As the battle against COVID-19 wages on around the world, scientists and researchers are turning to artificial intelligence (AI) to find new ways to detect, fight and treat the virus afflicting most of the planet.
Although there have been a lot of disappointing results from AI-driven clinical initiatives, one solution that now offers glimmers of hope is the use of speech-recognition technology for screening and detection of those likely positive for the virus.
Last month a team of researchers from Brazilian AI giant CyberLabs Group participated in the paralinguistics challenge at INTERSPEECH 2021, where they took home top honors for their promising new algorithm that is capable of detecting COVID-19 infections by analyzing audio clips of coughing and speech, based on early testing.
Using an advanced AI-based voice-recognition methodology, research teams analyzed several hours of audio recordings of both healthy and sick individuals to identify patterns that are consistent with COVID-19 infections. After teaching their deep-learning algorithm to recognize COVID-related sound patterns, CyberLabs was able to identify three out of four sick people correctly (75.9% of the time) when using audio recordings of their coughing, and 70.3% of the time when relying on recordings of speech.
While similar speech-recognition research was already being conducted before the INTERSPEECH 2021 event, participating in the paralinguistics challenge gave CyberLabs and other research teams a valuable opportunity to test their methodologies and advance them further, said Anderson Soares, chief scientist on the CyberLabs research team and general manager of Centro de Excelência em Inteligência Artificial (CEIA) & Deep Learning Brasil.
“This is the first time that the results reached an acceptable level and showed a viability level for possible products,” he added.
More than two million audio samples in 527 categories were provided to teams participating in the paralinguistics challenge. These recordings were purposefully collected across multiple countries and from people who speak different languages so researchers could test their theories and solutions using the same robust data set. Having access to this diverse audio library improved the feasibility of solving the problem through AI, and greatly increased the confidence in results, said Lucas Gris, another CyberLabs research team member.
“[What set our work apart] is in the compilation of various research and innovation efforts that involve speech analysis using neural networks. Deep down it’s about pattern recognition in someone’s voice. We were able to take advantage of several AI architectures used in other tasks related to voice and adapt them to this specific challenge,” Gris said.
CyberLabs Group is the largest artificial intelligence and cybersecurity company in Latin America, with nearly 200 researchers and engineers spread across Brazil and the U.S. It’s also one of the biggest champions of AI research at the university level. The company’s top AI research center is led by the Federal University of Goiás in midwestern Brazil, and CyberLabs supports additional researchers and students at the Federal Technological University of Paraná and University of São Paulo.
So what’s the next step for CyberLabs’ COVID-detecting AI algorithm?
“Once the technological feasibility of this type of solution has been demonstrated, it is possible to advance the development of products for health or beyond. The world has experienced advances in the digital transformation of healthcare, including remote care and AI-based solutions that can be used as a screening method. We will assess where this type of technology makes sense and look for partners for evolutions,” said Edresson Casanova, CyberLabs research team member.