Endophasia: Utilizing Acoustic-Based Images for Dropping Contact-Free Silent Speech Commands

• Main Authors: Wei-Hsiang Huang, 黃威翔
• Other Authors: Chih-Yuan Yao
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/u533m3

碩士 === 國立臺灣科技大學 === 資訊工程系 === 107 === Using silent speech to drop commands has received growing attention, as users can utilize existing command set from voicebased interface without evoking other people's attention. Such interaction keeps the privacy and social acceptance from others. However, current solutions for recognizing silent speech mainly relies on vision-based data or attaching the microphone on the throat. These solutions are either power-consuming and have potential privacy issues. In this paper, we propose a sensing technique that only needs a pair of microphone and speaker, which not only consumes only few powers but also have less privacy concerns. We chose 10 commands for experimentation and used the built-in speaker and microphone to detect the mouth movement of the user in front of the phone. Through the deep learning of CNN and the command data collection from 15 users, we trained within-user and cross-user model, and through the accuracy of 81.41\% and 81.80\%, the feasibility of using ultrasonic technology to drop silently command to the mobile phone is verified.