaterasu.is.moelicious.be/blog/?p=771).

Rules-based translators with vastly less training data than Google's tend to give far better results... and even then, they tend to be wrong constantly: "As you can see, four of the seven lines in this edited machine translation are almost the complete opposite of what the Japanese actually meant."

Chinese to English translation is technically very hard (I've worked on an open source translation engine -> http://adsotrans.com/downloads). While there are definitely lots of easy ways that Google can improve its translation (introducing a rules-based pre-processor to stop it from screwing up names and numbers would be a good start), it's worth remembering that Google is the start-up when it comes to machine translation. They lured Franz Och (author of GIZA++) away from academia and gave him the computing power to single-handedly crush everyone working on MT in the government-academia complex and in DOD funded companies like Language Weaver. In response, NIST changed the competition rules to penalize Google for not using NIST data.

Google needs to improve its Chinese-English machine translation for the same reason it needs to keep working on Chinese -- you can't organize the world's information if you don't understand its largest language. But the underlying problem is that Google isn't engaging with China and doesn't seem capable of hiring people who actually straddle both worlds. Not that it hasn't followed Systran down the rules-based path.

Realistically, I don't see any other company out there which is making the sort of progress that they are in MT. The fact they are still wrong on occasion when tackling distant language pairs seems more a result of their being held back by a universal architecture for developing MT systems, not because they need to go back to an exclusively rules-based system. The future is hybrid anyway.

Translation between European languages works great for me. Nothing I would ever dare to call “beyond terrible.”

The linked Translation Experiment article was fascinating - thanks :)

Ok but, the speech to text ability really sucks right now. I use google voice, and the transcriptions are ALWAYS mangled.

So you have the transcription, which sucks right now. Almost unusable.

Then you have the translation which can be okay depending on how simple the phrase is. So you're going to inputting a garbage transcription into a dubious translation engine.

Then you can output the voice, which I know works well now, great.

But the other pieces doesn't seem to be good enough yet in real world situations.

This is some serious Star Trek stuff.

Or maybe the TV show "Timecop" (I think that's the one), where the protagonist, using little more than his wits and a handheld credit card computer kept chrononaut baddies from altering the past. Considering how close my phone is to the size of a credit card, we may be less than a decade away from that kind of thing.

You mean "Time Trax": http://www.imdb.com/title/tt0106154/

Timecop was a Van Damme movie.

That's it. Darn my faulty memory.

I thought the idea was ridiculous at the time...but now I look at my phones and realize how close it is to being realized.

Reply on news.ycombinator.com