The invention relates to a method for generating a data flow from variable-length code words which are subdivided into a number of code word sets, whereby a raster comprising segments is determined for the data flow in which two adjacent raster points (41, 42) define a segment (40). The code words (1-6) from the first set are written into the data flow, beginning at raster points. Code words from the second set are then written into the data flow according to a predetermined allocation specification, whereby each code word from the second set is allocated to a separate segment. Whole code words or portions of code words which cannot be written according to their allocation are saved and entered into the data flow in additional writing attempts, whereby the allocation according to a predetermined specification is changed from one attempt to the next. This procedure is repeated in a similar manner for any additional data set that may exist. In this way, the ends of code words from the second set are decoupled from the beginnings of following code words from the second set, as the corresponding code words in a set are written in a segment by segment manner. This reduces the propagation of errors.