A unified speaker-dependent speech separation and enhancement system based on deep neural networks