Analysing Korean children's speech data for early childhood educational services: age-specific insights from text and audio analysis

초록

As speech-based artificial intelligence (AI) becomes integrated into educational contexts, attention is growing towards its role in supporting child-centred learning environments. This study offers insights for developing child-friendly conversational AI systems by analysing age-specific linguistic and acoustic features in the speech of Korean-speaking children aged 4-9 years. The study was conducted in three phases: linguistic analysis of transcribed text, acoustic analysis of recorded utterances and automatic speech recognition (ASR) analysis. In the ASR phase, we benchmarked two modern models (Whisper and wav2vec2) using character error rate and performed a classification analysis to identify factors influencing recognition success, excluding age-related variables from model inputs. The results revealed age-related differences in vocabulary diversity, syntactic complexity, pitch, intensity and articulation rate, with younger children exhibiting more frequent pronunciation errors and lower ASR performance. Acoustic features, such as articulation patterns and pitch variability, were found to significantly influence recognition performance. These findings highlight the importance of designing AI systems that reflect children's developmental speech characteristics. Overall, this study provides an empirical foundation for improving speech-based AI interactions in early learning environments.

키워드