People recognize familiar words without difficulty regardless of the voice and intonation it is pronounced with. “Six” or “eight” remain six and eight for a person no matter how they are pronounced – in a loud voice or in a whisper, in an excited or a calm voice, by the voice of an old man or a child, by that of a man or a woman. Our brains separate the semantic part from the mass of background sound.
As for a machine, each variant of a voice is unique. That is why speech recognition programs usually require training. As a result of training, an enormous library is built up in the memory of the silicon brain, where thousands of possible options of pronunciation of the same words (for example, numerals) are stored. Having heard a word, the computer would look through the library and almost certainly something similar to the heard word will be found in it.
The approach suggested by the scientists from the Institute of Radio Engineering and Electronics, at the Russian Academy of Sciences, is rather more human than machine: a computer under the researchers’ guidance filters individual peculiarities. It picks out the most basic things and rejects all immaterial ones. As a result, the machine even acquires the ability to discern individual sounds and to put together in its “mind” familiar words from these sounds.
As a result, 1 kilobyte would be sufficient for a processor to confidently recognize all numerals and some simple commands, however, pronounced (although only in Russian at the moment). Several dozen people with far-from-ideal articulation – tried to confuse the quick-witted program, by pronouncing numerals either in a whisper or in a voice trembling with excitement. However, the computer successfully rejected emotional frequencies as irrelevent.
“The prototype software interface developed and established by our specialists for the system of data and management commands voice input is intended for mass mobile electronic devices, says the project manager, Vyacheslav Anciperov.
Perhaps, the most important and fundamentally new aspect of this work is that they have managed to single out essential elements of speech being guided by the notion of hierarchical structure of speech. Like in a musical composition, one can recognize more or less high levels of organization – rhythm, main theme, arrangement, so they have also learned to single out the ranges in the speech flow (i.e. in the wide frequency spectrum), which carry the major semantic loading. It has turned out that this is a very small part of human speech sounds – only up to 1 KHz. “Thus we simplified the task for the computer to the maximum. And one more thing – we have taught the computer to recognize individual sounds, which is sometimes far from easy. As a result, our system wins in processing speed and in processor time and memory consumption as compared to those of all known similar systems. This is the path to efficient speech processors that nobody has passed yet,” he adds.
SOURCE: informnauka.ru press release