I just learned last week that we have software that can recognize your voice and write text. I know we have had software to convert text to verbal information for a few years. So does anyone have any idea how long until we have functioning translational software that allows you to speak in english and have it translated into something else?
But that device needs to be programmed to understand your dialect and personal voice, and I think current voice to text programs require that too. Is 5 years a good ballpark figure for how long it will take?
How long until we have software that allows you to speak in one language (English), be heard in another (German, for example) then allows the computer to understand the other persons language and convert it into your language without either of you having to program the software to understand your personal voices first and foremost, or at the very least requiring a very easy method of programming (maybe talking for 10 seconds and saying a list of phonetic sounds into a microphone)?
It’s still pretty far off. Even the best translation programs are only about 50% accurate.
Now, it might be easier to do limited-domain translations (for instance, there are only so many things to talk about as part of a business deal), but I don’t know of anyone who’s actually doing that.
According to the experts, good translation software is ten years away and has been ten years away at least since the 50s.
What has become clearer and clearer as time goes on is that, while we are getting a handle on natural language syntax, any real understanding of natural language semantics is probably a long way off and, without that understanding, translation is impossible.
Note that there are a lot of text translation programs out there. I have seen the output of such programs, and depending on the “idiom-ness” of the original, you can get readable but stilted text to something that looks like stream-of-conciousness ramblings from a schizophrenic.
In order to do a somewhat decent job with text, you have to have solved “natural language”, which is incredibly difficult (despite the AI people claiming to be “ten years away” since the 50s).
Throw in speech recognition on top of that and you have one amazingly difficult problem here.
I would be astonished if anything of reasonable quality could be done in 50 years.
The failure of the AI people to recognize that saying “ten years away” over and over is Not A Good Way To Approach A Hard Problem means that even 100 years would be overly optimistic.
According to xtisme’s link we already have one way translators that are perfectly understandable. You talk, the device writes the text, translates it and speaks it to people who can understand it. If you had the other person program the voice recognition program to their own voice you could have a 2 way conversation in two different languages. So I don’t see why it would be 50 years away.
I work on natural language systems, and we have done some generalized translation systems that work very well. The problem is that what we are doing is generating output in (for instance) Japanese from the internal, semantic representation of what the speaker said in (again, for instance) English. This means that no matter what way you ask what the time is, the system will choose one of the its ways of expressing that to the listener. It also means that the system has to have domain knowledge of the subject at hand, no mean task for very broad areas. It only works at all because there is no direct connection between the language parsed into our internal representation (which is canonical and not related to language) and how the system responds.
We don’t have speech recognition systems; we use other companies’ technology. Currently, our phone-based products are using Nuance. Our most recent launch was the first Nuance dictation-based product in Japan, and the first general FAQ system integrating their technology in the world.
Umm, its Discovery Channel hype of some company’s product. It has nothing to do with reality. Also, crop circles aren’t made by aliens and psychics can’t talk to the dead.
Here are some examples of translations into English from Mandarin Chinese news broadcasts from a state-of-the-art fully asutomatic speech recognition + machine translation system:
This is a research system; production systems would typically have hand-coded rules to improve performance. Nevertheless, you can see that we’re still a long way away from being able to do this reliably.
I am cautiously optimistic that a good way to solve the natural language problem is not by designing clever algorithms and really trying to compute language structures, but with truly massive lookup tables.
The kind that Google and Wikipedia are amassing. If Google succeeds in its proposed plans to scan and index a huge amount of literature in multiple languages, they may have the necessary data to provide very effective translation services. Of course, this still requires clever algorithms to make searching quick and cheap, but there are many well-known AI methods to do that kind of thing effectively. I believe that Robert X. Cringely wrote a column about this at some point in the last year or so, but I can’t seem to find it.
Is anyone else reminded of the famous quip about fusion power, that it’s “30 years in the future, and always will be”?
I had to use Google’s “translate this page” function a lot for a project at work last year. Happily I only had to learn the subject matter of a lot of scientific journals; sentences like “Everything approximately around the turtle is here topic” were sufficient for the purpose. That particular translation was from German, hence the passing resemblance to an English sentence; translations from Chinese, Japanese etc. resembled the kind of word salad carterba has already posted.
One problem I ran into was context; words that mean one thing in general usage often mean quite different things to scientists or engineers, so a simple word-for-word substitution may substitute the wrong word altogether. “Molar” means one thing to a dentist and another to a chemist. “Mole” means one thing to a gardener, another at the makeup counter, another to a chemist and yet another thing to a civil engineer (who also has another understanding of the word “groin”). I’m sure there are plenty of similar examples in other languages as well; I ran into some in German during a later project.
We have done a robot that includes translation as part of teaching English to Japanese children, as well as numerous research systems in Japan (where Natural Language Processing research is focused almost entirely on translation systems). However, as I said before, the limitation to our technology is that it can only translate within the domain of content that’s been entered. The canonical concepts have to be linked to specific facts by an author. This means that you can’t just have a web page that will translate an arbitrary string of text; it can only translate text that it can successfully parse into concepts.