Google has developed a special algorithm that is able to directly translate speech into another language, without using text and retaining voice data and the speech rate of the speaker. This was reported in the blog of an American company.
It is noted that the management of Google decided to abandon the translation of speech to text in Google Translate and sound it after. The fact that this leads to errors in translation. From now on, the speaker immediately will be transferred from one to another language.
For this neural network with long short-term memory, will accept voice input and process it as desired, then generate on this basis new spectrogram in a different language.
The developers of the program have been able to keep the original speaker’s voice even in a translated speech. Now, the experts teach the algorithm to understand the vector representation of words in both languages, and grammatically correct forms and combinations of words in a large sample.
Earlier “MIR 24” has written of the Italian, who has been on trial in Austria because of an error of the electronic translator.