Predicting and Avoiding Dog Barking Behaviour through Deep Learning

Excessive and intense barking can be a problem for dog owners and households more generally. Dogs can react to (be triggered by) a wide range of environmental noises and situations, e.g., the sound of the postman, the doorbell amongst a whole range of other sounds. In many situations this behaviour results in reprimands and/or punishments, e.g., demands for the dog to be quiet. Veterinary scientists and dog behaviour experts increasingly recognise that dogs respond more positively to positive reinforcement. Thus, instead of shouting at the dog, if the sounds can result in treats being given, then it is possible to change the dog's behaviour, i.e., the dog recognises the sound as a positive and not a negative thing hence it does not bark. Modifying a dog's behaviour when the owner is present is possible, but many pets are left alone for periods of time hence a solution is needed to automate the positive reinforcement. This is the focus of this paper. We demonstrate how rich and diverse urban audio datasets can be used as the basis for automatically capturing Mel-Frequency Cepstral Coefficients and Mel Spectrogram features. These are then used in conjunction with Convolutional Neural Networks to learn and hence identify and predict environmental sounds that precede a dog's barking. We achieve an overall accuracy of up to 87.6% based on 50 representative environmental sound classes.