Systems, methods, and computer-readable storage devices are disclosed for improved audio feature discovery using a neural network. One method including: receiving a trained neural network model, the trained neural network configured to output an audio feature classification of audio data; deconstructing the trained neural network model to generate at least one saliency map, the at least one saliency map providing a successful classification of the audio feature; and extracting at least one visualization of the audio feature the trained neural network model relies on for classification based on the at least one saliency map.