That could be, but depending on the model architecture, it might depend on multi stroke sequences (similar to autocorrect on a phone) rather than identifying ea

That could be, but depending on the model architecture, it might depend on multi stroke sequences (similar to autocorrect on a phone) rather than identifying each individual keystroke in isolation. If that's the case, the layout may be baked into the model.