There have been many articles recently, like this one, about advances in the automated translation of speech, and I’ve even read stories about armies using them. I find the latter news very worrying.
Automated translation of speech basically combines two previously existing technologies: speech recognition and machine translation. The problems with the latter are well publicised, and despite the advances made, the problems remain. Google’s corpus-based translations mean that sentences tend to be more coherent nowadays, but a coherent sentence can also be an incorrect translation.
Voice recognition has come on leaps and bounds recently. I use it myself when translating. But as every user of such technology knows, you have to train it to your voice, and even then it makes mistakes that you have to correct. The article from The Times I’ve provided a link to discusses the problem of understanding “high-speed Glaswegian slang”. Current technology would no doubt be absolutely useless at understanding this. But what about more standard forms of English?
I decided to test how Google’s new speech-recognition tool would cope with the Queen’s English — literally the Queen’s English — a speech made by Queen Elizabeth II to parliament in 2009. As I expected, because the tool is not trained to the individual’s voice, the results are pretty awful. To see the video, click on this link. Pause the video, move your mouse over the “CC” button at the bottom of the video, then click on “Transcribe Audio” (don’t click on “English”, as that just gives you captions provided by a human, rather than the automated transcription), click on OK, and the video begins. The Queen tells us how she “was a man that’s in the house of common” [sic].
We can, if we wish, have these captions translated into another language. Just go to the “CC” box and click on “Translate Captions”, then choose your language. But the machine translation will only translate what it’s asked to translate, so we are still likely to get told that the Queen is a man. The translations into the three other languages I work with begin like this:
Catalan: “Jo era un home que està a la Cambra dels Comuns”
Spanish: “Yo era un hombre que está en la Cámara de los Comunes”
French: “J’étais un homme qui est dans la Chambre des communes”
As you can see, there is a very high risk of misunderstanding when using this technology. If the army wants to communicate with people in other languages, I’m afraid they’re just going to have to hire trained interpreters.