In this paper we introduce a new bearing fault detection and diagnosis scheme based on hidden Markov modeling (HMM) of vibration signals. Features extracted from amplitude demodulated vibration signals from both normal and faulty bearings were used to train HMMs to represent various bearing conditions. The features were based on the reflection coefficients of the polynomial transfer function of an autoregressive model of the vibration signals. Faults can be detected online by monitoring the probabilities of the pretrained HMM for the normal case given the features extracted from the vibration signals. The new technique also allows for diagnosis of the type of bearing fault by selecting the HMM with the highest probability. The new scheme was also adapted to diagnose multiple bearing faults. In this adapted scheme, features were based on the selected node energies of a wavelet packet decomposition of the vibration signal. For each fault, a different set of nodes, which correlates with the fault, is chosen. Both schemes were tested with experimental data collected from an accelerometer measuring the vibration from the drive-end ball bearing of an induction motor (Reliance Electric 2 HP IQPreAlert) driven mechanical system and have proven to be very accurate.