Remote Speech Reconstruction Based on Convolutional Neural Network and Laser Speckle Images

• Published in: 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC) pp. 1039 - 1044
• Main Authors: Hao, Xueying, Guo, Lianbo, Zhu, Dali, Wang, Xianlan, Yang, Long, Zeng, Hualin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.10.2022
• Subjects: convolutional neural network; Convolutional neural networks; Laser modes; laser speckle; Real-time systems; Reconstruction algorithms; remote speech reconstruction; Speckle; Speech processing; Training
• ISSN: 2577-1655
• DOI: 10.1109/SMC53654.2022.9945090

Remote speech reconstruction is widely used in counter-terrorism, medical science and engineering. In order to obtain reconstructed speech with high accuracy, we propose a speech reconstruction method. This method consists of two parts. Firstly, some optical devices are used to collect speckle images. Secondly, the convolutional neural network is used to detect the subtle motion of speckles. The results show that the lowest mean absolute error of the sinusoidal signal reconstructed by the method is 0.0489, and the lowest mean absolute error of the real speech is 0.0271. Compared with the convolutional neural network proposed before, the error of reconstructed speech is small, and the number of parameters is significantly reduced, with 0.73M for our model compared to 11.45M for the previous model. Besides, the time cost of training on some datasets is reduced to less than 1 hour, which is much lower than the previous model. The experimental results prove that our model is a lightweight, high-accuracy model for remote speech reconstruction with short training time.