Daily Stress Detection Using Artificial Neural Network Based on Acoustic and Semantic Information from Speech

Cumulative daily stress is harmful to the health of people and leads to productivity loss. Hence, timely detection of daily stress is of vital importance. Natural speech from real life is the recommended information to detect stress as a non-invasive way. This study aims to improve stress detection accuracy by combing the acoustic and semantic information from speech. Based on the speech database with real daily stress, we fused the acoustic and semantic features and developed a daily stress detection model using artificial neural network. The results showed that the model accuracy using acoustic information is 65.50% with a F1-score of 60.21%. The model accuracy using semantic information is 80.00% with a F1-score of 76.65%. By combining the acoustic information and semantic information, the model accuracy was improved to 90.75% with a F1-score of 89.25%. These results indicated the complementary effect of acoustic and semantic information on the daily stress detection. This study validated the effectiveness of detecting daily stress based on the combination of acoustic and semantic information from real speech. The model developed in this study can be applied to daily stress monitoring in daily life, offering valuable insights for stress management intervention to mitigate adverse health impacts.