Google has quietly released a new artificial intelligence model that can listen to a person’s cough and predict a surprising amount about their health. The model, called Health Acoustic Representations or HeAR, is a foundation model trained on 300 million audio recordings of coughs, breaths, and other bodily sounds. It is designed to help researchers build tools that can screen for diseases simply by analyzing sounds people make.
From three hundred million sounds to a single diagnosis
The HeAR model was trained on a massive dataset of audio clips, each labeled with basic health information. The recordings came from a mix of public datasets and de-identified data collected through Google’s partner projects. By learning patterns across millions of coughs, the model can detect subtle acoustic differences that might indicate tuberculosis, chronic obstructive pulmonary disease, or other respiratory conditions. Google says the model can also identify non-respiratory signals, such as a person’s age and gender, from the way they cough or speak.
Google is not releasing HeAR as a product for consumers or doctors. Instead, it is open-sourcing the model through its research division. The goal is to let academic teams and healthtech startups build their own applications on top of this acoustic foundation. One early partner is Salcit Technologies, an Indian startup that uses smartphone microphones to screen for tuberculosis. Salcit plans to integrate HeAR into its existing app, which already listens to patients’ coughs and flags potential infections.
Why sound matters more than you think
Human ears can pick up certain cough qualities, like wetness or frequency, but we cannot reliably connect those sounds to specific diseases. Machine learning models, however, can identify patterns too subtle for human hearing. A shortness of breath caused by asthma sounds different from a breath halved by fluid in the lungs. Over time, models like HeAR could make these distinctions automatically, turning a standard smartphone into a diagnostic microphone.
This approach is not entirely new. Researchers have used audio AI to detect COVID-19 from coughs, but those models were narrow and trained on limited data. HeAR is a foundation model, meaning it is trained broadly first and then fine-tuned for specific tasks. That flexibility could let a single model screen for multiple conditions without needing separate training for each disease. Google says the model performs consistently across different microphones and background noise levels, which is critical for real-world use in clinics or homes.
The technology does raise privacy questions. A model that can infer health data from a few seconds of audio could be misused if deployed without consent. Google has published a detailed technical paper and is requiring that researchers who use HeAR agree to ethical guidelines. The company says it will not commercialize the model itself and is focused on supporting non-profit and academic research.
For now, the most immediate application is tuberculosis screening in low-resource settings. According to the World Health Organization, tuberculosis is one of the top infectious disease killers worldwide, and many cases go undiagnosed because traditional testing requires lab equipment. A smartphone app that listens for telltale cough signatures could help bridge that diagnostic gap.
Looking ahead, Google sees HeAR as a platform for much broader use cases. The same acoustic analysis could one day monitor asthma patients at home, track recovery after lung surgery, or even screen for conditions like Parkinson’s disease, which can affect voice and breathing patterns. Early detection through sound could reduce hospital visits and give patients a simple, non-invasive way to track their health daily. As the field of acoustic medicine matures, foundation models like HeAR might become the standard front door for a new generation of audio diagnostics.
For more on how AI is reshaping healthcare diagnostics, check out Mylistingo.
