A virtual keyboard takes a large portion of precious screen real estate. We have investigated whether an invisible keyboard is a feasible design option, how to support it, and how well it performs. Our study showed users could correctly recall relative key positions even when keys were invisible, although with greater absolute errors and overlaps between neighboring keys. Our research also showed adapting the spatial model in decoding improved the invisible keyboard performance. This method increased the input speed by 11.5% over simply hiding the keyboard and using the default spatial model. Our 3-day multi-session user study showed typing on an invisible keyboard could reach a practical level of performance after only a few sessions of practice: the input speed increased from 31.3 WPM to 37.9 WPM after 20 - 25 minutes practice on each day in 3 days, approaching that of a regular visible keyboard (41.6 WPM). Overall, our investigation shows an invisible keyboard with adapted spatial model is a practical and promising interface option for the mobile text entry systems.