Deep Learning Based Non-synchronous Sequential Measurement For Speech Localization

Acoustic beamforming is a widely used approach for sound field imaging and sound sources localization. The imaging accuracy and the range of workable frequency bandwidth are limited by the microphone density and the size of the microphone array. However, a larger and denser microphone array will lead to an ultra-high cost, thus not applicable in most cases. The NSM technique, which measures the sources at different sequential positions to form a sizeable virtual array, is proposed to overcome this problem. This approach will obtain an incomplete cross-spectral matrix due to the lack of information between different measurements. Thus, the lost information should be completed first before performing the localization. In this paper, two deep learning based methods are proposed to solve this completion problem. The simulation results verified the computation efficiency and accuracy of proposed approaches in sound field imaging.