Speech recognition using HMM with decreased intra-group variation in the temporal structure