FUTO Swipe: Neural decoder works on any keyboard layout, no retraining

Researchers David Lee Miller and Aleksandras Kostarevas from FUTO have published FUTO Swipe, a neural swipe decoder that works on any contiguous mobile keyboard layout without retraining. Traditional swipe decoders are tied to a specific layout, requiring new data and training for each new keyboard. FUTO Swipe solves this by encoding each point along the swipe trajectory and predicting both whether the user is indicating a character and where on the keyboard that character lies. The keyboard layout is supplied only at inference time, so the model never memorizes positions.

To train this flexible model, the team released what they claim is the largest MIT-licensed swipe corpus, containing over 1 million donated swipes from more than 12,000 donor sessions. They apply geometric augmentations to both swipe trajectories and keyboard layouts at every training step, forcing the model to learn gesture characteristics rather than layout-specific patterns. The resulting model generalizes to layouts completely absent from training, in some cases achieving higher accuracy than the layout it was trained on. This combines the flexibility of algorithmic decoders with the accuracy of neural models. Pre-trained models and the corpus are publicly available, enabling broader adoption in custom keyboards and accessibility tools.