Comet Talking Translator, 180 euros.

Let's imagine I am abroad, looking for a toilet, but cannot speak a word of English.

No problem. My smartphone has a handy translation app, which produces an English translation of any phrase I utter in Finnish.

So let's ask: "Onko teillä kylpyhuonetta?"

The translation app sends my question to Microsoft's Bing service via the Internet to be translated. After a couple of seconds, a hostile-sounding synthesised voice announces:

"This works."

What an earth? Clearly it doesn't work. I want the toilet!

Despite having come on in leaps and bounds in recent years, machine translations are still likely to produce bouts of laughter in the listener. Attempts to translate spoken text fail more often than not. But why is that?

Translating speech has three phases. Firstly, the program has to recognise the words spoken to the machine and convert the sounds into written text. After this, the text is translated into another language, a step that takes place in a blink of an eye with the help of a complex translation algorithm.

Even if the application succeeds in recognising the spoken words, the original message may get lost in translation if the translation program fails to understand the phrase that was used. This phase poses a serious challenge to a program trying to cope with the Finnish language as Finnish word conjugations can be complicated.

And even producing a fitting English translation of the Finnish phrase does not guarantee that the final result will be correct, as in the third phase of the process, the program has to convert the translated text back into speech. It, however, is not always able to pronounce all the words correctly.

A multitude of difficulties

And this multi-step process involved in speech-to-speech translation is what makes it so difficult. Even though the hit rate for each separate phase is around 80 per cent, the spoken phrase produced by the mobile app is correct in only half of the cases.

A consumer may be able to live with an 80-per cent success rate, but a 50-per cent rate is just too hit-and-miss to rely on.

But even if translation programs still get their tongues in a twist when translating speech, they fare much better with written text, thanks to statistical machine translation methods.

Before the 2000s, the algorithms of translation programs were designed around information on the grammar and vocabulary of a certain language. This approach was not successful because the programs failed to understand the text's context and in most cases struggled with colloquial expressions. Language is too complex to fit into such rigid patterns.

The new statistical machine translations are based on automatic analysis of huge quantities of written text, with programs trawling through texts and learning the properties of various languages independently.

Many options

The Internet is a real treasure trove for the developers of translation programs, containing an almost endless amount of material that can be used to hone the translation algorithms.

Launched in 2006, the Google Translate service is the best-known and most successful statistical machine translation program, which currently works with around 6,500 different language pairs.

Google's translation application is also fairly successful at translating spoken text as its speech recognition system is among the best, even occasionally understanding colloquial Finnish expressions. But even Google Translate only hits the target in two out of three cases – a bit more often when translating between two big languages.

For Finnish speakers, statistical translation methods have two major problems, which researchers are working to solve. The statistical methods cope well when at least one of the languages they are working with is widely spoken but when both languages are small, such as Finnish and Romanian, there just is not enough reference material available.

Another problem is presented by Finnish word conjugations, which the programs find difficult to get the hang of by just analysing material on the Internet.

(Helsinki Times)