Abstract: Exploiting the novel human-machine interaction paradigm called brain–computer interface, we constructed a mental speller for those with neuromuscular disorders and motor disabilities. The carriers of communication were induced by a so-called “imitating natural reading” paradigm. Single-trial feature estimating should be used in this online paradigm instead of grand average that usually used in the cognitive or clinical experiments. we explored the single-trial estimation of P300 from channel Pz using support vector machine in three subjects, and gained a satisfied classification accuracy, which is 91.3%, 88.9% and 91.5% respectively. These results demonstrated the advantages of the inducing paradigm in constructing our mental speller.