A method includes detecting a touch input by a user at a position on a virtual keyboard. The method further includes applying a spatial model to map the touch input to one or more first candidate keys of the virtual keyboard, wherein the spatial model maps the touch input to the one or more first candidate keys based on a heatmap of prior touch inputs by the user. The method further includes applying a language model to map the touch input to one or more second candidate keys of the virtual keyboard, wherein the language model maps the touch input to the one or more second candidate keys based on comparing the position of the touch input to one or more predicted next keys. The method further includes selecting a particular key of the one or more first and second candidate keys as indicated by the touch input.