Free-form spoken input would be the easiest and most natural way for young children to communicate to an intelligent tutoring system. However, achieving such a capability poses a challenge both to instruction design and to automatic speech recognition. To address the difficulties of accepting such input, we adopt the framework of predictable response training, which aims at simultaneously achieving linguistic predictability and educational utility. We design instruction in this framework to teach children the reading comprehension strategy of self-questioning. To filter out some misrecognized speech, we combine acoustic confidence with language modeling techniques that exploit the predictability of the elicited responses. Compared to a baseline that does neither, this approach performs significantly better in concept recall (47% vs. 28%) and precision (61% vs. 39%) on 250 unseen utterances from 34 previously unseen speakers. We conclude with some design implications for future speech enabled tutoring systems. (Contains 2 tables, 3 figures and 1 footnote.)