Keyword Spotting with Deep Neural Network on Edge Devices

Abstract

Interconnected devices are becoming attractive so-lutions to integrate physical parameters and make them more accessible for further analysis. The exploding number of sensors are demanding huge bandwidth and computational capabilities in the cloud. Additionally, continuous transmission of information to the remote server would hamper privacy. Deep Neural Network (DNN) is proving to be very effective for cognitive tasks and can easily be implemented on edge devices. In this paper, a DNN model was trained and implemented for Keyword Spotting (KWS) on two types of edge devices: a bare-metal embedded device (microcontroller) and a single board computer (Jetson Nano™) based robotic car. The unnecessary components and noise from audio samples were removed and speech features were extracted using Mel-Frequency Cepstral Coefficient (MFCC). A Depthwise Separable Convolutional Neural Network (DS-CNN) based model was proposed and trained with about 721 thousand trainable parameters. After implementing the DNN on a bare-metal platform, the converted model took only 11.52 Kbyte of RAM and 169.63 Kbyte of flash memory. The model was also utilized to develop a voice-controlled robotic vehicle, Jetbot. Thus, this research demonstrates an efficient implementation of DNN on edge devices in two major platforms.