Application of fault tolerant calculation in small-footprint keyword spotting neural network

EasyChair Preprint 1021

Abstract

In recent years, many applications of voice wake-up technology have entered people’s field of vision. The key technology is Keyword Spotting. The system needs to detect the ambient voice waiting for a wake-up at any time, so it requires a low hardware energy and high recognition accuracy. This paper aims at real-time speech keyword detection applications. Based on Google’s open source speech commands dataset and Librispeech dataset, combined with various fault-tolerant calculations, a deep neural network that suitable for low-power integrated circuits are constructed and trained. The main structure of the network is the Depthwise Convolution Network (DSC). The energy consumption and resource overhead of the model in hardware implementation is reduced by combining various fault-tolerant calculation methods such as approximation addition, quantification, and binarization. The fault tolerance of the model is improved through retraining method. We proved that the fault-tolerant calculation method of quantization with approximation addition has great potential in small-footprint keyword spotting neural network.

Keyphrases: Approximate addition, Depthwise convolution, keyword spotting, quantification

@booklet{EasyChair:1021,
  author    = {Yicheng Lu and Weiwei Shan and Jiaming Xu},
  title     = {Application of fault tolerant calculation in small-footprint keyword spotting neural network},
  howpublished = {EasyChair Preprint 1021},
  year      = {EasyChair, 2019}}