Shadowing was proposed originally in psycholinguistics to investigate listeners' process of perceiving speech, and now it is used widely in language education in Japan as a training method for learners to acquire a better skill in perceiving and producing L2 sounds. Recently, a shadowing-based assessment method was proposed to quantify learners' skill in perception and production. By integrating the two methods, this study aims at tracking gradual improvements of learners' skill during shadowing practices. We held a special program of 42-day Shadowing Marathon, where Japanese learners of English participated in shadowing practices every day for six weeks. Four new oral passages were presented daily, each of which was shadowed repeatedly. From the obtained data, we analyzed gradual improvements in learners' perception and production of both segments and prosody through repeating shadowing within a day, and through continuing the shadowing practices over 42 days. The results of analysis showed that, while learners' perception of segments and prosody improved significantly even with no explicit instructions, their production did not show significant improvement in a self-learning condition. The data also showed in what way learners depend on written input and auditory input when they learn perception and production of L2 sounds.