アクティブ音響センシングを用いた無発声母音認識

Silent speech recognition is suitable for use in public spaces due to privacy and security considerations compared to traditional speech recognition. Various silent speech recognition methods have been proposed, including those using cameras, inertial sensors, and electromyography. However, these methods often face challenges such as occlusion issues or the sensors being large and conspicuous. Therefore, this study focuses on active acoustic sensing using small and lightweight piezoelectric elements. Active acoustic sensing is a sensing method that uses the frequency transfer characteristics of objects. Two piezoelectric elements are used as a speaker and a microphone. An ultrasonic swept sine signal is emitted from the speaker to vibrate the object, and the microphone measures the propagated vibration. This allows for the calculation of frequency transfer characteristics based on the material and shape of the object. This study proposes a silent speech recognition technique using active acoustic sensing. Using the prototype of the proposed method, I conducted experiments with four volunteers, including the author, to investigate the optimal fixed position of the piezoelectric elements. Next, using the improved prototype, I recognized six mouth shapes, including closed-mouth formations and the five Japanese vowels (’a’, ’i’, ’u’, ’e’, ’o’). I achieved an accuracy of 88.4% when I conducted experiments with seven volunteers. Finally, for noise reduction in signals, I created an attachment for the piezoelectric element using Hitohada gel, aiming to apply the proposed method to wearable devices. The experiments revealed that the positional errors of the piezoelectric elements significantly affect the recognition accuracy. Therefore, I will create a program to calibrate positional errors. Additionally, I will develop a music playback application using the proposed method.