This paper describes a speaker-independent recognition system for continuous German speech based on semicontinuous Hidden-Markov-Models which produces a phonetic transcription of the spoken sentence. The recognition units are parts of syllables while the output is a phoneme level transcription. During recognition, the phonotactic constraints of German are taken into account by a micro syntax constrained Viterbi algorithm. A maximum likelihood training procedure based on Viterbi training together with a simple but efficient seed model generation algorithm is presented.

Keywords: phonotactic constraints, semicontinuous HMMs, seed model generation, Viterbi training.