The Covid-19 pandemic and the recent outbreaks of influenza have seen many infected people not showing symptoms associated with cold. They needed blood tests to confirm the infection. Soon pathbreaking research at Sardar Vallabhbhai National Institute of Technology in Surat might change that. The researchers have been able to recognise from a person's voice if she has a cold or not.
Surat-based researcher Professor Suman Deb and students Pankaj Warule and Siba Mishra collaborated with Jarek Karjewski of RFH Koln in Germany to develop an algorithm that can distinguish between the voice of a person with cold and one without.
The AI model was fed voice recordings of people with cold and those without, provided by Karjewski. Those who participated for the study were asked to read short stories like 'The North Wind and the Sun' and the German passage ‘Die Ostergeschichte’.
Deb said they achieved 70 per cent accuracy, and the model would detect voices in Indian languages as well. It will initially be available for English and Hindi speakers. Other Indian languages are being developed.
Warule said, after recording the voice, a harmonic peak extraction was done by calculating the Fourier transform of speech, followed by NHPF, NHPM and SHPR feature extraction. NHPF and NHPM stand for normalised harmonic peak with respect to the first harmonic peak and normalised harmonic peak with respect to the maximum value of the harmonic peak, and SHPR is successive harmonic peak ratio. Then comes the prediction of cold or healthy speech using machine learning or deep learning classifier. This is done using python (a computer language) and an AI model, and it involves mathematical and signal processing calculation. The speech spectrum amplitude and frequency of a healthy person's speech, and that of a person with cold, are different.
The Indian researchers had begun work in 2015 and published a paper in 2019. The work then continued and their latest study, 'Sinusoidal model-based diagnosis of the common cold from the speech signal', has been published in the journal Biomedical Signal Processing and Control.
At the SVNIT, Surat, the researchers are building a sound-proof room. Deb said a sound-proof room would increase the accuracy to 80 per cent to 90 per cent, and help deduce which parameters were affected by noise and to what extent. They are also trying to develop a mobile app, which would be a signal-based non-invasive diagnostic technique that can work remotely.
The use of the technology would mostly be for remote areas that do not have easy access to doctors and labs, said Deb. In future, we might also be able to detect lung and heart disease from a person’s speech.
Deb has already found a common application for the new technology. On a lighter note, he said it would be the quickest and easiest way to catch those who call in sick faking cold.