Neural networks have become an attractive choice for audio applications, but they are known to suffer from adversarial examples. In this work, we propose a method for detecting adversarial examples of a keyword spotting system. We design convolutional neural networks for metric learning to map the internal representation of each layer of an input audio to a low-dimensional feature space. We then extract the distance information from the feature space of each layer and feed it into an LSTM network to determine whether the input audio is clean or adversarial. Promising experimental results are shown to support our detector.