Diagnosis of laryngeal cancer is invasive and expensive. What if we could use speech to find patients at high risk for the disease? This would help doctors focus on those who need urgent attention.

In a perfect environment, researchers can accurately detect disease from speech. However, most people don’t have recording studios in their own homes! We came up with a way to test how well a computer program can detect disease from speech when there’s background noise. We did experiments using this method with different computer programs and noise reduction techniques. Our best program was accurate 81.2% of the time when the speech was clear. But when there was noise, the accuracy dropped to 63.8%. While noise reduction techniques did improve this it still could not perform as well as it did on recording studio data.

Dealing with noise is tricky, and we need to solve this issue before these systems can be used in real medical practice. We showed that our method lets us test how well these computer programs work with noise.

This is available as a full publication