Addressing Feature Imbalance in Sound Source Separation