Evolving Multi-Resolution Pooling CNN for Monaural Singing Voice Separation

Monaural singing voice separation (MSVS) is a challenging task and has been extensively studied. Deep neural networks (DNNs) are current state-of-the-art methods for MSVS. However, they often designed manually, which time-consuming error-prone. They also pre-defined, thus cannot adapt their structures to the training data. To address these issues, we first multi-resolution convolutional network (CNN) MSVS called pooling CNN (MRP-CNN), uses various-sized operators extract features. We then introduced Neural Architecture Search (NAS) extend MRP-CNN evolving (E-MRP-CNN) automatically search effective using genetic algorithms optimized in terms of single objective taking into account only performance multiple objectives both model complexity. The E-MRP-CNN multi-objective algorithm gives set Pareto-optimal solutions, each providing trade-off between Evaluations on MIR-1 K, DSD100, MUSDB18 datasets were used demonstrate advantages over several recent baselines.