A study was undertaken to investigate the variation in the speech of child second language learners as a function of the different discourse constraints imposed by three participant structures: narrator-audience, child-adult conversation, and child-child conversation. The subjects were 18 children aged 7 to 12, temporary residents in the United States, from nine native language groups. Data consisted of video and audiotaped interviews and videotaped interactions in a classroom of English as a second language. In the interviews, each child was questioned about his experiences learning English, everyday activities after school, and vacations, and was asked to narrate a familiar folktale. Portions of the discourse were analyzed for complexity, grammatical errors, dysfluency, and cohesion. Results show greatly varying communicative constraints in the three structures. In child-child conversations, communication involved short, routinized forms. In child-adult conversations the adult provided much support, and much of the child's speech was elliptical. However, in the narrative structure, the child is charged with providing coherent, cohesive discourse within minimum input, and it offers the greatest linguistic challenge. In this situation, only the most proficient speakers could elaborate syntax and use cohesive elements without increasing error and hesitation. It is proposed that the narrative structure provides the best communicative situation for measuring linguistic proficiency. (MSE)