A novel automatic method is introduced for classifying speech segments into broad phonetic categories using one or more hidden Markov models (HMMs) on long speech utterances. The general method is based on prior analysis of the acoustic features of speech and the properties of HMMs. Three example algorithms are implemented and applied to voiced-unvoiced-silence classification. The main advantages of the approach are that it does not require a separate training phase or training data, is adaptive, and that the classification results are automatically smoothed because of the Markov assumption of successive phonetic events. The method is especially applicable to speech recognition.