Recent studies have demonstrated that combinations of logarithmic durations can classify duration-sensitive phonemes, such as Japanese singleton and geminate consonants, at various speaking rates. The acoustic features of the Japanese fricative /s/ and affricate /ts/ are related to duration; therefore, a combination of logarithmic durations can likely classify these consonants. To examine this possibility, discriminant models using linear and logarithmic durations, with and without a speaking-rate-related variable, were compared in terms of their performance in classifying /s/ and /ts/ at word-initial position at various speaking rates. The results indicate that the discriminant model using logarithmic duration with a speaking-rate-related variable can classify the consonants better than the other models, indicating the importance of logarithmic duration. The results are considered in the framework of logarithmic information processing in the brain.