Deep-Learning-Based Speech Enhancement with Rough-Focused Optical Laser Microphone by Reconstructing Complex Spectrum

Rough-focused recording with an optical laser microphone allows for recording that is wide ranging and robust against changes in the position of a vibrating object. However, the recorded speech suffers from noise due to laser diffusion and missing signal components. To solve this problem, we propose a speech enhancement method for rough-focused recordings that reconstructs a complex spectrum. The conventional speech enhancement method for rough-focused optical laser microphones reconstructs only an amplitude spectrogram without considering phase components, resulting in lower quality speech enhancement. In contrast, the proposed method simultaneously reconstructs both amplitude and phase components of rough-focused recordings by reconstructing a complex spectrum. We compared the method with the conventional method. The results show that the proposed method performed equivalently to or better than the conventional method in objective evaluations.