A team of researchers in artificial intelligence has built a mobile app capable of spotting covid-19 infections from a person’s voice. The approach aims to be more sensitive and faster than the antigen tests currently in use.
The system, set to be showcased at the European Respiratory Society International Congress in Barcelona, promises a price advantage over antigen testing. This could enable deployment in low-income countries where tests remain costly or hard to obtain.
Maastricht University’s Institute for Data Science in the Netherlands notes an accuracy rate around 89 percent for the AI model, though the exact figure can vary with the specific test brand.
Project leaders say the results indicate audio data and carefully tuned AI algorithms can reliably identify covid-19 infections.
Fast and virtual response
The method provides a free, easily interpreted assessment. It relies on remote, virtual testing with response times under a minute, making it suitable for high-traffic entry points at events to quickly identify positive cases in the population.
Covid-19 infections typically impact the upper airways and vocal cords, leading to detectable changes in a person’s voice.
To explore feasibility, researchers including a Maastricht University Medical Center pulmonologist and a data science colleague at Maastricht collaborated with a Cambridge University open-source app named Covid-19 Sounds. They analyzed a database of 893 audio samples submitted by 4,352 participants, of whom 308 tested positive for covid-19.
The app runs on a user’s mobile device, collects basic demographic and medical-history information, and records breathing, coughing, and voice samples. Researchers then apply spectrogram analysis to the audio, describing loudness, power, and variation to extract voice features that may signal infection.
AI models were created and tested to determine which configuration best distinguished covid-19 voices from those of healthy individuals.
89% accuracy
The Long Short-Term Memory (LSTM) neural network model demonstrated an 89 percent accuracy in detecting positive cases and an 83 percent accuracy in correctly identifying negatives.
These findings are part of an ongoing larger study that will include more than 50,000 audio samples from tens of thousands of participants to strengthen the evidence base.
In another line of work, Henry Glyde from the University of Bristol reported that an AI system, delivered through the app myCOPD, can forecast acute exacerbations in people living with chronic obstructive pulmonary disease (COPD).
MyCOPD is an interactive platform developed with input from patients and clinicians and supported by the UK National Health Service since 2016, now helping thousands manage the condition. A dataset comprising hundreds of patient records collected between 2017 and 2021 was used to train the latest AI models.
The most recent AI model reported a sensitivity of about 32 percent and a specificity of roughly 95 percent. This means it is particularly effective at ruling out exacerbations when the model indicates none, potentially helping to avoid unnecessary treatments and interventions.
Overall, researchers emphasize that AI-driven voice analysis offers a promising complement to existing testing methods, with ongoing work aimed at validating results across broader populations and real-world settings, including repeated monitoring and integration with healthcare pathways.