Design of real and complex recurrent neural networks for sound source localisation

This study explores a pruning technique based on the singular value decomposition for both real- and complex-valued recurrent neural networks (RNNs). Building on prior work on multilayer perceptrons (MLPs), the method reduces hidden-layer dimensions during training to reduce computation with minimal performance loss. Challenges unique to RNNs are addressed and compared to those of MLPs. Pruned models are compared with original architectures and a benchmark pruning method in a sound source localisation task. Results show real- and complex-valued RNNs can retain strong performance while using substantially fewer computational resources.