Stethoscope-Sensed Speech and Breath-Sounds for Person Identification With Sparse Training Data

A novel person identification (PID) technique is developed in this study, which exploits a new biometric called bronchial breath sound and speech signal acquired by a stethoscope. In addition to investigating the acoustic characteristics of breath sounds for PID, we evaluate three identification methods, including support vector machines (SVM), artificial neural networks (ANN), and i-vector approach. Recognizing the requirement that the amount of sound data collected from each person should be as small as possible, this work studies data augmentation (DA) techniques that avoid the system training process from the overfitting problem when the training sound data is insufficient. In addition, we apply feature engineering techniques to find the informative subset of breath sound features which is beneficial for PID. Our experiments were conducted using a dataset composed of 16 subjects, including an equal number of male and female participants. In the test phase, both Support Vector Machine combined with feature selection and Artificial Neural Networks approaches yielded the promising accuracies of 98%.