Deep Learning-Based Non-synchronous Measurement for Broadband Sound Source Localization

Array-based acoustic beamforming methods are prevalently used in many practical fields. However, limited by the size of the array and sensor elements, these methods cannot achieve the best performance. To mitigate the limitations, Non-synchronous Measurement (NSM), which synthesizes a virtual array with a larger aperture and denser sensor elements by shifting a prototype array to different positions, was proposed to enlarge the working frequency band of the prototype array. Since the prototype microphone array measures the sound signals without time-overlap between different measurements, some phase information of NSM is lost. Thus, the generated cross-spectral matrix (CSM) of NSM is incomplete. Although some rank-minimum CSM completion approaches recover the lost information, they rely on localization algorithms to estimate the positions of the sound sources, which is time-consuming. There-fore, we proposed a computationally friendly NSM based on deep learning to achieve effective broadband-sound source localization (broadband-SSL), which directly estimates the coordinates of the speech source. Some simulations are implemented, and the results validate the efficiency and accuracy of our proposed method.