We aim to build a vast database (up to 9 million individuals) from the handwritten tabular nominal census of Paris of 1926Paris of , 1931Paris of and 1936, each composed of about 100,000 handwritten simple pages in a tabular format. We created a complete pipeline that goes from the scan of double pages to text prediction while minimizing the need for segmentation labels. We describe how weighted finite state transducers, writer specialization and self-training further improved our results. We also introduce through this communication two annotated datasets for handwriting recognition that are now publicly available, and an opensource toolkit to apply WFST on CTC lattices.