Even at the beginning of the corona pandemic, researchers suspected that the Sars-CoV-2 coronavirus would be betrayed by the voice – because since the Covid-19 disease caused by the virus is primarily a respiratory disease, the voice must also to be affected. Experienced doctors can often use their voice, breath, and cough to draw conclusions about a disease, such as whooping cough or asthma. With many diseases, however, voice changes are barely audible – it is beyond human ability to perceive them.
Artificial intelligence recognizes voice characteristics of a disease
Computers have a number of advantages over humans: They are undividedly attentive, never tire and can learn from an infinite number of examples. In order to be able to detect Covid-19 with artificial intelligence, the researchers first need voice recordings from people who have been infected with the virus. The computers learn through countless examples to recognize the voice characteristics of the disease. At the beginning of the pandemic, the scientists used voice recordings from Wuhan – the fact that these were in Chinese did not bother, because clear patterns can be recognized in different languages that can be simplified. At some point the computer will recognize these patterns itself.
Computer discovers patterns
The computer brain is fed with as many voice recordings as possible. During the learning phase, he receives information with every audio sample whether he has Covid-19 disease or not. The computer breaks down the voice recordings into curves, bits and columns of numbers. Through the comparison, his neural network recognizes patterns that speak for or against a disease. With each example, the computer becomes smarter – until at some point it no longer discovers any additional differences in the patterns. Then the learning curve flattens out and the system has finished learning.
Voice samples from Germany
The Augsburg researchers are now being supported by the local university hospital and some corona test centers in collecting the voice samples. After the usual PCR or rapid test, people there are asked to provide an additional voice sample for research.
The app’s hit rate increases
The computers continue to learn through the numerous recordings, while the researchers are simultaneously testing how well their prototype app is already working. According to the researchers, the hit rate is already 85 percent, and the trend is rising.