This study compares the performance of gear fault detection using Artificial Neural Networks (ANNs) and Adaptive Neuro-Fuzzy Inference System (ANFIS). The feature vector is extracted from vibration signals by standard deviation of wavelet packet coefficients at the fourth level. An improved distance evaluation technique is proposed for feature selection, and with it, the four salient features are selected from the original feature set. These features are normalized in the range zero to one and then fed into ANN and ANFIS. The gear conditions were considered to be healthy gearbox, slightly worn, and medium worn and broken-teeth gears faults. The output layer of ANN and ANFIS consists of one node indicating the status of the gearbox by four labels. A 3-layer Multi-Layer Perceptron (MLP) neural-network and an ANFIS were designed to carry out the task. The results show that the classification accuracy of ANFIS is better than MLP. Also, the effectiveness of the proposed feature selection method is demonstrated by the test results.