This work proposes a method to detect repetitions and corrections of words in children’s reading aloud tasks, through speech recognition using task-specific lattices. To be able to automatically evaluate the reading performance of children, as a replacement or complement for the current manual method of reading evaluation in schools, the deviations to the intended correct reading (disfluencies) should be detected. Here, we focus on detecting any extra events related to a word in a sentence such as repetitions or mistakes that are followed by correction. First, we build speech recognition models trained with a collected and annotated database of children’s speech. Using those models, our approach then decodes utterances by using a specific task lattice where the expected words are present but in a flexible way that also allows repetitions and syllable-based false starts, to find any word-related extra events. We achieve 15% miss rate for a 5% false alarm rate of these events, which shows we can provide improved time alignments for future annotations tasks.