This study examines the effects of a computer-assisted learning (CAL) program in which syllabic units were highlighted inside words in comparison with a CAL program in which the words were not segmented, i.e. one requiring whole word recognition. In a randomised control trial design, two separate groups of French speaking poor readers (2 * 14) in first grade were constituted. They were matched on a range of reading measures and non verbal intelligence and trained intensively over a short period (10 h over a period of 5 weeks). Three tasks were proposed using a classical pre-test/training/post-test design, written word recognition, word reading aloud and word spelling. In addition, three post-test sessions were conducted: one just after training, one after 4 months, and a last one after 9 months. The experimental group trained with the CAL using syllabic units outperformed the control group using CAL with whole word recognition in all the three tasks and there were important lasting effects. The results are discussed in the light of the self-teaching hypothesis and phonological recoding. (Contains 5 figures and 1 table.)