This paper presents a new concept for a demisyllable-based speech recognition system for German where final consonant clusters are subdivided into rudiments and suffixes. These units are represented by Semicontinuous HMMs. All equations for modeling the time duration with Gaussian or Gamma distributions are given. An economic solution applies simple "post-processing" by a temporal weighting of each HMM state. The experimental results show an increase in recognition rate with duration modeling from 53.8% to 56.3% for initial consonant clusters, from 68.3% to 69.4% for rudiments and from 78.3% to 80.8% for suffixes, all from continuous speech in a speaker-independent mode.