Embedded System Implementation of Wake-Up Word Recognition

• Main Authors: Maystya Tri Handono, 韓多諾
• Other Authors: Jia-Ching Wang
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/8zsb73

碩士 === 國立中央大學 === 資訊工程學系 === 106 === Wake-up word system is used to put an intelligent device in a state of alert so that it expects further spoken commands. The wake-up word system allows for hands-free operation of devices such as smart phones, multimedia systems in cars or home automation system. In this thesis, we implement the wake-up word system using deep learning architecture. The system is implemented using Tensorflow framework for training. In the testing, we implements the inference without Tensorflow framework. The reason is to give some embedded system device that has not support with Tensorflow. The wake-up word system will be evaluated using False Alarm Rate (FAR), False Rejection Rate (FRR), and the accuracy. The baseline has 95% of accuracy, 0.09% of FAR, and 0.02% of FRR. The testing result with Tensorflow framework has 0.03% FAR, 0.07% FRR, and the accuracy is 96%. The testing inference without tensorflow resulting 0.06% FAR, 0.11% FRR and the accuracy is 95.8%. The different between the two implementation is around 3%.