Testing out Google Translate

A few days ago Maria Perramon alerted us to the launch of new automatic translation services from Google and Microsoft (Skype) and speculated about their impact on the profession. If you read the blurb and look at some of the online videos (Google on YouTube and this BBC test) that have been published, the technology certainly looks powerful: rapid, easy to use, precise. Just point and click!

Nonetheless, this sounds too good to be true… and probably is!

To see for ourselves, we did some small-scale experiments with the new Google Translate application, which is an integration of the Word Lens augmented reality technology acquired some months ago. Google Translate now includes several technologies that can be daisy-chained together in various combinations:

• Optical Character Recognition (OCR) for Image > Text
• Machine Translation (MT) for Text (L1) > Text (L2)
• Voice recognition (VR) for Sound > Text
• Voice synthesis (VS) for Text > Speech

First we tried to read (OCR) and translate two newspaper headlines.

As you can see in the video, the result here was:

• OCR: perfect, but not very fast.
  “Podemos cumple un año y ya es alternativa para gobernar.”
• MT: problematic, unintelligible.
  “We can meet a year and it is an alternative to govern.”
  The first problem here is translation of the name “Podemos”.
  Next, “cumple” is “meet” in expressions such as “meet objectives” but not here.
  Finally, “it is an alternative to govern” is not very natural in English but it is intelligible.

Having a name in the text that is also a lexical word (Podemos) is an obvious problem for machine translation. We decided to give the system a second chance with another headline on the same page.

This time the result was:

• OCR: perfect, but not very fast.
  “El coladero de la frontera de Ceuta, gran reto de la UE.”
• MT: problematic, semi-intelligible.
  “The loophole Ceuta border, challenge EU.”
  The first problem here is translation of the name “coladero” as “loophole”. This might be a good translation in another context, such as tax avoidance, but here it is wrong.
  Next, “challenge EU” reads like a verb and complement, which is difficult to  make sense of, particularly with the preceding comma.

On the basis of these two simple texts, we would have to conclude that the OCR is effective but slow and the machine translation is a hotch-potch of the good, the bad and the ugly!

Finally, we did a short experiment to check the latest sound recognition and voice synthesis functionality of Google Translate. Here the input is sound, rather than an image. We used the same headlines as above and in each case I made three attempts, trying to improve my pronuciation each time. To hear the tests, click on the play button.

“Podemos cumple una año y ya es alternativa para gobernar.”

“We can meet a year and is already native to take anything.”
My Spanish with an Anglo-Catalan accent can give rise to all sorts of confusion. Here the VR apparently hears “nativo” instead of “alternativo” and “coger nada” instead of “gobernar”!

“We can meet a year and AIS (?) now an alternative to govern.”
I tried to separate the words and pronounce more carefully but there is still a problem with “ya es”.

“We can meet a year and is now an alternative to govern.” Perfect VR but not much of a translation!

“El coladero de la frontera de Ceuta, gran reto de la UE.”

“??? Ceuta border, challenge of UEI.”
I can’t understand the start of the Spanish version.

“The loophole Ceuta border, challenge of EUI.”
“EUI”?

“The loophole Ceuta border, challenge EU.”
Perfect VR, but once again a mediocre translation.

On the basis of what we have seen here, Google Translate is truly amazing, but it’s practical use is still limited for some functions to just a handful of languages. The applications could be useful for tourists abroad trying to read signs in the street or in shop windows and product names in supermarkets. Also, perhaps, to have simple conversations with people that know how to speak slowly separating the words.

Alternatively, the VR functionality is a great pronunciation trainer, as you can hear in the recordings above! In that case, not much good at translation, but great for language learning.