DIFFERENTIABLE ARCHITECTURE SEARCH FOR KEYWORD SPOTTING

Abstract

With the popularity of smart mobile devices, the need to control mobile devices using voice is increasing. Also, there is greater expectation for the accuracy of keyword spotting. Many existing researches have applied neural networks to keyword spotting, and have great performances. However, at the same time, for keyword spotting on mobile devices, there is still a need to further reduce the parameter size and improve the recognition accuracy simultaneously. In this thesis, I apply the neural network architecture search approach to keyword spotting, and proposed a Differentiable Architecture Search Approach for keyword spotting. This approach can design multiple neural network models for keyword spotting through search. In this thesis, I proposed eight specific neural network models designed by this approach. All models beat the state-of-the-art model based on the evaluation on Google Commands Dataset, with similar or much smaller parameter sizes.