Why is children's ASR so difficult? Analyzing children's phonological error patterns using SSL-based phoneme recognizers

Children's automatic speech recognition (child ASR) is generally more challenging than adult ASR, as children's speech differs from that of adults and continuously evolves with a child's development. To investigate the causes of this difficulty and leverage the findings to improve the child ASR performance in the future, this study examines phoneme recognition error patterns using self-supervised learning (SSL)-based phoneme recognizers (PRs). First, we apply several SSL-based child PRs to English children's speech from kindergarten to grade 10, confirming that the WavLM-based model performs best. Then, using this model, we examine whether misrecognitions are primarily due to child mispronunciations or model errors. Furthermore, analyzing the results obtained by the WavLM-based model, we investigate how error patterns change along with age. The results show that mispronunciations persist longer than previously reported, particularly in certain manners of articulation.