Fusion of deep shallow features and models for speaker recognition