Microphone Array-Based Sound Source Localization Using Convolutional Residual Network

Microphone array-based sound source localization (SSL) is widely used in a variety of occasions such as video conferencing, robotic hearing, speech enhancement, recognition and so on. The traditional SSL methods cannot achieve satisfactory performance adverse noisy reverberant environments. In order to improve performance, novel algorithm using convolutional residual network (CRN) proposed this paper. spatial features including time difference arrivals (TDOAs) between microphone pairs steered response power-phase transform (SRP-PHAT) spectrum are extracted each Gammatone sub-band. different sub-bands with frame combine into feature matrix the input CRN. employ CRN fuse features. Since introduces structure on basis network, it reduce difficulty training procedure accelerate convergence model. A model learned from data various reverberation noise environments establish mapping regularity azimuth. Through simulation verification, compared deep neural can better task, provide generalization capacity untrained reverberation.