Drive train system of a wind turbine is often susceptible to failures owing to its exposure to harsh & aggressive environmental conditions. Fault diagnosis at an early stage of failure is becoming a vital task to ensure gearbox reliability. Conventionally, the monitoring of gears using various condition monitoring technologies in conjunction with advanced signal processing constitutes an effective diagnostic scheme. As it is able to deal with non-stationary and multi-component signals, Wavelet transform finds prolific application in various fields. Especially, the implementation of discrete wavelet transform (DWT) and wavelet packet transform (WPT) has become extensive for feature extraction from the signal. The present investigation attempts to compare the feature extraction abilities of the above mentioned approaches in terms of their effectiveness in the fault diagnosis of local gear faults. Experiments are carried out on a laboratory scaled model of a wind turbine gearbox consists of a three-stage spur gear train having a gear ratio of 48:1. Two gear defects namely, tooth root crack and tooth chipping with two different severity levels are seeded. Vibration analysis is done and the acquired vibration signatures are subject to further decomposition using DWT as well as WPT to extract the wavelet coefficients. A plethora of statistical features alongside the energy contained by wavelet coefficients are computed. Decision tree (J48 algorithm) is performed to identify the dominant features among the computed features and these are channeled as input for Support Vector Machine (SVM) algorithm for classification. The investigation results reveal that WPT approach shows improved performance compared to DWT as a feature extraction technique under the experimental conditions.