Shiqi Zhang

The amount of information in sound, for not only humans but also other species, is often underestimated, which could be utilized for species classification, inter-specie communication, individual behavior or status analysis, and more. However, creating an easy-to-use and effective system remains challenging due to complex real-world environments, lack of pure sound sources, high expert annotation costs, and extremely unbalanced sampling of rare species. Even powerful backend models struggle to address these issues. One potentially effective approach is to design multiple frontend systems to tackle individual tasks separately, such as DoA estimation, dereverberation, noise reduction, source counting, and separation, before integrating them. Regarding backend models, the pressing issue is efficiently training models with rare and unbalanced sample-label pairs datasets. In my PhD project, I aim to find a robust acoustic representation and a low-cost, adaptable active learning method for unbalanced data, effectively applying them to bioacoustic signal monitoring.