This paper presents a large scale study in a public museum setting, where a back-projected robot head interacted with the visitors in multi-party dialogue. The exhibition was seen by almost 8000 visitors, out of which several thousand interacted with the system. A considerable portion of the visitors were children from around 4 years of age and adolescents. The collected corpus consists of about 10.000 user utterances. The head and a multi-party dialogue design allow the system to regulate the turn-taking behaviour, and help the robot to effectively obtain information from the general public. The commercial speech recognition component, supposedly designed for adult speakers, had considerably lower accuracy for the children. Methods are proposed for improving the performance for that speaker category.