For use in offline speech processing systems a novel algorithm has been developed, that classifies clean speech segments robustly as voiced, unvoiced, or silence respectively. This decision is needed e.g. in source controlled speech coders which treat voiced, unvoiced, and silent segments differently, to increase the coding efficiency. The classifier is based on a combination of several features, extracted from the speech signal in the time domain. Besides the Energy, a novel measure, representing the unsteadiness of the speech signal, is proposed. Non-realtime coding allowes iterative refinement of the classification, taking the cepstral distance into account. The new classification scheme has been tested with the popular AMR codec.