Multi-stream Convolutional Neural Network with Frequency Selection for Robust Speaker Verification