This paper addresses the problem of automatic discrimination of rhythms in ECG signals. In particular we attempt at discriminating normal sinus rhythm (NSR), ventricular tachycardia (VT) and ventricular fibrillation (VF). The solution to this problem plays an important role in determining whether or not a subject should be given an electroshock. In performing the discrimination we extract features, computed from segments of the ECG, that are used as inputs to the learning vector quantization (LVQ) classification algorithm, possibly with a modified distance measure. In this paper we show the performance of three different features for this purpose: the pitch period, a pulse train extracted from a regular-pulse excitation (RPE) model of the ECG signal, and, finally, autoregressive (AR) parameters extracted from the signal. After presenting classification results, we also comment on some fundamental problems in testing such classification algorithms.