Towards robust audio spoofing detection: a detailed comparison of traditional and learned features