Interpretable Spectrum Transformation Attacks to Speaker Recognition