Carnegie University - Mellon Develops COVID-19 Patient Voice Analysis Application

A team of researchers from Carnegie Mellon University is working on a project that they said will help diagnose coronavirus without routine tests. We are talking about identifying certain characteristics of the voice of the sick person - if the disease has affected the lungs strongly, then it can be determined by certain parameters of the voice.

Benjamin Streiner, a graduate student at the university who came up with the idea, thinks that the need for inexpensive and effective tests for the presence of the disease is long overdue. But what is now is expensive and not very reliable. Of course, diagnosis of a disease by voice cannot be called a 100% reliable method, but now we are talking about analyzing the very possibility of identifying a disease by voice recording.

The developers are not talking about any serious results. An application needs a microphone on a computer or smartphone to record voice. For voice analysis, a neural network developed by university specialists is used. It is reported that they analyzed the voices of patients with coronavirus (not all in a row, but only those who had lung complications).

“We are trying to develop a voice service that, as shown by preliminary experiments and our expertise, can show a fairly accurate result. True, the final is still far away, ”the developers said.

The service “listens” to a person’s voice, analyzes and sets a point. The higher the score, the higher the likelihood that a person is sick. According to scientists, the number of points is an indicator of the coincidence of the signatures of the subject's voice with the signatures of the voices of patients with coronavirus patients. Now experts are collecting a large database of voice recordings that can be used to train a neural network.

For a short time, the service was open to everyone, now the opportunity to try the service in operation has been closed - probably because the number of voice records needed to train the neural network has reached the desired value .



All that is needed to work with the service is to cough several times, pronounce the alphabet and perform a couple of other simple actions. According to some doctors, coughing patients with complications from coronavirus is not like coughing caused by other reasons. Of course, the difference is not very noticeable by ear, but machine learning makes it possible to catch the difference and fix it.

To collect primary data, scientists asked for help from colleagues from around the world. Doctors, with the consent of their patients, recorded their voices, cough, and sent an audio recording to the university. Further, the records were used to train the neural network and configure it to identify diseased patients.

The developers do not claim that their service will be 100% working. As mentioned above, tests are currently underway. The results are pretty encouraging, but to improve the accuracy of the diagnosis, work needs to be continued for some more time.

Now it is planned to work on cases of positive operation of the algorithm, which in fact turned out to be false. Scientists will study all these cases and try to understand why the algorithm did not work correctly.