Acoustic emissions from tribological contacts have become an interesting field of science in recent years. This study focuses on predicting the friction power of a given system (lubricated ring-on-disc geometry), independently of the used sliding material and lubricant, from the acoustic emissions emitted from the system. The low-frequency (1 Hz), continuously measured RMS value of the acoustic data is combined with short-duration and high-frequency (850 kHz) signal data in form of the power spectra and hit rate with three prominence levels. The classification system then predicts the friction power of the test system continuously over the whole test time. Prediction is achieved by four different machine learning methods (tree-type, support vector machine, K-nearest-neighbor, neural network) trained with data from 54 ring-on-disc tests with high variation in material and oil combinations. The method allows for the quantifiable and step-free prediction of absolute values of friction power with accuracy of 97.6% on unseen data, with a weighted K-nearest-neighbor classifier, at any point in time during an experiment. The system reacts well to rapid changes in friction conditions due to changes in load and temperature. The study shows the high information degree of acoustic emissions, concerning the actual friction mechanisms occurring and the quantitative, and not only qualitative, information that one can gain about a tribological system by analyzing them.