Three feature extraction methods for a recogniser for telephone speech are compared : LPC-cepstra, RASTA-PLP and LPC with cepstral mean subtraction (CMS). Training of the discrete-density HMM's happened on a database collected over the telephone, while recognition was done on TI46-Word, played back via a hardware telephone line emulator. The robustness against realistic variations in the telephone line transfer function and in the signal-to-noise ratio is tested.