Five different algorithms for speaker adaptation are investigated under realistic conditions: The new speaker can use the system immediately for his purpose without knowing about the adaptation, there is no announcement of a new speaker, a change of speakers can take place any time and the amount of computation and memory for adaptation is less than 1 % of the recognition task itself The adaptation is carried out by re-estimating the codebooks of a phoneme based word recognizer using semicontinuous hidden Markov models (SCHMM). The movement of the mean vector in the codebook according to his euclidean distance to the observed vector achieved the best results during very short adaptation sessions of 20 spoken words and for the long term adaptation in sessions of 100 spoken words.

Keywords: Speaker adaptation, codebook, semicontinuous EMM, realistic conditions, LVQ1, LVQ2