On separating glottal source and vocal tract information in telephony speaker verification