Overcoming The Impact of Different Materials on Optical Microphones For Speech Capture Using Deep Learning

This study analyzes changes in vibration characteristics produced by different materials for the same audio signal through acoustic analysis. A deep learning model combined with training data is employed to mitigate problems and reduce the impact of spectral differences in speech signals recorded using laser Doppler vibrometer on different materials and convert them into high-quality speech output. Results demonstrate that (1) different materials cause variability in the vibration signals produced by the same audio signal and (2) deep learning models can address the challenges posed by material variability through appropriate training data. Furthermore, integrating deep learning with suitable training data helps overcome these challenges. This study suggests that the speech signal variability caused by materials in optical microphones is analogous to the channeleffect problem in traditional microphone signal processing. Future studies should reference methods for addressing channel effects in microphones to enhance the efficiency of speech-capture technology in optical microphones.