A word wise search is performed on an MPEG-2 stream. For every word, the invention finds word-aligned patterns of 0×00 0×00 or 0×00 0×01. The algorithm applied by the invention examines the input stream buffer for the first word aligned 0 in which further testing determines is the first byte of a valid start code, and sets the sub-buffer defined by the start of the search to the location of this discovered start-code as the zero-word reach. A second search is performed in the same part of the input stream buffer, this time looking for word aligned 1's (i.e. byte pattern 0×00 0×01). For each word aligned 1 that is a start code, an entry is made into a start code list. When all of these have been found, the offset of the start code ending the current zero-word reach is added to the list of start-code offsets. This process is repeated from the 0 word reach until the end of the buffer is encountered. If 0's are not word aligned, the start codes are found on the 1's search, otherwise the start codes are found on the first search.