Low Bits: Binary Neural Network for Vad and Wakeup

Abstract

Nowadays, with the development of mobile Internet, more and more applications require small memory footprint, low power and high precision. In this paper, we first apply the state-of-the-art Binarized Neural Network (BNN) for voice activity detection (VAD) and Binarized Weight Network (BWN) architecture for wakeup. Unlike traditional networks, the weights and activations of BNN and weights of BWN are binary during training and real time processing, which drastically reduce memory, real time and energy. Our VAD system can achieve a good recall 91.1% with only 1k (bit) memory and nearly 44k (pJ) energy. Our wakeup system can achieve 90% accuracy and 4/h (four times per hour) false alarms with 230K (bit) memory and 4.99M (pJ) energy. Both of our system can get 32× memory saving and 5× faster on real time processing on GPUs. While during the real time testing, the Batch Normalization layer and posterior handing are too complex to get good performance, so we optimize them to make it possible to get much faster, smaller and low power VAD and wakeup system.