Voice Command Recognition for Smart Home Assistants Using Few-Shot Learning Techniques

Abstract

Voice command recognition is an essential part of intelligent and personalized development of the smart home assistants. Nevertheless, conventional deep learning techniques demand intensive labeled data and computational resources, and thus limiting the applicability in situations where commands are rare or personalized. This paper proposes a new few-shot learning framework designed for practical voice command recognition of smart homes. With the help of prototypical networks and data augmentation algorithms, our approach deals effectively with a small number of training examples. We test our model on data from the Google Speech Commands and a custom command collection for smart home applications. Results achieve Top-1 accuracy of 89.3% using just five examples per class besting baseline convolutional neural networks and other few-shot variants. On embedded systems, our framework can be deployed in a real-time manner, owing to the low latency it entails. This research offers a fresh opportunity to establish tree-specific voice control in smart houses with little training of the human user.