An experiment was performed to clarify the acquisition process of automaticity in reading letters. First, a set of artificial (modified Sanskrit) letters were displayed one by one in random order with utterance of their names. The task of the subject was to remember and spell out those letters with their associated phonetic names on a strip of answer sheet. Next, taking up the strip, the experimenter displayed each letter for 500 msec, which the subject was asked to read aloud as fast and correctly as possible. The reading latency was repeatedly measured so as to obtain the latency vs. trials curve, even after the subject got able to spell all the letters without any error. The latency curve showed, in the asymptotic state, the acquisition of automatic processing of letter identification. The curve was composed of two different phases, which might be related to correct identification and correct spelling. Thsose phases were different in the absolute value level as well as in the time constant. But not so systematic was the difference among letters used as learning materials which differed in the number of features. The traditional theory of simple skill-learning and also that of feature analysis seemed to be inappropriate to explain all those results. For better interpretation a new hypothesis of topo-plastic image matching was proposed. It assumes plastic matching between one synthesized visual image and one memory image, but not element-to-element matching between so called the perceptual and the memory ;vectors. In that sense, it is a modified template matching. Under the hypothetical model, the first phase of the latency curve is conceived of as the gradual development of automatization of the topo-plastic matching mechanism which would reach the asymptote if the memory image was built infallible. The second phase, on the other hand, was interpreted as that of vocal naming, the asymptote of which converged on to the latency value level for familiar letters.