Slashdot Mirror
Google Shooting For Smartphone Universal Translator
nikki4 writes to tell us that in giving some major improvement tweaks to its existing voice recognition tool for the Smartphone, Google is aiming for new translator software that will provide instant translation of foreign languages. "The company has already created an automatic system for translating text on computers, which is being honed by scanning millions of multi-lingual websites and documents. So far it covers 52 languages, adding Haitian Creole last week. Google also has a voice recognition system that enables phone users to conduct web searches by speaking commands into their phones rather than typing them in. Now it is working on combining the two technologies to produce software capable of understanding a caller’s voice and translating it into a synthetic equivalent in a foreign language."
Maybe my experience is atypical, but Google doesn't seem to translate pages very well. I can only imagine how bad it will be having a phone do this. "Did that guy's phone just call me what I think it did?"
Caution: not for use with Hungarian Tobacconists.
Does anyone use voice recognition software? Here are a couple of my voicemails transcribed by Google Voice:
Hey man, Hello, this is gonna ask you about Stockton uncle in a missed your call, so, so give well. Okay bye.
Hey it's me and I for me. Long, My of the day. So Hey Jared, Here doing. If you come for another anti, gimme a call before you go to sleep and stuff, so give me a favor you familiar with it. I love you bye.
Whale
Google is in the business of collecting data and applying it to practical problems. I imagine the voice-to-text will be vastly improved over its generations by users accepting/rejecting the vtt result and them pooling the results data. The same thing could be done for translation from one language to another.
I see it as crowdsourcing the algorithm accuracy checks among millions of people, allowing them to improve the algo at a much faster rate than they (or their competitors) would otherwise be able to do in a closed testing environment.
This is all speculating on the fact that google pools results of translations or VTT and whether the user accepts/declines them. I wouldn't be surprised in the slightest if they did.
Tablespoons, by an Apple Newton
or [allegedly] what happens when you run Jabberwocky through a handwriting recognition program.... :-)
-----------
Teas Willis, and the sticky tours
Did gym and Gibbs in the wake.
All mimes were the borrowers,
And the moderate Belgrade.
'Beware the tablespoon my son,
The jaws that bite, the Claus that catch.
Beware the Subjects bird, and shred
The serious Bandwidth!'
He took his Verbal sword in hand:
Long time the monitors fog he sought,
So rested he by the Tumbled tree,
Long time the monitors fog he sought,
And as in selfish thought he stood,
The tablespoon, with eyes of Flame,
Came stifling through the trigger wood,
And troubled as it came!
One, two! One, two! And through and through,
The Verbal blade went thicker shade.
He left it dead, and with its head,
He went gambling back.
'And host Thai slash the tablespoon?
Come to my arms my bearish boy.
Oh various day! Cartoon! Cathay!'
He charted in his joy.
Teas Willis, and the sticky tours
Did gym and Gibbs in the wake.
All mimes were the borrowers,
And the moderate Belgrade.
"Shaka, when the walls fell"
I tried to translate your sentence multiple times, then back to English so I could post the ridiculous result.
Except google's translation was actually pretty good.
Stephen Fry offers...
"Hi, Stephen, it’s Natasha from BBC Newsnight in London. Just to say I’ve sent you two texts. One is to say that we could do it at eleven am your time after the launch, or any time sooner after the launch, or we could do it at midday as we suggested earlier. I, er, if you could text me back about that, and I’ve sent you the details of Skype that you need to do too. If you could give me a call back. Enjoy the launch and I’ll speak to you after that. Thank you Bye."
I’ve transcribed it from the voicemail sound file that resides online on my inbox on the Google Voice site. All fine. I have also ticked the option for Google Voice to send me a text transcript of any voicemail. Below is their interpretation of Natasha’s message it’s rather endearing how hopelessly wrong the largest company on earth gets it.
"Hi Stephen. It’s Jeff from BBC needs in nuns. And just to say I sent 80 tax, one, if to say we could do it. I left in i a m your time off to go into any time soon, or the court and full we could grab me today as we suggested at. A. F. I. If you could text me back byebye. I’ve sent you the details of skylights that you need to 3 T if you could give me a call. Bye. Enjoy the loans. I’ll speak to you after that. Thank you. Bye"
On a more serious note, such transcripts at least allow you to get an idea of the rough content and tone of a message without having to stop and listen to it, a much more concentration-intensive task.
No kidding!!! What do you say at this point?
This seems like something that the NSA is probably salivating over. Imagine being able to translate intercepts in near real time with accurate voice recognition. I'm sure they already have imagined it. That technology is nothing short of a Manhattan Project for the SIGINT community.
I tried to translate your sentence multiple times, then back to English so I could post the ridiculous result.
Except google's translation was actually pretty good.
Try a more complicated example. For instance, starting here:
"It's probably pretty good at translating translations it produces back into the same source text. If you figure that a phrase structure in one language corresponds to a certain data structure in Google Translate, then it makes sense that this data structure would survive multiple passes through the same restructuring algorithm..."
translating to Japanese and back to English yields this:
"It is translated to produce translated text back to the very same source is probably a good thing. Cases, one single phrase structure of language specific data structures in the Google translation, it is this data structure makes sense and survival of multiple paths through the same algorithm structure corresponding figures ..."
Here you've got badly handled idiomatic phrases all around... Like the Google translation to Japanese used "seiseisuru honyaku no honyaku dewa ii koto da" at the end of the first sentence ("created-translation's translation is good" or something like that). On the translation back the connection between "good" and "translation" was lost - Google slapped on a fairly generic "is probably a good thing" - picking the bit of uncertainty out of the start of the Japanese sentence and combining that with the "dewa ii koto da" - but dropping the whole idea of what it is that's good... Which is something that can be kind of vague in the structure of Japanese... Meanwhile, the phrase "source text" was transliterated into katakana, but it got broken up in the translation back to English and wound up in two different locations in the sentence...
The whole conditional clause in the second sentence got kind of mangled. In the Japanese translation it starts with "baai wa": baai means "case" or "situation" - the structure of the sentence establishes this "case" being described as a possibility... Google lost all that, and just said "cases," Then, at the end of the sentence, after the ellipsis, "figure", from "if you figure" in the English original, was tacked on as "taiousuru zu" - "interacting drawing" or "interacting figures". In the return-to-English version this somehow wound up back before the ellipsis again.
The rest of the second sentence in Japanese is something like "if this data structure uses the same intermediary algorithm, several passes of the algorithm should be survived and it should make sense." The apparent problem there is something analogous to operator precedence in arithmetic. The "and" is meant to mean that the surviving translation should still make sense - but this clause apparently got broken up... like the reverse translation assumed that "uses the same intermediary algorithm... should be survived" was all one stand-alone clause - and so it assumed that clause had nothing to do with "this data structure", switched the order of the "and" around, etc...
My hobby is building Gundam models - one of the most comprehensive review sites for new Gundam kits is in Korean. Believe me, we all try using Google translate or Babelfish on Dalong's site from time to time, but the result is rarely worth the effort.
Bow-ties are cool.