A Speech-Noise-Equilibrium Loss Function for Deep Learning-Based Speech Enhancement