A number of text entry methods use a predictive completion based on letter-level n-gram model. In this paper, we investigate on an optimal length of n-grams stored in such model for a predictive keyboard operated by humming. In order to find the length, we analyze six different corpora, from which a model is built by counting number of primitive operations needed to enter a text. Based on these operations, we provide a formula for estimation of words per minute (WPM) rate. The model and the analysis results are verified in an experiment with three experienced users of the keyboard.