Sound field estimation using deep unsupervised learning for sound field reproduction

Many technologies of sound field reproduction have been proposed with multiple loudspeakers to physically recreate a target sound field. Pressure matching (PM) method, one such reproduction technique, requires dense arrangement of control points. Consequently, it requires many transfer function measurements between control points and loudspeakers, as well as recording of the original sound field using a large number of microphones. Previous studies have addressed this issue by employing data-driven deep learning to estimate multi-point transfer functions from a limited number of measurements, successfully achieving accurate reproduction. However, these approaches were constrained to experimental settings with point sources producing only direct sound. When considering reflected sounds and other acoustic phenomena, the amount of required training data becomes enormous, limiting their feasibility for real-world applications. In this study, we propose a sound field estimation method based on a deep unsupervised learning network. By learning latent structures and patterns within the measured signals themselves, the model can estimate sound pressure signals at positions different from the original measurement points. In simulation experiments, we evaluate the estimation accuracy by comparing the proposed method with conventional approaches. Work partially supported by Research Institute for Science and Technology of Tokyo Denki University Grant No. Q24J-04/Japan.