Generally, children with speech sound disorders (SSDs) are clinically diagnosed by speech-language pathologists who transcribe the child's speech and manually analyse pronunciation errors. Although there have been several attempts to automatically analyse pronunciation errors in child SSD speech [1, 2], traditional automatic speech recognition (ASR) training methods were unable to achieve the desired recognition performance to replace human annotations. Training traditional ASR models requires a substantial amount of accurately annotated speech data. In addition, traditional ASR models require a pronunciation dictionary called a lexicon for combining in an acoustic model with learned speech features via phonetic symbols. A language model (LM) that has computed the probability of word chains based on correct grammar and vocabulary is also required [3]. For child SSD speech data, it is difficult and timeconsuming not only to gather a transcribed speech but also to build hand-designed pronunciation dictionaries with several pronunciations having the same spelling. However, over a few short years, the development of the end-to-end-based (e2e-based) model training method generated new possibilities in the spectrum of ASR [4, 5].