Monitoring health of dynamic systems includes using speaker recognition techniques. Some embodiments include determining a system-independent statistical (first) model, and determining a healthy system (second) model based on data representing vibrations of multiple healthy systems and the first model and speaker recognition techniques. Vibration data are obtained from a particular system. It is determined whether the particular system is unhealthy based on the vibration data from the particular system and the first model and the second model and speaker recognition techniques. Some embodiments include obtaining training data that represents vibrations of multiple healthy systems. A damage-sensitive parameter is based on the training data. A threshold value that separates damaged systems from healthy systems is based on the training data and the parameter. It is determined whether a particular system is healthy based on the threshold value and a value for the parameter for vibration data from the particular system.