We have developed a user-trainable query by humming (QBH) system that develops an error probability model of a user's singing. While the training is effective, it is also tedious and time consuming, requiring the user to sing dozens of melodies to the system before the system can be trained. To make training fun, we introduce a new interactive, distributed karaoke game, called Karaoke Callout, played over a cell phone. The user selects a song and sings it into the cell phone. The audio is sent to a server which rates the quality of the singing by measuring how closely it resembles a canonical example of the song stored in the server database, sending a score back to the user. The user may then challenge anyone in the phone's contact list. An SMS text challenge is sent to the challenged person's cell phone. The challenged person sings the song, attempting to better the performance of the challenger. This challenge may then be repeated, with either party selecting a new song with which to "call out" the other party. Over the course of an interaction, numerous examples of each party's singing are created and stored. These may then be used to train a QBH to the idiosyncrasies of each user's singing, as well as providing new query targets for the system.