A main problem in continuous speech recognition consists of the coarticulation effects which make the recognition of phoneme-sized segments difficult. This problem can be reduced if larger units are used which contain the main coarticulation effects. An approach is presented which starts from an explicit localization of the syllabic nuclei, demisyllables are chosen as basic recognition units. The demisyllables are represented by Hidden Markov Models (HMM). Three different types of HMM's are exploited: discrete HMM's (vector quantization) , histogram HMM's and mixture Gaussian HMM's. The syllable boundaries are implicitly determined during the classification of the demisyllables. These methods constitute the acoustic-phonetic decoding module in a complete speech recognition system.