Abstract: Acoustic signals emitted by bearing contain lots of important information about its operation status.Compared with vibration signals，acoustic signals can be collected with non-contact sensors，so as to be convenient and cheap.By abstracting the mel-frequency-cepstrum-coefficients（MFCC） from acoustic signals of engine bearing as the characteristic parameters，and using Hidden Markov Model（HMM） with the good performance of recognizing and anti-noise to analyze the operation status，this paper presents a new fault diagnosis method by acoustic signals based on HMM.In order to achieve diagnosis of bearing fault in real-time，this paper adopts discrete HMM（DHMM） with low computational complexity to fasten the speed of training and identifying.Experiments results prove that，with an average recognition rate for all bearing operating status of near 90% and recognition time of about 31ms，the presented method is effective and feasible and has a great prospect.