The Holy Grail of Voice Translation, Now on Skype

If you followed the computer revolution from the inception of the microprocessor, you would know that there are a number of early promises that have yet to be fulfilled.

One of these is on-the-fly translated speech, where you say something in English and it is repeated in another language, like Spanish, in near real-time. Though still a work in progress, Microsoft is now one step closer with Skype Translator, the first phase of which went live this week.

If this comes close to working and isn't a joke, it is the product of the decade. Right now, Skype Translator supports Spanish-to-English and English-to-Spanish translations, but more languages are set to follow.

The typical problem with these sorts of things is the outrageous difficulty level. I have yet to see a decent translation package that does text-to-text translations that work well. Speech translation has an entirely different level of difficulty, especially in real time.

Microsoft does have a movie showing two kids speaking in Spanish and English that makes it look like it works well. But I saw a similar system demonstrated by IBM almost 20 years ago that never saw the light of day. IBM's version was developed in the 1990s, when there was a speech-recognition mania led by Lernout & Hauspie, a company that went for broke—before going broke over a fraud scandal in 2001. Before that, though, it was buying every speech technology company it could, including Dragon Systems, Berkeley Speech Technologies, Fonix, Dictaphone, and others. Microsoft had an 8 percent stake in L&H and ended up with some of the technologies as a bankruptcy parting gift.

Now Skype Translator suddenly appears. Coincidence? Whatever the case, this entire technology has taken too long to get to this point.

Look at the miserable text-to-text computer translations done by Google and others. Text does not and cannot avoid the most obvious of mistakes. None of this approaches the complexity of speech translation, which entails heavy signal processing.

I'm a wine collector and frequently visit French winery websites. My French is only okay, so I often turn on the translation to expedite reading the site. Here we are in 2014 and none of the translators can figure out that the translation of Château Margaux is Château Margaux, not Castle Margaux. How difficult is it to refrain and not translate a commonly used word used in a proper name, such as château, usually referring to a specific winery, into the word castle? Apparently it is impossible. They all do it.

How hard is it to write some exception into the code that tells the translator that it is on a website about Bordeaux wine? While on that site the word château does not mean castle. Often the translator will try and decode the rest of the château name as well, with ridiculous results. The best you can generally do with text translation is get an inkling as to what the site says.

Ask yourself: if text is so difficult, how will Microsoft manage speech?

It's next to impossible even in the same language. Take the Google Voice phone service. It has a speech-to-text message-taking function. I have never received a voice-to-text message that even comes close to what the person actually said. Not once.

I will admit that in a quiet room when you talk distinctly and clearly, voice recognition does well. I use it for text messages on my Android phone. But in a real conversation, nobody talks like that. I admire Microsoft for rolling this out. But it will not work as advertised.

That said, it is at least something to play with. It might even trigger a new generation of research. So I won't be complaining too much.

Everybody wants this. Let's get back to work on it.