To each his own dubbing speaker: electronic translation system to render directly into foreign languages
In the meantime, there are the first portable translation systems that work with voice recognition: the tourist speaks his wish into the device and it quacks the translation in the foreign language out of the loudspeaker. It may work, but it makes a rather damned impression and stops at the beginning. In the future the translation will be "live", almost without delay.
In the near future, americans and englishmen will be able to speak spanish or german even if they do not speak the language at all. This is the plan of tanja schultz, a linguist at carnegie mellon university in pittsburgh, usa, as reported in new scientist. The first systems of this type are already available for purchase: with these, the text must first be spoken into the device in one’s own native language, then a button must be pressed, whereupon the translation is reproduced in the foreign language. This may be better than the long leafing through a vocabulary book, but it is not really practical for longer conversations – and even if you only want to buy a ticket, it looks worse to the bystanders than a real handicap.
Schultz already sees the step of having to speak the sentences aloud in one’s own language as an annoying detour: in this way, the time needed for a conversation is doubled, because the translator can only either listen or speak, and the other person can’t cope with you babbling away in your own language at the same time as an electronic device is talking in another language. Even the interviews that are often held in germany with dubbed sound and the original sound underneath (which is raised in volume during pauses in the translator’s speech) are too much for some tv viewers.
Mute articulation for the computer
Instead, schultz has developed a system that uses electrodes on the face and neck to interpret the electrical impulses that the brain sends to the facial muscles and tongue to pronounce the appropriate words. The user of the translator does not have to actually pronounce the words, since no speech recognition is used, but just make the appropriate mouth movements and speak the words silently. The translation would then come out of the device with a slight delay while the user is still "speaking"" – almost like a dubbed movie.
The ultimate goal is that with the help of such a device, a normal flowing conversation becomes possible, according to the linguist alan w. Black, who also works at carnegie mellon university. Black. However, both parties had to use such an installation – simply handing the translator over to the other party and switching the direction of translation, as with previous devices, could not go far here.
In october 2005, tanja schultz and her colleague at carnegie mellon university demonstrated dr. Alex waibel of the technical university of karlsruhe, germany, developed the first automatic translator that could pick up electrical signals from the face and throat muscles and convert them into written text or synthetic speech. The technique in question is called sub-vocal voice recognition.
Vocabulary without borders
The software during the demonstration ran on a normal notebook and translated mandarin chinese into english or spanish, but its vocabulary was very limited, with only about 100 words, which the user also had to train the system to use. In the meantime, the vocabulary is theoretically unlimited, since not only words, but also phonemes are recognized, from which the words can be composed. For english speech recognition 45 phonemes are supposed to be sufficient. This is done in a similar way as it is done with normal speech recognition software ("the hot samantha wants to text you right now") – a list is used, which contains probabilities, which phonemes are usually closely related and in which order. If it doesn’t work at first go with the full speech recognition, the most probable result is chosen on the basis of these tables.
However, the results are still meager: if the system is presented with words that it has never heard before in this order, it is only 62% likely to choose the correct translation, which is nevertheless a major step forward, according to chuck jorgensen, who is working on just this kind of speech recognition for robot control at nasa ames research center in california.