This paper presents Humsher -- a novel text entry method operated by the non-verbal vocal input, specifically the sound of humming. The method utilizes an adaptive language model for text prediction. Four different user interfaces are presented and compared. Three of them use dynamic layout in which n-grams of characters are presented to the user to choose from according to their probability in given context. The last interface utilizes static layout, in which the characters are displayed alphabetically and a modified binary search algorithm is used for an efficient selection of a character. All interfaces were compared and evaluated in a user study involving 17 able-bodied subjects.