True accessibility requires minimizing the scanning time to find a particular piece of information. Sequentially reading web pages do not provide this type of accessibility, for instance before the user gets to the actual text content of the page it has to go through a lot of menus and headers. However if the user could navigate a web page based through semantically classified blocks then the user could jump faster to the actual content of the page, skipping all the menus and other parts of the page. We propose a transcoding engine that tackles accessibility at two distinct, yet complementary, levels: for specific known sites and general unknown sites. We present a tool for building customized scripts for known sites that turns this process in an extremely simple task, which can be performed by anyone, without any expertise. For general unknown sites, our approach relies on statistical analysis of the structural blocks that define a web page to infer a semantic for the block.