In reading support applications that assist children’s oral reading learning by detecting misread words, highly accurate recognition at the phoneme or mora level, rather than the word level used in conventional automatic speech recognition, is required. To achieve high-accuracy phoneme recognition, it is essential to analyze error tendencies, such as which phonemes are likely to be misrecognized. However, children’s speech has not been sufficiently investigated to date. In this study, we developed an end-to-end phoneme recognizer for children’s English speech and analyzed its phoneme recognition results to investigate age-dependent differences and characteristic patterns in phoneme recognition errors. The results revealed that phonemes produced with particular places or manners of articulation showed slower improvement in recognition accuracy with age compared to other phonemes. In addition, many substitution errors were found to originate from specific phonological processes. These findings suggest that actual pronunciation errors made by children are a dominant factor underlying the error tendencies of current children’s phoneme recognizers. (This work was conducted in collaboration with NTT Communication Science Laboratories.)