Automatic recognition of Japanese voiceless fricative consonants [s], [&Imoust;] and [h] is discussed terms of acoustic measurements, and an effective algorithm for the recognition is presented. First, discrimination of voiceless fricatives from other phoneme groups, in particular voiceless stops and affricates, is investigated. It is shown that the voiceless fricatives in isolated CV's can be separated from the stops and affricates by the rising patterns of the second and third local peak amplitudes and voiceless interval length. Next, classification within the voiceless fricatives is investigated using power spectral parameters obtained from isolated CV's uttered by five speakers. Effective parameters are the frequency of the center of gravity of the power spectrum, the second local valley frequency, and the first and second local peak frequencies. Based on their distributions, a classification algorithm is constructed and its feasibility is confirmed by recognition experiments. Voiceless fricatives in CV's and VCV's uttered by three speakers chosen from the five speakers used for training are correctly recognized at rates higher than 95%. Also, those uttered by four unknown speakers are recognized with no deterioration.