Competitive karuta is a Japanese sport in which players listen to a poem read by the reciter and take the corresponding card as fast as possible. In competitive karuta, the decisive part of each poem is called kimariji and it is written in hiragana. However, some players seem to identify the poems earlier than kimariji. In this study, we verify this hypothesis by Bayesian modeling and statistical speech analysis. We formulate the poem identification process as a sequential Bayes estimation and describe the recitations of poems by hidden Markov models. The results demonstrate that the three poems with kimariji “ooe”, “ooke”, and “ooko” can be identified already at the “oo” part.