In this paper, we propose two versions of the HMM-trellis calculation algorithms, STA-I and STA-II, suitable for continuous speech recognition systems. The Segmented Trellis Algorithms make use of the HMM's property that each state models a particular interval of the acoustic signal. The STA-I was implemented in our system, SUSKIT-IL It achieved three times faster processing speed without decreasing the average sentence recognition rate.