Sberbank learned to identify covid-infected people by voice, breathing and cough

According to the first deputy chairman of the bank's board, Alexander Vedyakhin, quoted by Interfax, the likelihood of infection is determined based on the results of a short survey of symptoms, as well as three sound patterns - voice, breathing and coughing.

The sound files of the test subjects are converted into a spectrogram, and it shows the distribution of signal power at different frequencies. After that, the spectrogram is analyzed by a deep convolutional neural network, which was trained on the basis of over a thousand samples of breath and cough sounds collected from patients in Russian clinics.

It is indicated that the area under the "error curve" of the created model is now equated to 0.8. However, Sberbank suggests that the quality of the model will improve due to the growth in the amount of data that is also collected by the mobile application. It will soon appear in the App Store and Google Play.

Sberbank experts decided to create the algorithm in November 2020.

- Of course, our model does not yet achieve the accuracy of biological PCR, which is not surprising, but now it has comparable characteristics. At the same time, it makes it possible to make customizable sensitivity, much easier to use, more convenient and, importantly, cheaper. This is not a medical diagnostic tool, but rather a personal daily checker - taking the test and getting the result takes only 60 seconds! - said Vedyakhin.

Recall that in Japan, back in July last year, they presented a mask that does not complicate communication between people, like ordinary medical masks, but vice versa. The high-tech mask is called the C-Face and costs 4,400 yen (about $ 40). The mask translates speech from voice into text, and at the same time is capable of translating it into eight languages.