Stress detection just by listening to your voice...Development of voice-based AI measurement model
Mar 06, 2025
A research team led by Professor Kim Jung-hyun of the Department of Mental Health at Seoul National University Bundang Hospital (joint research by Bundang Seoul National University Hospital and Seoul National University's New Media and Communication Research Institute) developed a stress measurement technology through voice analysis. There is a way to measure and manage stress in real time through personal digital devices such as smartphones.
Proper stress has a positive effect on increasing energy and concentration, but chronic stress can cause health problems such as mental illness, cardiovascular disease, and cancer. Existing stress measurements have relied on subjective surveys or hormone tests, but the research team developed a deep learning-based model that detects stress conditions using non-verbal voice biomarkers and verified their effectiveness using Korean data, noting that stress can be detected by analyzing the effects of muscle tension and respiratory changes on voice tone.
The research team induced a stress state with the SECPT technique, which allows 115 healthy office workers to undergo social evaluation while immersing their hands in cold water through a clinical study in Korea. In this process, an artificial intelligence model was developed that can collect voice data before and after stress and closely compare and analyze frequency, speech speed, and speech pattern to predict stress levels.
In this study, we used a high-performance deep learning model ECAPA-TDNN that can accurately analyze the features of different voices for each person to increase accuracy and cross-validation with cortisol testing in parallel to increase research reliability. As a result, the model developed by the research team was found to be able to distinguish stress states with a high accuracy of 70%. In the future, we plan to further improve performance through large-scale datasets.
The stress detection model developed by Professor Kim's team excluded verbal information such as conversation content among voices and analyzed only non-verbal elements such as voice tone. Through this, it was possible to build a universal model that was not affected by education level, cultural background, and growth environment. In addition, all data is processed locally and not transferred to an external server, thereby minimizing the risk of privacy leakage.
Professor Kim Jung-hyun said, "If you can check your stress level periodically on your personal mobile device, it will help you manage your mental health with appropriate measures, such as using mitigation techniques such as deep breathing, meditation, and exercise, or visiting hospitals if necessary."
Meanwhile, the results of this study were published in the latest issue of the Korean Neuropsychiatric Association's academic journal 'Psychiatry Investigation'.
Proper stress has a positive effect on increasing energy and concentration, but chronic stress can cause health problems such as mental illness, cardiovascular disease, and cancer. Existing stress measurements have relied on subjective surveys or hormone tests, but the research team developed a deep learning-based model that detects stress conditions using non-verbal voice biomarkers and verified their effectiveness using Korean data, noting that stress can be detected by analyzing the effects of muscle tension and respiratory changes on voice tone.
The research team induced a stress state with the SECPT technique, which allows 115 healthy office workers to undergo social evaluation while immersing their hands in cold water through a clinical study in Korea. In this process, an artificial intelligence model was developed that can collect voice data before and after stress and closely compare and analyze frequency, speech speed, and speech pattern to predict stress levels.
In this study, we used a high-performance deep learning model ECAPA-TDNN that can accurately analyze the features of different voices for each person to increase accuracy and cross-validation with cortisol testing in parallel to increase research reliability. As a result, the model developed by the research team was found to be able to distinguish stress states with a high accuracy of 70%. In the future, we plan to further improve performance through large-scale datasets.
The stress detection model developed by Professor Kim's team excluded verbal information such as conversation content among voices and analyzed only non-verbal elements such as voice tone. Through this, it was possible to build a universal model that was not affected by education level, cultural background, and growth environment. In addition, all data is processed locally and not transferred to an external server, thereby minimizing the risk of privacy leakage.
Professor Kim Jung-hyun said, "If you can check your stress level periodically on your personal mobile device, it will help you manage your mental health with appropriate measures, such as using mitigation techniques such as deep breathing, meditation, and exercise, or visiting hospitals if necessary."
Meanwhile, the results of this study were published in the latest issue of the Korean Neuropsychiatric Association's academic journal 'Psychiatry Investigation'.
|
This article was translated by Naver AI translator.
