End-of-Sentence Token Modeling for Streaming Conformer-Based Korean Children’s Speech Recognition Applied to a Social Robot

This paper proposes an end-of-sentence (EOS) token modeling approach for streaming conformer-based Korean children’s speech recognition applied to a social robot. The baseline system for speech recognition is a streaming conformer-based speech recognition model with the joint connectionist temporal classification attention-based encoder-decoder architecture, and it is trained with the AIHub Child•Infant Free-Talk dataset. However, it is difficult to determine sentence boundaries in children’s speech because the speaking style of children is somewhat different from that of adults. Thus, the proposed approach is to insert EOS tokens into each sentence of a large children’s dataset to detect sentence boundaries. Finally, we evaluated the performance of the proposed approach on a dataset collected under scenarios of a voice interaction between children and a social robot. Consequently, the proposed EOS token modeling decreased character error rate by 1.7%, compared to the baseline.