Bending fatigue crack is a dangerous and insidious mode of failure in gears. As it produces no debris in its early stages, it gives little warning during its progression, and usually results in either immediate loss of serviceability or greatly reduced power transmitting capacity. This paper presents the applications of vibration-based techniques (i.e. conventional time and frequency domain analysis, cepstrum, and continuous wavelet transform) to real gear vibrations in the early detection, diagnosis and advancement monitoring of a real tooth fatigue crack and compares their detection and diagnostic capabilities on the basis of experimental results. Gear fatigue damage is achieved under heavy-loading conditions and the gearbox is allowed to run until the gears suffer badly from complete tooth breakage. It has been found that the initiation and progression of fatigue crack cannot be easily detected by conventional time and frequency domain approaches until the fault is significantly developed. On the contrary, the wavelet transform is quite sensitive to any change in gear vibration and reveals fault features earlier than other methods considered.