Researchers from the University of Carnegie - Mellon taught the neural network to recognize people by sighing between words. The researchers published their research on arXiv.org. The work used a neural network with a long short-term memory. It was downloaded more than 100 hours of news in English. Based on this information, the neural network was trained to recognize speech, and then separate phonemes.

Later, scientists separated from the general flow of sounds between words. One of the authors of the study Rita Singh explained that sighs differ from other speech components in that they are much more difficult to control and much more difficult to forge. The use of a neural network for their recognition made it possible to significantly improve the accuracy of human recognition.

Another way of recognizing by sighs is through super-vectors. Scientists have turned to a technique that recognizes feature super-vectors (the technique is used in most voice recognition systems). The method allows you to convert individual fragments of speech into vectors. Among them, regions with condensations (super-vectors) are sought. These data are then analyzed using classifiers.

In the course of the experiment, scientists have established that identity recognition with the help of supervectors is not as accurate as using a neural network. The result for the first method is 72-74%, and for the second method - more than 91%.