Entropy Based Feature Pooling in Speech Command Classification

In this research a novel deep learning architecture is proposed for the problem of speech commands recognition. The problem is examined in the context of internet-of-things where most devices have limited resources in terms of computation and memory. The uniqueness of the architecture is that it uses a new feature pooling mechanism, named entropy pooling. In contrast to other pooling operations, which use arbitrary criteria for feature selection, it is based on the principle of maximum entropy. The designated deep neural network shows comparable performance with other state-of-the-art models, while it has less than half the size of them.