Isolated Word Speech Recognition based on BNN and Its Hardware Implementation

Abstract

In this paper, a binary convolution neural network (BNN) is proposed to realize isolated word speech recognition task, which greatly reduces the model training parameters and training time. For isolated word data sets, the rectangular convolution kernel is designed to replace the traditional square convolution kernel, and batch normalization layer is integrated into the convolution layer to realize the lossless acceleration of the inference process. The binary convolution neural network is deployed on FPGA to realize the edge calculation.